JP2012164026A - Image recognition device and display device for vehicle - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technology capable of performing processing for recognition while keeping high accuracy of recognizing an image of a recognition object from a captured image.SOLUTION: A night view device 100 is provided with an image recognition circuit 20 which includes: an image acquisition part 21 for acquiring a forward image 50 generated by a near-infrared camera 10; a cutout part 23 for sequentially cutting out cutout images 56 from the forward image 50; a first identification part 25 for determining whether or not there is a possibility that a pedestrian image 51 is included in each of a plurality of the cutout images 56 based on a first learning model 66; a second identification part 27 for determining whether or not the pedestrian image 51 is included in the cutout images 56 which is determined there is a possibility that the pedestrian image 51 is included based on a second learning model 77; and a recognition part 29 for recognizing a position in the forward image 50 of the cutout image 56 determined the pedestrian image 51 is included as the position where the pedestrian is photographed.

Description

  The present invention relates to an image recognition device for recognizing an image representing a recognition object from a captured image generated by an imaging means, and a vehicle display device including the image recognition device.

  2. Description of the Related Art Conventionally, a technique for recognizing an image of a recognition object such as a pedestrian and a vehicle shown in a captured image generated by an imaging unit such as a camera is known. For example, in the technique disclosed in Non-Patent Document 1, images of a predetermined size are cut out in order from an input image generated by an imaging unit. Then, based on a learning model created from a plurality of sample images (see FIG. 5 of Non-Patent Document 1) prepared as a reference for learning, images of pedestrians and the like that are assumed as recognition objects in each cut out image Is included. As a result of identification, the position in the input image of the image identified as including the pedestrian image is recognized as the position where the pedestrian image appears (see Non-Patent Document 1, FIG. 4 and the like).

  In the method disclosed in Non-Patent Document 1, the learning model as a discriminator used for discrimination repeats the process of selecting the feature amount most common to a plurality of sample images, and superimposes the plurality of feature amounts. A combination of feature quantities that can identify a recognition object is learned. Various methods for constructing a learning model by learning a feature amount of a recognition target object by repeating such a process are disclosed.

  As such a technique, for example, Patent Document 1 discloses pattern recognition for identifying a recognition target object based on a learning model created by learning feature quantities of sample images such as numbers using principal component analysis. An apparatus is disclosed. Patent Document 2 discloses an image recognition device that recognizes an abnormal state of a monitoring target based on a learning model created by learning a feature amount of a reference image in a normal state using a neural network. ing. Patent Document 3 discloses a vehicle image for recognizing a vehicle image based on an SVM classifier created by learning a feature amount of a learning sample such as a vehicle image using a support vector machine (SVM). A recognition device is disclosed. Further, in Patent Document 4, a face image in an input image is identified based on a pattern identification parameter created by learning feature amounts of a face image and a non-face image prepared as sample images. A pattern identification device is disclosed.

Japanese Patent Laid-Open No. 05-40852 Japanese Patent Laid-Open No. 11-7536 JP 2007-235951 A JP 2009-86749 A

Paul Viola, Michael J. Jones, Daniel Snow, “Detecting Pedestrians Using Patterns of Motion and Appearance”, In iccv03 (Nice, France, 2003), pp. 734-741

  Now, in the input image input to the pattern identification device disclosed in Patent Document 4, for example, some object such as a face or a torso is included, and there is a possibility that an image of the face is included. The difference from the background, etc., which clearly contains only different non-faces, is clear. Therefore, even a simple pattern identification parameter that has learned only a small amount of features can be identified. However, in the input image, the difference between a face image that includes a face image and a non-face image that does not include a face image but includes an image similar to a face is ambiguous. Therefore, since it is necessary to learn detailed feature amounts, the parameters for pattern identification must be complicated.

  Therefore, when trying to recognize a pedestrian with the feature amount disclosed in Non-Patent Document 1 using the pattern identification device described in Patent Document 4, for all the cut-out images cut out in order from the captured image It is necessary to identify whether or not a pedestrian is included by using a pattern identification parameter. Similar to the face recognition described above, the parameter for pattern identification in which an ambiguous difference between a pedestrian and a non-pedestrian similar to the pedestrian is learned is complicated. Therefore, since the identification process based on the parameters is performed with a large amount of feature amounts even with respect to a clipped image that can be identified with a small amount of features, the processing becomes long and redundant. As described above, the entire processing amount from the captured image until the image of the recognition target object such as a pedestrian is recognized can be enormous.

  The present invention has been made in view of the above problems, and an object of the present invention is to perform recognition processing at high speed while maintaining high accuracy of recognizing an image of a recognition object from a captured image. An image recognition device that can be used, and a vehicle display device including the image recognition device.

  In order to achieve the above object, according to the first aspect of the present invention, an image acquisition unit that acquires a captured image generated by the imaging unit, and a captured image acquired by the image acquisition unit have a preset size. Based on a first learning model created from a first sample image group serving as a learning reference, a cutting unit that sequentially cuts out the cut image and a plurality of cut images cut out by the cutting unit are set in advance. A first identification means for identifying whether or not there is a possibility that an image representing the recognized object is included, and the first identification means identifies that there is a possibility that the image of the recognition object is included. A second identification means for identifying whether or not an image of a recognition object is included based on a second learning model created from a second sample image group serving as a learning reference for the cut image; Second identification The position in the captured image with the identified clipped images includes an image of the recognition object by stage to recognizing means for recognizing a position where the recognition object objects appear, the image recognition apparatus including a.

  In general, in each cut-out image cut out sequentially from the captured image by the cut-out means, there is only a non-recognition object that may include an image of a preset recognition object and a recognition object that is clearly different from the recognition object. The difference from the non-recognized object is clearer than the difference between the object including the image of the recognition object and the object including the non-recognition object similar to the recognition object. Therefore, the first learning model for identifying whether there is a possibility that the image representing the recognition target object is included in the cut-out image is a non-recognition target object that is clearly different from the recognition target object. Can be configured simply even if high accuracy is maintained. On the other hand, the second learning model for identifying whether the image of the recognition object is included in the cut-out image is a learning model for identifying the recognition object and a non-recognition object similar to the recognition object. For this reason, in order to maintain high accuracy, it must be complicated.

  According to the first aspect of the present invention, the cut-out image that is the target of the identification process by the second identification unit may include the image of the recognition target object by the identification process by the first identification unit. It is narrowed down to. In the identification processing by the first identification means, identification between the possibility of including an image of a recognition target object and that apparently including only a non-recognition target object is performed with high accuracy. Therefore, even if the cut-out image is narrowed down by the first identification unit, the accuracy of the entire process for recognizing the image of the recognition object from the captured image can be maintained high.

  In addition, the identification process by the first identification unit based on the simple first learning model can be processed at a higher speed than the identification process by the second identification unit based on the complicated second learning model. Therefore, by narrowing down the cut-out images by the identification processing by the first identification means, the cut-out images that are the targets of the identification processing by the redundant second identification means are reduced. As a result, the entire process from the captured image to the recognition of the recognition target image can be speeded up.

  Therefore, the image recognition apparatus can perform recognition processing at high speed while maintaining high accuracy in recognizing the image of the recognition target object from the captured image.

  In the invention according to claim 2, the first sample image group includes a plurality of positive sample images in which an assumed object assumed as a recognition object is captured, and a plurality of negative example samples in which a non-assumed object other than the assumed object is captured. The first learning model is created by learning a negative example sample image showing a non-assumed object similar to the assumed object among a plurality of negative example sample images as a positive example together with a plurality of positive example sample images. The first identifying means identifies whether or not there is a possibility that an image of the recognition object is included in each cut-out image depending on whether or not the positive example learned by the first learning model corresponds to the first example. It is characterized by doing.

  In the present invention, in the first sample image group, the difference between a positive example sample image in which an assumed object assumed as a recognition object is captured and a negative example sample image in which an unimagined object similar to the assumed object is captured is ambiguous. is there. On the other hand, in a plurality of negative sample images, the difference between a non-assumed object similar to the assumed object and an unimagined object similar to the assumed object such as the background is clear. Therefore, the first learning model contributes to the speeding up of the processing by the first identification means by learning the negative example sample image in which the non-assumed object similar to the assumed object is captured as a positive example together with the positive example sample image. It can be a simple learning model.

  In addition, the negative sample image that shows an unforeseen object similar to the assumed object is also included in the positive example of the first learning model, so that in addition to the cut-out image that includes the image of the recognition object, the recognition object A cut-out image including an image of an object similar to can also correspond to a positive example. These cut-out images are cut-out images that may include an image of the recognition object. Therefore, the 1st identification means can narrow down appropriately the cut-out image which may contain the image of a recognition target object by identifying the cut-out image corresponding to a positive example.

  Then, the second identification unit identifies whether or not the image of the recognition target object is included in the cut image narrowed down by the first identification unit, so that an image similar to the recognition target object is captured. Can be excluded. As described above, the cut-out image including the image of the recognition object is identified with high accuracy by the identification processing of the first identification unit and the second identification unit. Therefore, the image recognition apparatus can reliably maintain high accuracy in recognizing the image of the recognition target object from the captured image even if the processing for recognition is accelerated.

  In the invention according to claim 3, the first sample image group includes a plurality of positive sample images in which an assumed object assumed as a recognition object is reflected, and a plurality of negative example samples in which a non-assumed object other than the assumed object is reflected. The first learning model includes an image, and learns as a positive example an image group in which a positive example sample image in which a hypothetical object similar to an unanticipated object is captured from a plurality of positive example sample images. Created by learning as a negative example an image group that excludes negative example sample images in which an unforeseen object similar to the assumed object is shown from the image, the first identification means is a positive example learned by the first learning model. It is characterized by identifying whether or not there is a possibility that an image of a recognition object is included in each cut-out image depending on whether or not they are related.

  As in the present invention, in the first sample image group, the difference between the image of the assumed object similar to the unexpected object and the image of the unexpected object similar to the assumed object is ambiguous. Therefore, a positive example sample image in which an assumed object similar to the unexpected object is shown and a negative sample image in which an unexpected object similar to the assumed object is shown are excluded from the first sample image group. By creating the first learning model from such a first sample image group, the first learning model can be surely a simple learning model that contributes to speeding up the processing by the first identification means.

  In addition, negative sample images showing images of non-assumed objects similar to the assumed objects are excluded from the negative examples of the first learning model, so in addition to the cut-out images containing the images of the recognition objects, the recognition objects A cut-out image including an image of an object similar to an object can also correspond to a cut-out image related to the positive example. Therefore, the 1st identification means can narrow down appropriately the cut-out image which may contain the image of a recognition target object by identifying the cut-out image relevant to a positive example.

  Then, the second identification unit identifies whether or not the image of the recognition target object is included in the cut image narrowed down by the first identification unit, so that an image similar to the recognition target object is captured. Can be excluded. As described above, the cut-out image including the image of the recognition object is identified with high accuracy by the identification processing of the first identification unit and the second identification unit. Therefore, the image recognition apparatus can reliably maintain high accuracy in recognizing the image of the recognition target object from the captured image even if the processing for recognition is accelerated.

  In the invention according to claim 4, the first sample image group includes a plurality of positive sample images in which an assumed object assumed as a recognition object is captured, and a plurality of negative example samples in which a non-assumed object other than the assumed object is captured. The first learning model includes a plurality of positive example sample images as a positive example in the first sample image group, and the first identification unit relates to a positive example learned by the first learning model. Whether or not there is a possibility that the image of the recognition target object is included in each cut-out image.

  In the present invention, the first learning model learns a plurality of positive example sample images as positive examples in the first sample image group. Since such a first learning model does not learn an ambiguous difference between an assumed object and an unimagined object, it can be surely a simple learning model that contributes to speeding up the processing by the first identification means.

  The first learning model as described above does not learn as a negative example a negative example sample image in which an image of an unforeseen object similar to the assumed object is captured. Therefore, in addition to the cut-out image including the image of the recognition target object, the cut-out image including the image of the object similar to the recognition target object can also correspond to the cut-out image related to the positive example. Therefore, the 1st identification means can narrow down appropriately the cut-out image which may contain the image of a recognition target object by identifying the cut-out image relevant to a positive example.

  Then, the second identification unit identifies whether or not the image of the recognition target object is included in the cut image narrowed down by the first identification unit, so that an image similar to the recognition target object is captured. Can be excluded. As described above, the cut-out image including the image of the recognition object is identified with high accuracy by the identification processing of the first identification unit and the second identification unit. Therefore, the image recognition apparatus can reliably maintain high accuracy in recognizing the image of the recognition target object from the captured image even if the processing for recognition is accelerated.

  In the invention according to claim 5, the second sample image group includes a plurality of positive sample images in which an assumed object assumed as a recognition object is reflected, and a plurality of negative example samples in which a non-assumed object other than the assumed object is reflected. The second learning model includes an image and is created by learning a plurality of positive example sample images as positive examples and learning a negative example sample image as a negative example. Whether or not the image of the recognition object is included in the cut-out image is identified based on whether or not it corresponds to the learned positive example.

  According to the present invention, the second learning model is created from the second sample image group including the plurality of positive example sample images in which the assumed object is photographed and the plurality of negative example sample images in which the unimaginable object is photographed. I am learning in detail about ambiguous differences between unexpected and unforeseen things. Therefore, the second identification unit identifies the clipped image corresponding to the positive example of the second learning model, thereby identifying the recognition target from the clipped images that may include the image of the recognition target object. A cut image including an image of an object similar to an object can be accurately excluded. As described above, the cut-out image including the image of the recognition object is accurately identified by the identification processing of the first identification unit and the second identification unit.

  In the invention according to claim 6, the first identification means identifies whether or not there is a possibility that the image of the pedestrian that is the recognition object is included in the plurality of cut-out images, and the second identification means Is characterized by identifying whether or not the pedestrian image is included in the cut-out image identified as possibly having a pedestrian image included by the first identification means.

  As in the present invention, the pedestrian image shown in the captured image has various forms depending on the moving direction of the pedestrian and the relative position of the pedestrian with respect to the imaging means. Therefore, in order to accurately identify whether or not the image is a pedestrian image, the learning model has to learn the characteristics of the pedestrian image from a very large number of sample images. Then, since the learning model becomes complicated, the identification process may become further redundant.

  However, the image recognition apparatus described above reduces the overall processing amount for recognizing the recognition target object from the captured image by narrowing the cut-out image that is the target of the identification process of the second identification unit by the first identification unit. is doing. Therefore, even if the second learning model used by the second identification unit becomes a complex model, the image recognition apparatus can recognize a pedestrian at high speed while maintaining high accuracy. As described above, the configuration in which the first identification unit narrows the cut image is particularly suitable for an image recognition apparatus that recognizes a pedestrian image as a recognition target image.

  In the invention according to claim 7, the image of the recognition object existing in the peripheral area is recognized from the captured image generated by the imaging unit that captures the peripheral area of the vehicle, and the image of the recognition object is emphasized. In the display apparatus for vehicles which displays an emphasis image superimposed on a picked-up image, and acquires a picked-up image from an image pick-up means, In the image recognition device according to any one of claims 1 to 6, and a picked-up image, A display device for a vehicle is provided that includes a display unit that superimposes and displays an emphasized image at a position where a recognition object recognized by the image recognition device appears.

  As in the present invention, in the vehicle display device, it is necessary to promptly notify the operator of the vehicle of the presence of the recognition object located in the peripheral region of the vehicle. Therefore, an image recognition apparatus that recognizes an image of a recognition object from a captured image is strongly required to increase the processing speed for recognition. Therefore, an image recognition device that realizes high-speed recognition by narrowing the cut-out image by the identification processing of the first identification means based on a simple first learning model is used for a vehicle display device and is used in a vehicle peripheral region. This is particularly suitable as a configuration for recognizing a recognition object.

1 is a configuration diagram schematically illustrating a configuration of a night view apparatus according to a first embodiment of the present invention. It is a figure by which the front image displayed on a liquid crystal display is shown. It is a figure for demonstrating the 1st learning model by 1st embodiment, Comprising: (a) is a figure for demonstrating the detail of a 1st sample image group, (b) is a 1st sample image group. It is a figure for demonstrating that a 1st learning model is constructed | assembled by inputting into a learning model production | generation apparatus. It is a figure for demonstrating the 2nd learning model by 1st embodiment, Comprising: (a) is a figure for demonstrating the detail of a 2nd sample image group, (b) is a 2nd sample image group. It is a figure for demonstrating that a 2nd learning model is constructed | assembled by inputting into a learning model production | generation apparatus. It is a flowchart in which the process of the image recognition circuit which recognizes a pedestrian image from a front image is shown. It is the figure which compared the difference in the processing amount and recognition rate for recognition resulting from the presence or absence of narrowing down of the cut-out image by a 1st identification part. It is a figure for demonstrating the 1st learning model by 2nd embodiment, Comprising: It is a figure which shows the modification of FIG. It is a figure for demonstrating the 1st learning model by 3rd embodiment, Comprising: It is a figure which shows another modification of FIG.

  Hereinafter, a first embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a configuration diagram schematically showing the configuration of a night view apparatus 100 according to the first embodiment of the present invention. The night view device 100 is mounted on a vehicle and recognizes an image of a recognition object such as a pedestrian (hereinafter referred to as a pedestrian image 51) existing in a front area of the vehicle. And the night view apparatus 100 superimposes and displays the frame image 53 for emphasizing the pedestrian image 51 on the front image 50, as FIG. 2 shows. Hereinafter, the configuration of the night view apparatus 100 will be described in detail with reference to FIGS. 1 and 2.

  The night view apparatus 100 is connected to the near infrared camera 10. The night view apparatus 100 includes a night view control circuit 15 that acquires the front image 50 generated by the near infrared camera 10 and a liquid crystal display 40 that displays the front image 50 output from the night view control circuit 15.

  The near-infrared camera 10 is installed so that the traveling direction of the vehicle is directed, and images a region in front of the vehicle, in particular, of the vehicle peripheral region. The near-infrared camera 10 can photograph the state of the front region even in an environment with little visible light by detecting near-infrared light having a wavelength of about 0.7 to 2.5 micrometers. The vehicle on which the near-infrared camera 10 is mounted is provided with a near-infrared projector 11 that is configured integrally with, for example, a headlight module and that projects near-infrared rays toward the front region. The near-infrared camera 10 is connected to the near-infrared projector 11 and starts photographing the front area and causes the near-infrared projector 11 to start projecting light. The near-infrared camera 10 generates a front image 50 as a captured image by detecting near-infrared light reflected by an object in the front region. The near-infrared camera 10 sequentially outputs the generated front image 50 toward the night view control circuit 15.

  The night view control circuit 15 is connected to the near infrared camera 10 and the liquid crystal display 40. The night view control circuit 15 includes an image recognition circuit 20, a drawing circuit 30, and the like. The image recognition circuit 20 and the drawing circuit 30 are each a processor that performs various arithmetic processes for control and drawing, a flash memory that stores programs used for the arithmetic processes, and a RAM that functions as a work area for the arithmetic processes Etc. are constituted.

  The image recognition circuit 20 includes an image acquisition unit 21, a cutout unit 23, a first identification unit 25, a second identification unit 27, and a recognition unit 29 as functional blocks by causing a processor to execute a predetermined program. The image acquisition unit 21 acquires the front image 50 generated by the near-infrared camera 10 and outputs it to the cutout unit 23.

  The cutout unit 23 uses the cutout window 55 to cut out a cutout image 56 having a preset size in order from the front image 50. The cutout unit 23 moves the cutout window 55 gradually in the horizontal direction and the vertical direction of the front image 50 (see the dashed arrows in FIG. 2), and cuts the image in the range surrounded by the cutout window 55. The extracted image 56 is cut out. When the cutout window 55 reaches the lower right, the cutout unit 23 changes the size of the cutout window 55 and cuts out the cutout image 56 from the front image 50 again. It should be noted that the region where the sky is mainly photographed in the front image 50 may be excluded from the moving range of the cutout window 55.

  The first identification unit 25 identifies whether or not there is a possibility that the pedestrian image 51 is included based on the first learning model 66 for the plurality of cut images 56 cut by the cut unit 23. To do. As a result, the first identification unit 25 narrows down the cut-out images 56 to be subjected to the identification process by the second identification unit 27. First, the first identification unit 25 calculates the Haar-like feature amount of each cut image 56 cut out. The Haar-like feature is a combination of a white rectangular area and a black rectangular area (see Non-Patent Document 1 FIG. 1). Various shapes are defined as such Haar-like features. The Haar-like feature value is a value based on the difference between the average luminance of the white rectangular area of the Haar-like feature and the average luminance of the black rectangular area.

  In the first learning model 66, the shape, position, direction, threshold for identification, and the like of Haar-like features effective for identifying an image that may include the pedestrian image 51 are learned. Based on the first learning model 66, the first identification unit 25 calculates the Haar-like feature amount in the specific region of the cut-out image 56, and compares it with the learned threshold value. Based on the first learning model 66, the first identification unit 25 repeats the identification based on the Haar-like feature value, so that each cut-out image 56 corresponds to the positive example learned by the first learning model 66. Identify whether or not. Thereby, the 1st identification part 25 identifies that the pedestrian image 51 may be contained in the cut-out image 56 corresponding to a positive example.

  Based on the second learning model 77, the second identification unit 27 uses the second learning model 77 to identify the pedestrian image 51 for the clipped image 56 that has been identified by the first identification unit 25 as possibly including the pedestrian image 51. Identifies whether it is included. Thereby, the 1st identification part 25 excludes the cut-out image 56 in which the object similar to a pedestrian is reflected from the cut-out images 56 identified as the possibility that the pedestrian image 51 may be included. First, the second identification unit 27 calculates the Haar-like feature amount of each cut-out image 56 narrowed down by the first identification unit 25. In the second learning model 77, the shape, position, direction, threshold value for identification, and the like of the Haar-like feature effective for identifying the image including the pedestrian image 51 are learned. Based on the second learning model 77, the second identification unit 27 calculates the Haar-like feature amount in the specific region of the cutout image 56 and compares it with the learned threshold value. Based on the second learning model 77, the second identification unit 27 repeats the determination based on the Haar-like feature value, so that each cut-out image 56 corresponds to the positive example learned by the second learning model 77. Identify whether or not. And the 2nd identification part 27 identifies that the pedestrian image 51 is contained in the cut-out image 56 corresponding to a positive example.

  The recognizing unit 29 recognizes the position in the front image 50 of the cut image 56 identified as including the pedestrian image 51 by the second identifying unit 27 as the position where the pedestrian appears in the front image 50. By repeating the identification processing by the first identification unit 25 and the second identification unit 27, the recognition unit 29 can recognize all the positions where the pedestrian appears in one front image 50. The recognition unit 29 outputs position information of the recognized pedestrian image 51 in the front image 50 to the drawing circuit 30.

  When the pedestrian image 51 is recognized in the front image 50, the drawing circuit 30 performs a composition process for superimposing the frame image 53 on the front image 50. Specifically, the drawing circuit 30 draws the frame image 53 on a predetermined layer based on the position information of the pedestrian image 51 acquired from the recognition unit 29. Then, the drawing circuit 30 superimposes the layer on which the frame image 53 is drawn on the layer on which the front image 50 is drawn so that the frame image 53 surrounds the pedestrian image 51. The drawing circuit 30 sequentially outputs the front image 50 on which the frame image 53 is superimposed toward the liquid crystal display 40. When the pedestrian image 51 is not recognized by the image recognition circuit 20, the drawing circuit 30 sequentially outputs the front image 50 toward the liquid crystal display 40 without performing the frame image 53 combining process.

  The liquid crystal display 40 is a dot matrix display device capable of color display by controlling a plurality of pixels arranged on the display surface. The liquid crystal display 40 is disposed inside the instrument panel in the vehicle compartment with the front side shown in FIG. 2 facing the driver's seat. The liquid crystal display 40 can continuously display the front image 50 sequentially acquired from the drawing circuit 30, thereby showing the state of the front region to the viewer with a moving image.

  Next, the first learning model 66 used by the first identification unit 25 and the second learning model 77 used by the second identification unit 27 will be described in more detail based on FIGS.

  The first learning model 66 is a learning model for identifying whether or not the pedestrian image 51 may be included in the cutout image 56. The first learning model 66 is created by causing the learning model generation device 80 to learn the first sample image group 60 serving as a learning reference. The second learning model 77 is a learning model for identifying whether or not the pedestrian image 51 is included in the cutout image 56. The second learning model 77 is created by causing the learning model generation device 80 to learn the second sample image group 70 that serves as a learning reference.

  The first sample image group 60 and the second sample image group 70 include a plurality of positive sample images 61 and 71 in which pedestrians assumed as recognition objects are shown, vehicles other than pedestrians, signs, and backgrounds. Etc. (hereinafter referred to as non-pedestrians) and a plurality of negative example sample images 63 and 73 are included. The positive sample images 61 and 71 include positive sample images 61a and 71a that can clearly determine that a pedestrian is shown, and positive sample images 61b and 71b that show a pedestrian similar to a non-pedestrian. It is. Similarly, in the negative example sample images 63 and 73, the negative example sample images 63a and 73a that can clearly determine that no pedestrian is shown, and the negative example sample images 63b and 63b that show a non-pedestrian similar to the pedestrian. 73b is included. The second sample image group 70 may be the same image group as the first sample image group 60, or may be an image group different from the first sample image group 60.

  The learning model generation device 80 includes a processor that performs various arithmetic processes, a storage medium such as a hard disk drive that stores programs used for the arithmetic processes, and a RAM that functions as a work area for the arithmetic processes. . When creating the first learning model 66 and the second learning model 77, the learning model generation device 80 includes the first sample image group 60 or the second sample image group 70 and whether each sample image is a positive example or a negative example. Is input. The learning model generation device 80 calculates the feature amount of each input sample image while changing the shape, position, and direction of the Haar-like feature, and accumulates the calculation results. Then, the learning model generation apparatus 80 has a Haar-like feature that is common to the sample image associated with the teacher signal indicating the positive example and not common to the sample image associated with the teacher signal indicating the negative example, such as AdaBoost or the like. Learning using the boosting algorithm.

  As a feature of the first learning model 66, a negative example sample image 63 b showing a non-pedestrian similar to a pedestrian among a plurality of negative example sample images 63 is learned as a positive example together with a plurality of positive example sample images 61. ing. Specifically, the negative example sample image 63b showing a non-pedestrian similar to a pedestrian is associated with a teacher signal indicating that it is a positive example when input to the learning model generation device 80. Thereby, in the identification of the 1st identification part 25 based on the 1st learning model 66, in addition to the cut-out image 56 containing the pedestrian image 51, the cut-out image 56 in which the object similar to a pedestrian is reflected is also a positive example. Is identified as a cut-out image 56 corresponding to. The cutout image 56 including the pedestrian image 51 and the cutout image 56 in which an object similar to the pedestrian is captured are cutout images 56 that may include the pedestrian image 51. Therefore, the 1st discrimination | determination part 25 identifies the cutout image 56 corresponding to the positive example learned by the 1st learning model 66 mentioned above, and the cutout which may contain the pedestrian image 51 is included. The image 56 can be accurately narrowed down.

  In the second learning model 77, a plurality of positive example sample images 71 are associated with teacher signals indicating positive examples. Further, unlike the case of the first learning model 66, the plurality of negative example sample images 73 are all associated with teacher signals indicating negative examples. Therefore, in the identification of the second identification unit 27 based on the second learning model 77, a clipped image 56 that shows an object similar to a pedestrian is excluded, and a clipped pedestrian image 51 corresponding to the positive example is included. An outgoing image 56 is identified.

  When the first learning model 66 and the second learning model 77 are compared, the second learning model 77 has a larger number of searches for Haar-like features for identification due to the following reasons, and is a complicated learning model. Become. The reason is that it is difficult to distinguish these images because the difference between an image of a pedestrian similar to a non-pedestrian and an image of a non-pedestrian similar to a pedestrian is ambiguous. . On the other hand, the difference between an image that may include a pedestrian and an image that includes only a non-direction person that is clearly different from other pedestrians, for example, is relatively clear. Therefore, identification is easy. For the above reasons, the first learning model 66 that learns as a positive example the negative sample image 63b in which an image of a non-pedestrian that is similar to a pedestrian is captured is a Is a learning model for identifying Therefore, even if the first learning model 66 maintains high accuracy, the first learning model 66 can be simply configured to contribute to speeding up the identification processing by the first identification unit 25. On the other hand, the second learning model 77 for discriminating between an image showing a pedestrian similar to a non-pedestrian and an image showing a non-pedestrian similar to a pedestrian is a non-walking similar to a pedestrian and a pedestrian. In order to maintain high accuracy by using a learning model for identifying a person, it must be complicated.

  A process in which the night view apparatus 100 described so far displays the front image 50 on which the frame image 53 is superimposed on the liquid crystal display 40 will be described in detail with reference to FIG. The processing described below is performed by the image recognition circuit 20 when a switch for starting and stopping the operation of the night view apparatus 100 is operated by the user. The processing by the image recognition circuit 20 is repeated until the operation of the night view device 100 is stopped based on the user's operation.

  In S101, the front image 50 is acquired from the near-infrared camera 10, and it progresses to S102. In S102, the cutout image 56 is cut out from the front image 50 acquired in S101 using the cutout window 55, and the process proceeds to S103.

  In S <b> 103, whether or not there is a possibility of including the pedestrian image 51 is identified based on the first learning model 66 for the cut image 56 cut out in S <b> 102. If it is determined in S103 that there is no possibility of including the pedestrian image 51, the process proceeds to S106. On the other hand, if it is determined in S103 that the pedestrian image 51 may be included, the process proceeds to S104.

  In S <b> 104, whether or not a pedestrian is included is identified based on the second learning model 77 for the cut image 56 that has been identified as having a possibility of including the pedestrian image 51 in S <b> 103. If it is determined in S104 that the pedestrian image 51 is not included, the process proceeds to S106. On the other hand, if it is determined in S104 that the pedestrian image 51 is included, the process proceeds to S105.

  In S105, the position where the clipped image 56 is clipped in S102 is recognized as a position where a pedestrian appears in the front image 50, and the process proceeds to S106. In S106, it is identified whether or not the identification process has been completed for all the cut images 56 cut out from the front image 50 acquired in S101. If it is determined in S106 that the identification processing for all the cutout images 56 has not been completed, the process returns to S102, and the processing of S103 to S105 is performed for the next cutout image 56. If it is determined in S106 that the identification processing for all the cutout images 56 has been completed, the process proceeds to S107.

  In S107, the positional information in the front image 50 of the pedestrian image 51 recognized in S105 is output to the drawing circuit 30, and the process returns to S101. And the process for recognizing the pedestrian of S102-S106 is implemented about the front image 50 acquired next. The drawing circuit 30 that has acquired the position information of the pedestrian image 51 output in S107 superimposes the frame image 53 on the front image 50 in accordance with the position information. The drawing circuit 30 continuously outputs the front image 50 on which the frame image 53 is superimposed to the liquid crystal display 40. Thereby, the state of the front area is displayed on the liquid crystal display 40 as a moving image.

  In the first embodiment described so far, the clipped image 56 that is the target of the identification process of the second identification unit 27 may include the pedestrian image 51 by the identification process by the first identification unit 25. The extracted image 56 is narrowed down. In the identification processing by the first identification unit 25, since an image with a clear difference is identified, a pedestrian image 51 may be included and a non-pedestrian image may be included. The distinction is made with high accuracy. Therefore, even if the cut-out image 56 is narrowed down by the first identification unit 25, the accuracy of the entire process for recognizing the pedestrian image 51 from the front image 50 can be maintained high.

  In addition, the identification process by the first identification unit 25 based on the simple first learning model 66 can be processed at a higher speed than the identification process by the second identification unit 27 based on the complicated second learning model 77. Therefore, by narrowing down the cutout image 56 by the identification process by the first identification unit 25, the cutout image 56 that is the target of the identification process of the redundant second identification unit 27 is reduced. As a result, even when the first identifying unit 25 and the second identifying unit 27 perform overlapping identification processing on some of the cut-out images 56, the entire process until the pedestrian image 51 is recognized from the front image 50. This process can be speeded up.

  When narrowing the cut image 56 as in the first embodiment, and without narrowing the cut image 56 by the first identification unit 25, all the cut images 56 are targeted for identification processing of the second identification unit 27. The comparison result with the case is shown in FIG. By narrowing the cut-out image 56 by the first identification unit 25, the number of times that the Haar-like feature is searched for by the first identification unit 25 and the second identification unit 27 is reduced by about 75%. In addition, the recognition rate remains high.

  Therefore, the image recognition circuit 20 included in the night view device 100 can perform recognition processing at high speed while maintaining high accuracy of recognizing the pedestrian image 51 from the front image 50.

  In addition, the first learning model 66 of the first embodiment learns a negative example sample image 63b in which a non-pedestrian image similar to a pedestrian is captured as a positive example. Therefore, the first identification unit 25 that performs the identification process based on the first learning model 66 may include the pedestrian image 51 by identifying the cutout image 56 corresponding to the positive example. The cut image 56 can be accurately narrowed down.

  Further, the second learning model 77 learns a positive example sample image 71b showing a pedestrian similar to a non-pedestrian as a positive example, and takes a negative example sample image 73b showing a non-pedestrian similar to a pedestrian as a negative example. Learning. Therefore, the second learning model 77 learns in detail about ambiguous differences between pedestrians and non-pedestrians. As described above, the second identifying unit 27 identifies the clipped image 56 corresponding to the positive example of the second learning model 77, and thus the pedestrian image 51 may be included. Therefore, the cut-out image 56 in which an object similar to a pedestrian is captured can be accurately excluded. Therefore, the image recognition circuit 20 can reliably maintain high accuracy for recognizing the pedestrian image 51 from the front image 50 even if the processing for recognition is speeded up.

  In the first embodiment, the pedestrian image 51 included in the front image 50 has various forms according to the actual movement direction of the pedestrian and the relative position of the pedestrian with respect to the near-infrared camera 10. Therefore, in order to accurately identify whether or not the image is the pedestrian image 51, the learning model has to learn the characteristics of the pedestrian image 51 from a very large number of sample images. Then, since the learning model becomes complicated, the identification process may become further redundant.

  However, the image recognition circuit 20 of the first embodiment recognizes the pedestrian image 51 from the front image 50 by narrowing down the cutout image 56 that is the target of the identification processing of the second identification unit 27 by the first identification unit 25. To reduce the overall processing amount. Therefore, even if the second learning model 77 used by the second identification unit 27 becomes complicated, the image recognition circuit 20 can recognize the pedestrian image 51 at high speed while maintaining high accuracy. As described above, the configuration in which the cutout image 56 is narrowed down by the first identification unit 25 is particularly suitable for the image recognition circuit 20 that recognizes the pedestrian image 51 as the image of the recognition object.

  Furthermore, the night view apparatus 100 as in the first embodiment must promptly notify the operator of the vehicle of the presence of a pedestrian located in the peripheral area of the vehicle. Therefore, the image recognition circuit 20 that recognizes the pedestrian image 51 from the front image 50 is strongly required to speed up the process for recognition. Therefore, the image recognition circuit 20 that realizes high-speed recognition by narrowing the cut image 56 by the identification processing of the first identification unit 25 based on the simple first learning model 66 is used in the night view device 100. This is particularly suitable as a configuration for recognizing pedestrians in the surrounding area of the vehicle.

  In the first embodiment, the near-infrared camera 10 corresponds to the “imaging means” recited in the claims, the image recognition circuit 20 corresponds to the “image recognition device” recited in the claims, and the image The acquisition unit 21 corresponds to the “image acquisition unit” described in the claims, the cutout unit 23 corresponds to the “cutout unit” described in the claims, and the first identification unit 25 includes the claims. The second identification unit 27 corresponds to the “second identification unit” described in the claims, and the recognition unit 29 corresponds to the “recognition unit” described in the claims. The drawing circuit 30 and the liquid crystal display 40 correspond to the “display means” described in the claims, the front image 50 corresponds to the “captured image” described in the claims, and the pedestrian The image 51 is an “image of the recognition object” and “ The frame image 53 corresponds to the “enhanced image” described in the claims, and the pedestrian corresponds to the “recognition object” and the “assumed object” described in the claims. The night view device 100 corresponds to the “vehicle display device” recited in the claims.

(Second embodiment)
The second embodiment of the present invention is a modification of the first embodiment. In the second embodiment, the first identification unit 25 performs identification processing based on the first learning model 266 instead of the first learning model 66 used in the first embodiment. Hereinafter, the first learning model 266 used in the second embodiment and the identification processing of the first identification unit 25 based on the first learning model 266 will be described in detail based on FIGS. 7, 1, and 2.

  In creating the first learning model 266, the learning model generation device 80 includes a positive example sample image 61a associated with a teacher signal indicating a positive example and a negative example sample image 63a associated with a teacher signal indicating a negative example. Are input (see FIG. 7B). More specifically, the first learning model 266 is an image group obtained by excluding a positive example sample image 61b in which a pedestrian similar to a non-pedestrian is captured from a plurality of positive example sample images 61 included in the first sample image group 60. Learning as a positive example. In addition, the first learning model 266 learns, as a negative example, an image group obtained by excluding a negative example sample image 63b in which a non-pedestrian similar to a pedestrian is shown from a plurality of negative example sample images 63.

  As described above, the difference between an image showing a pedestrian similar to a non-pedestrian and an image showing a non-pedestrian similar to a pedestrian is ambiguous. Therefore, by excluding the positive sample image 61b showing a pedestrian similar to a non-pedestrian and the negative sample image 63b showing a non-pedestrian similar to a pedestrian from the first sample image group 60, a first learning model is obtained. 266 does not learn these ambiguous differences in detail. Therefore, the first learning model 266 can be a simple learning model that contributes to speeding up the processing by the first identification unit 25.

  In the second embodiment, the first identification unit 25 may include the pedestrian image 51 in each cut-out image 56 depending on whether or not it is related to the positive example learned by the first learning model 266. Identifies whether or not there is. Since the negative sample image 63b of a non-pedestrian similar to a pedestrian is excluded from the negative examples, an image of an object similar to a pedestrian is included in addition to the cut-out image 56 including the pedestrian image 51. The cut image 56 to be displayed may correspond to the cut image 56 related to the positive example. Therefore, the 1st identification part 25 identifies the cutout image 56 relevant to the positive example of the 1st learning model 266, and can extract the cutout image 56 which may contain the pedestrian image 51 at high speed. It can narrow down accurately.

  And from the cut-out image 56 narrowed down by the 1st identification part 25, the 2nd identification part 27 includes the image of the object similar to a pedestrian by identifying whether the pedestrian image 51 is contained. The cut image 56 is excluded. As described above, the cut image 56 including the pedestrian image 51 is identified with high accuracy by the identification processing of the first identification unit 25 and the second identification unit 27. Therefore, even when the first learning model 266 according to the second embodiment is used, the image recognition circuit 20 reliably maintains high accuracy for recognizing the pedestrian image 51 from the front image 50 while speeding up the processing for recognition. can do.

(Third embodiment)
The third embodiment of the present invention is another modification of the first embodiment. In the third embodiment, the first identification unit 25 performs the identification process based on the first learning model 366 instead of the first learning model 66 used in the first embodiment. Hereinafter, the first learning model 366 used in the third embodiment and the identification processing of the first identification unit 25 based on the first learning model 366 will be described in detail based on FIGS. 8, 1, and 2.

  When creating the first learning model 366, only a plurality of positive example sample images 61 in the first sample image group 60 are input to the learning model generation device 80 in association with teacher signals indicating positive examples. (See FIG. 8 (b)). The first learning model 366 learns only the positive example sample image 61 and does not learn the negative example sample image 63. Thus, the 1st learning model 366 constructed | assembled based on the positive example sample image 61 is not learning in detail the ambiguous difference between a pedestrian and a non-pedestrian. Therefore, the first learning model 366 can be surely a simple learning model that contributes to speeding up the processing by the first identification unit 25.

  The first identification unit 25 of the third embodiment may include the pedestrian image 51 in each cut-out image 56 depending on whether or not it is related to the positive example learned by the first learning model 366. Or not. Since the first learning model 366 has not learned the negative example image 63 in which the non-pedestrian is photographed as a negative example, in addition to the cut-out image 56 including the pedestrian image 51, an image of an object similar to the pedestrian is obtained. The included cutout image 56 may also correspond to the cutout image 56 related to the positive example. Therefore, the 1st identification part 25 identifies the cutout image 56 relevant to the positive example of the 1st learning model 366, and can extract the cutout image 56 which may contain the pedestrian image 51 at high speed. It can narrow down accurately.

  And from the cut-out image 56 narrowed down by the 1st identification part 25, the 2nd identification part 27 includes the image of the object similar to a pedestrian by identifying whether the pedestrian image 51 is contained. The cut image 56 is excluded. As described above, the cut image 56 including the pedestrian image 51 is identified with high accuracy by the identification processing of the first identification unit 25 and the second identification unit 27. Therefore, even when the first learning model 366 according to the third embodiment is used, the image recognition circuit 20 reliably maintains high accuracy for recognizing the pedestrian image 51 from the front image 50 while speeding up the process for recognition. can do.

(Other embodiments)
Although a plurality of embodiments according to the present invention have been described above, the present invention is not construed as being limited to the above embodiments, and can be applied to various embodiments and combinations without departing from the gist of the present invention. can do.

  In the first embodiment, the first learning model 66 is constructed as a simple learning model by learning, as a positive example, a negative example sample image 63b showing a non-pedestrian similar to a pedestrian. In the second embodiment, the first learning model 266 includes a positive sample image 61b showing a pedestrian similar to a non-pedestrian and a negative sample image 63b showing a non-pedestrian similar to a pedestrian as a first sample. By excluding from the image group 60, it was constructed as a simple learning model. However, as long as it can be constructed as a simple learning model, the method for constructing the first learning model is not limited to the method of the above embodiment. For example, the first learning model learns a positive example sample image 61b showing a pedestrian similar to a non-pedestrian among the plurality of positive example sample images 61 as a negative example together with the plurality of negative example sample images 63. May be. Such a first learning model does not learn in detail an ambiguous difference between a pedestrian and a non-pedestrian, and thus can be a simple learning model that contributes to speeding up the processing by the first identification unit 25. . When such a first learning model is used, the first identification unit sets the cut-out image 56 that may include the pedestrian image 51 by setting a threshold value for identifying a positive example low. It can narrow down accurately.

  In the above embodiment, the Haar-like feature amount is used when the learning model generation device 80 learns the feature of the pedestrian image 51 and when the image recognition circuit 20 recognizes the pedestrian image 51. However, a feature value other than the Haar-like feature value may be used as a feature value for image recognition. For example, the luminance distribution of the positive sample image or the wavelet coefficient having a high appearance frequency when the positive sample image is wavelet transformed may be used as a feature amount for learning and identification. Alternatively, a Histograms of Oriented Gradients (HOG) feature value, a Joint HOG feature value, a Joint Haar-like feature value, etc. for calculating a luminance gradient between adjacent pixels are used as a feature value for learning and identification. Also good.

  In the above embodiment, AdaBoost is used for learning the Haar-like feature value by the learning model generation device 80. However, the feature amount learning method is not limited to AdaBoost. For example, when learning the luminance distribution and the wavelet coefficient as the feature amount, the learning model generation apparatus 80 can use a neural network, SVM, or the like as a learning method. Further, when learning the HOG feature value, the Joint HOG feature value, and the Joint Haar-like feature value as the feature value, the learning model generation apparatus 80 can use Real AdaBoost or the like as a learning method.

  In the above embodiment, when the first learning model 66 and the second learning model 77 are constructed, the Haar-like feature amount is used. However, the feature quantity used when constructing the first learning model may be different from the feature quantity used when constructing the second learning model. For example, for the construction of the first learning model for narrowing down the cut-out image 56, Haar-like feature quantities that are relatively easy to simplify the learning model are used. On the other hand, the construction of the second learning model 77 for discriminating between pedestrians and non-pedestrians uses a Joint HOG feature that can be expected to have high discrimination accuracy although the learning model is complicated. By constructing the first learning model and the second learning model as described above, the process of recognizing the pedestrian image from the front image can be further increased in speed and accuracy. It should be noted that the feature amount when the learning model is constructed must correspond to the feature amount used by the identification unit when performing the identification process based on the learning model. Therefore, when the 1st learning model is constructed | assembled using Haar-like feature-value, a 1st discrimination | determination part implements identification processing based on the Haar-like feature-value of a cut-out image. Further, when the second learning model is constructed using the Joint HOG feature value, the second identification unit performs the identification process based on the Joint HOG feature value of the clipped image. As described above, if the feature amount when the learning model is constructed and the feature amount at the time of the identification processing by each identifying unit correspond to each other, the image recognition circuit determines various feature amounts for image recognition. Can be used.

  In the above embodiment, the near-infrared camera 10 is used as an imaging unit for capturing the front image 50. However, as long as the front image 50 can be captured, the imaging means is not limited to the near-infrared camera 10. The imaging means may be, for example, a visible light camera that detects visible light and generates an image, or a far-infrared camera that detects far infrared and generates an image. Moreover, the night view apparatus 100 of the said embodiment was the structure which acquires the front image 50 from the near-infrared camera 10. FIG. However, the night view apparatus may include an imaging unit.

  In the above embodiment, the image recognition circuit 20 and the drawing circuit 30 are each configured as a circuit including a processor and the like. However, the night view control circuit may have a functional block corresponding to the image recognition circuit 20 and the drawing circuit 30 by executing a program. In addition, each functional block 21, 23, 25, 27, 29 included in the image recognition circuit 20 may be configured as an independent circuit. Furthermore, an analog circuit not depending on the execution of the program may be formed in the night view control circuit as a configuration corresponding to each functional block.

  In the above embodiment, the night view apparatus 100 displays the front image 50 on which the frame image 53 is superimposed on the liquid crystal display 40. However, the display means of the night view device is not limited to the liquid crystal display 40. For example, the head-up display device that projects the front image 50 on which the frame image 53 is superimposed on the windscreen of the vehicle may be the display means of the night view device. Alternatively, the display unit of the night view device may be an output unit that outputs a signal of the front image 50 toward, for example, a liquid crystal display included in the navigation system.

  In the embodiment described above, the night view apparatus 100 acquires the front image 50 from the near-infrared camera 10 that captures the front area. However, the night view apparatus may recognize an image of a recognition target object such as a pedestrian from an image of a rear region or a side region of the vehicle, not limited to the front region of the vehicle. Moreover, in the said embodiment, the image recognition circuit 20 recognized the pedestrian as a recognition target object. However, the recognition target recognized by the image recognition circuit 20 is not limited to a pedestrian. For example, a person who rides a vehicle, a bicycle or a two-wheeled vehicle, an animal such as a deer or a spider, and the like may be set in advance as recognition objects.

  Furthermore, the present invention can be applied not only to the vehicle night view device 100 but also to all image recognition devices for recognizing recognition objects. For example, the image recognition apparatus may be used as a part of a monitoring apparatus that recognizes a person as a recognition target object by being connected to a monitoring camera installed in an elevator hall or a store in a condominium.

DESCRIPTION OF SYMBOLS 10 Near-infrared camera (imaging means), 11 Near-infrared projector, 15 Night view control circuit, 20 Image recognition circuit (image recognition apparatus), 21 Image acquisition part (image acquisition means), 23 Cutout part (cutout means), 25 First identification unit (first identification unit), 27 Second identification unit (second identification unit), 29 Recognition unit (recognition unit), 30 Drawing circuit (display unit), 40 Liquid crystal display (display unit), 50 Front Image (captured image), 51 Pedestrian image (image of recognition object), 53 frame image (emphasized image), 55 cut window, 56 cut image, 60 first sample image group, 70 second sample image group, 61, 61a, 61b, 71, 71a, 71b Positive example sample image, 63, 63a, 63b, 73, 73a, 73b Negative example sample image, 66, 266, 366 First learning model 77 Second learning model, 80 learning model generating device 100 night view apparatus (vehicle display device)

Claims (7)

Image acquisition means for acquiring a captured image generated by the imaging means;
From the captured image acquired by the image acquisition means, a cutting means for cutting out a cut image of a preset size in order,
For the plurality of cut-out images cut out by the cut-out means, an image representing a recognition object set in advance based on a first learning model created from a first sample image group serving as a learning reference is included. First identifying means for identifying whether there is a possibility of being
A second learning model created from a second sample image group serving as a learning reference for the clipped image that has been identified by the first identification means as being likely to contain the image of the recognition object. Based on the second identification means for identifying whether or not the image of the recognition object is included,
Recognizing means for recognizing a position in the captured image of the clipped image identified as including the image of the recognition object by the second identification means as a position where the recognition object appears;
An image recognition apparatus comprising: The first sample image group includes a plurality of positive example sample images in which an assumed object assumed as the recognition object is captured, and a plurality of negative sample images in which a non-assumed object other than the assumed object is captured,
The first learning model learns, as a positive example, the negative example sample image in which the non-imaginary object similar to the assumed object is captured among the plurality of negative example sample images together with the plurality of positive example sample images. Created by
Depending on whether the first identification means corresponds to the positive example learned by the first learning model, there is a possibility that an image of the recognition object may be included in each cut-out image The image recognition apparatus according to claim 1, wherein: The first sample image group includes a plurality of positive example sample images in which an assumed object assumed as the recognition object is captured, and a plurality of negative sample images in which a non-assumed object other than the assumed object is captured,
The first learning model learns, as a positive example, an image group obtained by excluding the positive example sample image in which the assumption object similar to the non-expected object is captured from the plurality of positive example sample images, and the plurality of negative example images. It is created by learning from the example sample image as a negative example an image group excluding the negative example sample image in which the non-imaginary object similar to the assumed object is reflected,
Whether the first identification means may include an image of the recognition object in each of the cut-out images depending on whether the first example is related to the positive example learned by the first learning model. The image recognition apparatus according to claim 1, wherein: The first sample image group includes a plurality of positive example sample images in which an assumed object assumed as the recognition object is captured, and a plurality of negative sample images in which a non-assumed object other than the assumed object is captured,
The first learning model learns the plurality of positive sample images as positive examples in the first sample image group,
Whether the first identification means may include an image of the recognition object in each of the cut-out images depending on whether the first example is related to the positive example learned by the first learning model. The image recognition apparatus according to claim 1, wherein: The second sample image group includes a plurality of positive sample images in which an assumed object assumed as the recognition object is captured, and a plurality of negative sample images in which a non-expected object other than the assumed object is captured,
The second learning model is created by learning the plurality of positive example sample images as positive examples and learning the negative example sample images as negative examples,
The second identifying means identifies whether or not the image of the recognition object is included in the cut-out image depending on whether or not it corresponds to the positive example learned in the second learning model. The image recognition apparatus according to any one of claims 1 to 4, wherein the image recognition apparatus is characterized. The first identifying means identifies whether or not there is a possibility that an image of a pedestrian that is the recognition object is included in the plurality of cut-out images,
The second identification means determines whether or not the pedestrian image is included in the clipped image identified by the first identification means as being likely to include the pedestrian image. The image recognition apparatus according to claim 1, wherein the image recognition apparatus is identified. From the captured image generated by the imaging unit that captures the surrounding area of the vehicle, the captured image is an enhanced image that recognizes the image of the recognition object existing in the surrounding area and emphasizes the image of the recognition object. A vehicle display device that superimposes and displays
The image recognition apparatus according to any one of claims 1 to 6, wherein the captured image is acquired from the imaging unit.
Display means for superimposing and displaying the emphasized image at a position where the recognition object recognized by the image recognition device appears in the captured image;
A vehicle display device comprising: