CN106095830A - A kind of image geo-positioning system based on convolutional neural networks and method - Google Patents

Abstract

The invention provides a kind of network information identification system based on convolutional neural networks and method, relate to field of satellite location, it is characterized in that, described system includes: video upload module, picture upload module, picture frame extraction module, GEOGRAPHICAL INDICATION detection module, decomposing module, characteristic extracting module, activation primitive, pond module, full link block, grader analyze judge module, study module, area data storehouse, pixel data storehouse and result display module.This system and method has that image analysis is fast, image recognition accurately, possess the positioning function to video information and possess learning capacity all advantages.

Description

A kind of image geo-positioning system based on convolutional neural networks and method

Technical field

The present invention relates to field of satellite location, particularly relate to a kind of image of based on convolutional neural networks geo-location system System and method.

Background technology

For framing, if the scenery in pushing to is well-known sight spot or landmark, then we can pass through Special scene labeling just can very clear be known.Photo for common place positions, and everybody is generally by photo The method adding GEOGRAPHICAL INDICATION indicates.A lot of Android phone are taking pictures when such as now, activate in camera laggard Entering to shoot menu and just have " geographical position " function, we open this function before having only to shoot.

So photo of shooting will add the geographical location information of upper locality, automatically when we check this on computers During a little photo, it is switched to " details ", under GPS project wherein it is seen that photograph taking actual geographic position, this In use GPS longitude and latitude be marked, cellular base station, Wi-Fi etc. certainly can also be used to be marked.Certainly digital camera etc. Also having similar positioning function, so by adding GEOGRAPHICAL INDICATION in photo Exif information, we just can be the most easily for shining Sheet positions.But do not has geography information in actual photographed or online treated a lot of photos, then for this The location of a little photos just seems trouble particularly.

Existing image location system is primarily present following defect:

1, image recognition accuracy is inadequate: in existing image location system, is mostly directly to be stored in data base Substantial amounts of sample, and different images has various different feature.So directly sample being compared, then can cause comparison Effect is the poorest, and accuracy rate is the lowest.

2, do not possess learning capacity: existing framing lacks learning capacity in actual use, regardless of Use which kind of algorithm and analyze determination methods, always causing framing that deviation occurs, if can not be the most continuous Improve in study, whole image location system can be caused to stagnate.

3, image analysis is slower: existing image location system, the image analysis algorithm of employing the most more tradition, it then follows often Picture breakdown, deep analysis isotype.The speed causing a pictures location is very slow, and due to the office of algorithm Sex-limited, cause the result the most usually inaccuracy of location.

4: do not possess the location to video information: video information is not the most carried out fixed by existing image location system The function of position and means.

Summary of the invention

For the defect of above-mentioned anti-plug-in technical method, the invention provides a kind of image based on convolutional neural networks ground Reason alignment system and method, this system and method has that image analysis is fast, image recognition accurately, possess the location to video information Function and possess learning capacity all advantages.

The technical solution used in the present invention is as follows:

A kind of image geo-positioning system based on convolutional neural networks, it is characterised in that described system includes: on video Carry module, picture upload module, picture frame extraction module, GEOGRAPHICAL INDICATION detection module, decomposing module, characteristic extracting module, swash Function alive, pond module, full link block, grader analyze judge module, study module, area data storehouse, pixel data storehouse With result display module；

Described video upload module signal is connected to picture frame extraction module；Picture frame extraction module signal respectively is connected to Video upload module and GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module signal respectively is connected to picture frame and extracts mould Block, analysis judge module, picture upload module, decomposing module and characteristic extracting module；Described characteristic extracting module, respectively signal It is connected to GEOGRAPHICAL INDICATION detection module, activation primitive and pond module；Described pond module signal respectively is connected to feature extraction Module, activation primitive and full link block；Described full link block signal respectively is connected to grader and pond module；Described point Class device signal respectively is connected to full link block and analyzes judge module；Described analysis judge module signal respectively is connected to decompose Module, GEOGRAPHICAL INDICATION detection module, grader, result display module and study module；Described study module signal respectively connects In analyzing judge module, pixel data storehouse and area data storehouse.

Described video upload module, for uploaded videos information, sends video information to picture frame extraction module；Described Picture upload module, for uploading pictures information, sends pictorial information to GEOGRAPHICAL INDICATION detection module；Described picture frame extracts Module, for video information being decoded, intercepts the GEOGRAPHICAL INDICATION of video and obtains picture frame complete in video information, will GEOGRAPHICAL INDICATION and picture frame send to GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module, for judging that uploads regards Frequently whether information and pictorial information there is GEOGRAPHICAL INDICATION, if comprising GEOGRAPHICAL INDICATION, directly sending extremely to analyze by pictorial information and sentencing Disconnected module, then sends picture to decomposing module and characteristic extracting module without comprising GEOGRAPHICAL INDICATION.

Described decomposing module, the Pixel Information for the pictorial information received is resolved into Pixel Information, after decomposing Send to analyzing judge module；Described characteristic extracting module, for the picture received is carried out feature extraction, the spy extracted Reference breath sends to activation primitive；Described activation primitive, for being attached characteristic extracting module and pond module；Described pond Change module, for the characteristic information received is processed, reduce the characteristic vector of characteristic extracting module output, improve simultaneously Result；Described full link block, for being attached grader and final characteristic information；Described grader is for according to spy Reference breath carries out classification process.

Described pixel data storehouse internal memory contains sampled pixel information；Described sampled pixel information is comprised GPS ground by 1.5 hundred million The pixel composition that the picture resource of reason positional information decomposes；

In described area data storehouse, the data of storage are small area labelling；Described small area is labeled as: will filter out Population density superlatively district, 30000, the whole world in each regional decompasition become 30000 square areas not of uniform size, and pin These square areas are added the data message after not isolabeling.

The learning method of described study module comprises the following steps:

Step 1: by the small area labelling in GPS geographical location information corresponding for sampled pixel information and area data storehouse Compare；

Step 2: the sampled pixel labelling being mutually matched and small area labelling are associated；

Step 3: for the sampled pixel information of all small areas labelling in the data base of area cannot be mated, create new Small area labelling carries out interrelated.

Described analysis judge module, for judging the geographical position of this picture or video capture according to the information received； The method that described analysis judge module analysis judges comprises the steps:

Step 1: if receiving the GEOGRAPHICAL INDICATION that GEOGRAPHICAL INDICATION detection module directly transmits, then directly according to geography Label information judges the shooting ground of picture or video；

Step 2: if receiving the Pixel Information that decomposing module sends over, then send this Pixel Information to learning mould Block, study module, according to the correlation analysis in pixel data storehouse and area data storehouse, show that the pixel of this Pixel Information is geographical Position；This geographical position is stored temporarily；

Step 3: the same transmission of image feature information that grader is sended over to study module, study module according to Correlation analysis in pixel data storehouse and area data storehouse, draws the convolution geographical position of this image feature information；

Step 4: compared in this convolution geographical position and pixel geographical position, if there is difference, then by pixel ground Reason position and convolution geographical position show to result display module simultaneously as result transmission；

Step 5: user may determine that in result display module which result is accurately, feeds back to result learn mould Block；

Step 6: study module can adjust in pixel data storehouse and area data storehouse according to the correct result of feedback before Relatedness.

The sorting technique of described grader comprises the following steps:

Step 1: set hypothesis function as:

$h_{\mathrm{\θ}}\left(x\right)=\frac{1}{1+\exp (-{\mathrm{\θ}}^{T}x)}$

Step 2: draw cost function

$j\left(\mathrm{\θ}\right)=-\frac{1}{n}\[\underset{i=1}{\overset{n}{\Σ}}{y}^{\left(i\right)}{\mathrm{logh}}_{\mathrm{\θ}}\left(x^{(i+1)}\right)+(1-y^{(i+1)})log(1-h_{\mathrm{\θ}}(x^{(i+1)}\left)\right)\]$

Step 3: use and assume that function calculates probit for the classification of each characteristic information and is:

$\frac{1}{{\mathrm{\Σ}}_{j=1}^k{e}^{{\mathrm{\θ}}_j^T}{e}^{x^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\θ}}_1^T}{e}^{x^{\left(i\right)}}\\ e^{{\mathrm{\θ}}_2^T}{e}^{x^{\left(i\right)}}\\ ...\end{array}\right]$

Step 4: carry out classification process according to the probability calculated.

Use technique scheme, present invention produces following beneficial effect:

1, image recognition accuracy is high: employing convolutional neural networks is as image recognition, simultaneously also by traditional image Reason technology is applied to wherein, carries out bi-directional matching, can maximize accuracy and the precision promoting image recognition.

2, possess learning capacity: in the process to output result, user can be evaluated for result each time, According to evaluation result, whole system can carry out learning and improving, upper once run into analogue when promote location accuracy And precision.

3, image analysis is fast: for having the image information of GEOGRAPHICAL INDICATION, can directly carry out localization process, improve not Necessary processing procedure.Simultaneously for there is no the picture of GEOGRAPHICAL INDICATION, use convolutional neural networks to process, compare traditional images Speed is more accelerated.

4: possess the location to video information: outside Chu Ci, it is also possible to video information is carried out frame extraction, enters for image Row goes out the shooting ground of video information here after analyzing.

Accompanying drawing explanation

Fig. 1 is a kind of image geo-positioning system based on convolutional neural networks and method in the embodiment of the present invention.

Detailed description of the invention

All features disclosed in this specification, or disclosed all methods or during step, except mutually exclusive Feature and/or step beyond, all can combine by any way.

Any feature disclosed in this specification (including any accessory claim, summary), unless specifically stated otherwise, By other equivalences or there is the alternative features of similar purpose replaced.I.e., unless specifically stated otherwise, each feature is a series of An example in equivalence or similar characteristics.

The embodiment of the present invention 1 provides a kind of image geo-positioning system based on convolutional neural networks, system structure As shown in Figure 1:

A kind of intelligent medical supersonic image processing equipment, it is characterised in that described equipment includes: image capture device, Image receiving apparatus, image sorting device, image general purpose processing device, the first algorithm data-base, analysis judgement equipment, display set Standby, ultrasonoscopy processing equipment and the second data base；

Described image capture device signal is connected to image receiving apparatus；Described image receiving apparatus signal is connected to image Sorting device；Described image sorting device signal is connected to image general purpose processing device；Described image general purpose processing device signal It is connected to ultrasonoscopy processing equipment, image sorting device, the first algorithm data-base and display device；Described first algorithm data Storehouse signal respectively is connected to image general purpose processing device and analyzes judgement equipment；Described analysis judgement equipment signal respectively is connected to Display device, the first algorithm data-base and the second algorithm data-base；Described ultrasonoscopy processing equipment signal respectively is connected to figure As general purpose processing device, display device and the second algorithm data-base；Described second algorithm data-base signal respectively is connected to ultrasonic Image processing equipment and analysis judgement equipment.

Described ultrasonoscopy processing equipment, the image after processing general image processing equipment is carried out at ultrasonoscopy Managing, it includes: ultrasound image processor and signal are connected to the ultrasonoscopy processing system of processor；Described image general procedure Equipment, for processing all images sended over, if the result after ultrasonoscopy then will process sends to super Acoustic image processing equipment proceeds to process, if general image, then the image transmission after processing is carried out to display device Showing, it includes: general image processor and signal are connected to the common image processing systems of processor.

Described image capture device includes: acquiring ultrasound image equipment and general image collecting device；Described image-receptive Equipment, for receiving the image information sended over from image capture device, sends image information to image sorting device； Described image sorting device, for classifying image information according to different image capture devices, sends classification results To image general purpose processing device.

Described display device, for showing the figure after processing general image processing equipment and the process of ultrasonoscopy processing equipment Picture；Described analysis judgement equipment, for artificially judging image that display device shows the most accurately, according to judged result, to the In one algorithm data-base and the second algorithm data-base, the algorithm priority of storage is adjusted.

The embodiment of the present invention 2 provides a kind of image geo-positioning system based on convolutional neural networks, system structure Scheme as shown in Figure 1:

A kind of image geo-positioning system based on convolutional neural networks, it is characterised in that described system includes: on video Carry module, picture upload module, picture frame extraction module, GEOGRAPHICAL INDICATION detection module, decomposing module, characteristic extracting module, swash Function alive, pond module, full link block, grader analyze judge module, study module, area data storehouse, pixel data storehouse With result display module；

Described video upload module signal is connected to picture frame extraction module；Picture frame extraction module signal respectively is connected to Video upload module and GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module signal respectively is connected to picture frame and extracts mould Block, analysis judge module, picture upload module, decomposing module and characteristic extracting module；Described characteristic extracting module, respectively signal It is connected to GEOGRAPHICAL INDICATION detection module, activation primitive and pond module；Described pond module signal respectively is connected to feature extraction Module, activation primitive and full link block；Described full link block signal respectively is connected to grader and pond module；Described point Class device signal respectively is connected to full link block and analyzes judge module；Described analysis judge module signal respectively is connected to decompose Module, GEOGRAPHICAL INDICATION detection module, grader, result display module and study module；Described study module signal respectively connects In analyzing judge module, pixel data storehouse and area data storehouse.

Described video upload module, for uploaded videos information, sends video information to picture frame extraction module；Described Picture upload module, for uploading pictures information, sends pictorial information to GEOGRAPHICAL INDICATION detection module；Described picture frame extracts Module, for video information being decoded, intercepts the GEOGRAPHICAL INDICATION of video and obtains picture frame complete in video information, will GEOGRAPHICAL INDICATION and picture frame send to GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module, for judging that uploads regards Frequently whether information and pictorial information there is GEOGRAPHICAL INDICATION, if comprising GEOGRAPHICAL INDICATION, directly sending extremely to analyze by pictorial information and sentencing Disconnected module, then sends picture to decomposing module and characteristic extracting module without comprising GEOGRAPHICAL INDICATION.

Described decomposing module, the Pixel Information for the pictorial information received is resolved into Pixel Information, after decomposing Send to analyzing judge module；Described characteristic extracting module, for the picture received is carried out feature extraction, the spy extracted Reference breath sends to activation primitive；Described activation primitive, for being attached characteristic extracting module and pond module；Described pond Change module, for the characteristic information received is processed, reduce the characteristic vector of characteristic extracting module output, improve simultaneously Result；Described full link block, for being attached grader and final characteristic information；Described grader is for according to spy Reference breath carries out classification process.

Described pixel data storehouse internal memory contains sampled pixel information；Described sampled pixel information is comprised GPS ground by 1.5 hundred million The pixel composition that the picture resource of reason positional information decomposes；

In described area data storehouse, the data of storage are small area labelling；Described small area is labeled as: will filter out Population density superlatively district, 30000, the whole world in each regional decompasition become 30000 square areas not of uniform size, and pin These square areas are added the data message after not isolabeling.

The learning method of described study module comprises the following steps:

Step 1: by the small area labelling in GPS geographical location information corresponding for sampled pixel information and area data storehouse Compare；

Step 2: the sampled pixel labelling being mutually matched and small area labelling are associated；

Step 3: for the sampled pixel information of all small areas labelling in the data base of area cannot be mated, create new Small area labelling carries out interrelated.

Described analysis judge module, for judging the geographical position of this picture or video capture according to the information received； The method that described analysis judge module analysis judges comprises the steps:

Step 1: if receiving the GEOGRAPHICAL INDICATION that GEOGRAPHICAL INDICATION detection module directly transmits, then directly according to geography Label information judges the shooting ground of picture or video；

Step 2: if receiving the Pixel Information that decomposing module sends over, then send this Pixel Information to learning mould Block, study module, according to the correlation analysis in pixel data storehouse and area data storehouse, show that the pixel of this Pixel Information is geographical Position；This geographical position is stored temporarily；

Step 3: the same transmission of image feature information that grader is sended over to study module, study module according to Correlation analysis in pixel data storehouse and area data storehouse, draws the convolution geographical position of this image feature information；

Step 4: compared in this convolution geographical position and pixel geographical position, if there is difference, then by pixel ground Reason position and convolution geographical position show to result display module simultaneously as result transmission；

Step 5: user may determine that in result display module which result is accurately, feeds back to result learn mould Block；

Step 6: study module can adjust in pixel data storehouse and area data storehouse according to the correct result of feedback before Relatedness.

The sorting technique of described grader comprises the following steps:

Step 1: set hypothesis function as:

$h_{\mathrm{\θ}}\left(x\right)=\frac{1}{1+\exp (-{\mathrm{\θ}}^{T}x)}$

Step 2: draw cost function

$j\left(\mathrm{\θ}\right)=-\frac{1}{n}\[\underset{i=1}{\overset{n}{\Σ}}{y}^{\left(i\right)}{\mathrm{logh}}_{\mathrm{\θ}}\left(x^{(i+1)}\right)+(1-y^{(i+1)})log(1-h_{\mathrm{\θ}}(x^{(i+1)}\left)\right)\]$

Step 3: use and assume that function calculates probit for the classification of each characteristic information and is:

$\frac{1}{{\mathrm{\Σ}}_{j=1}^k{e}^{{\mathrm{\θ}}_j^T}{e}^{x^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\θ}}_1^T}{e}^{x^{\left(i\right)}}\\ e^{{\mathrm{\θ}}_2^T}{e}^{x^{\left(i\right)}}\\ ...\end{array}\right]$

Step 4: carry out classification process according to the probability calculated.

The embodiment of the present invention 3 provides a kind of image geo-positioning system based on convolutional neural networks, system structure Scheme as shown in Figure 1:

A kind of image geo-positioning system based on convolutional neural networks, it is characterised in that described system includes: on video Carry module, picture upload module, picture frame extraction module, GEOGRAPHICAL INDICATION detection module, decomposing module, characteristic extracting module, swash Function alive, pond module, full link block, grader analyze judge module, study module, area data storehouse, pixel data storehouse With result display module；

Described video upload module signal is connected to picture frame extraction module；Picture frame extraction module signal respectively is connected to Video upload module and GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module signal respectively is connected to picture frame and extracts mould Block, analysis judge module, picture upload module, decomposing module and characteristic extracting module；Described characteristic extracting module, respectively signal It is connected to GEOGRAPHICAL INDICATION detection module, activation primitive and pond module；Described pond module signal respectively is connected to feature extraction Module, activation primitive and full link block；Described full link block signal respectively is connected to grader and pond module；Described point Class device signal respectively is connected to full link block and analyzes judge module；Described analysis judge module signal respectively is connected to decompose Module, GEOGRAPHICAL INDICATION detection module, grader, result display module and study module；Described study module signal respectively connects In analyzing judge module, pixel data storehouse and area data storehouse.

Described video upload module, for uploaded videos information, sends video information to picture frame extraction module；Described Picture upload module, for uploading pictures information, sends pictorial information to GEOGRAPHICAL INDICATION detection module；Described picture frame extracts Module, for video information being decoded, intercepts the GEOGRAPHICAL INDICATION of video and obtains picture frame complete in video information, will GEOGRAPHICAL INDICATION and picture frame send to GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module, for judging that uploads regards Frequently whether information and pictorial information there is GEOGRAPHICAL INDICATION, if comprising GEOGRAPHICAL INDICATION, directly sending extremely to analyze by pictorial information and sentencing Disconnected module, then sends picture to decomposing module and characteristic extracting module without comprising GEOGRAPHICAL INDICATION.

Described decomposing module, the Pixel Information for the pictorial information received is resolved into Pixel Information, after decomposing Send to analyzing judge module；Described characteristic extracting module, for the picture received is carried out feature extraction, the spy extracted Reference breath sends to activation primitive；Described activation primitive, for being attached characteristic extracting module and pond module；Described pond Change module, for the characteristic information received is processed, reduce the characteristic vector of characteristic extracting module output, improve simultaneously Result；Described full link block, for being attached grader and final characteristic information；Described grader is for according to spy Reference breath carries out classification process.

Described pixel data storehouse internal memory contains sampled pixel information；Described sampled pixel information is comprised GPS ground by 1.5 hundred million The pixel composition that the picture resource of reason positional information decomposes；

In described area data storehouse, the data of storage are small area labelling；Described small area is labeled as: will filter out Population density superlatively district, 30000, the whole world in each regional decompasition become 30000 square areas not of uniform size, and pin These square areas are added the data message after not isolabeling.

The learning method of described study module comprises the following steps:

Step 1: by the small area labelling in GPS geographical location information corresponding for sampled pixel information and area data storehouse Compare；

Step 2: the sampled pixel labelling being mutually matched and small area labelling are associated；

Step 3: for the sampled pixel information of all small areas labelling in the data base of area cannot be mated, create new Small area labelling carries out interrelated.

Described analysis judge module, for judging the geographical position of this picture or video capture according to the information received； The method that described analysis judge module analysis judges comprises the steps:

Step 1: if receiving the GEOGRAPHICAL INDICATION that GEOGRAPHICAL INDICATION detection module directly transmits, then directly according to geography Label information judges the shooting ground of picture or video；

Step 2: if receiving the Pixel Information that decomposing module sends over, then send this Pixel Information to learning mould Block, study module, according to the correlation analysis in pixel data storehouse and area data storehouse, show that the pixel of this Pixel Information is geographical Position；This geographical position is stored temporarily；

Step 3: the same transmission of image feature information that grader is sended over to study module, study module according to Correlation analysis in pixel data storehouse and area data storehouse, draws the convolution geographical position of this image feature information；

Step 4: compared in this convolution geographical position and pixel geographical position, if there is difference, then by pixel ground Reason position and convolution geographical position show to result display module simultaneously as result transmission；

Step 5: user may determine that in result display module which result is accurately, feeds back to result learn mould Block；

Step 6: study module can adjust in pixel data storehouse and area data storehouse according to the correct result of feedback before Relatedness.

The sorting technique of described grader comprises the following steps:

Step 1: set hypothesis function as:

$h_{\mathrm{\θ}}\left(x\right)=\frac{1}{1+\exp (-{\mathrm{\θ}}^{T}x)}$

Step 2: draw cost function

$j\left(\mathrm{\θ}\right)=-\frac{1}{n}\[\underset{i=1}{\overset{n}{\Σ}}{y}^{\left(i\right)}{\mathrm{logh}}_{\mathrm{\θ}}\left(x^{(i+1)}\right)+(1-y^{(i+1)})log(1-h_{\mathrm{\θ}}(x^{(i+1)}\left)\right)\]$

Step 3: use and assume that function calculates probit for the classification of each characteristic information and is:

$\frac{1}{{\mathrm{\Σ}}_{j=1}^k{e}^{{\mathrm{\θ}}_j^T}{e}^{x^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\θ}}_1^T}{e}^{x^{\left(i\right)}}\\ e^{{\mathrm{\θ}}_2^T}{e}^{x^{\left(i\right)}}\\ ...\end{array}\right]$

Use convolutional neural networks as image recognition, also be applied to wherein, enter by traditional image processing techniques simultaneously Row bi-directional matching, can maximize accuracy and the precision promoting image recognition.

In the process to output result, user can be evaluated for result each time, according to evaluation result, whole Individual system can carry out learning and improving, upper once run into analogue when promote location accuracy and precision.

For having the image information of GEOGRAPHICAL INDICATION, can directly carry out localization process, improve unnecessary process Journey.Simultaneously for there is no the picture of GEOGRAPHICAL INDICATION, use convolutional neural networks to process, more accelerate than traditional images speed.

In addition to this it is possible to video information is carried out frame extraction, after being analyzed, go out video information for image here Shooting ground.

The invention is not limited in aforesaid detailed description of the invention.The present invention expands to any disclose in this manual New feature or any new combination, and the arbitrary new method that discloses or the step of process or any new combination.

Claims (7)

1. an image geo-positioning system based on convolutional neural networks, it is characterised in that described system includes: video upload Module, picture upload module, picture frame extraction module, GEOGRAPHICAL INDICATION detection module, decomposing module, characteristic extracting module, activation Function, pond module, full link block, grader analyze judge module, study module, area data storehouse, pixel data storehouse and Result display module；

Described video upload module signal is connected to picture frame extraction module；Picture frame extraction module signal respectively is connected to video Upload module and GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module signal respectively be connected to picture frame extraction module, Analyze judge module, picture upload module, decomposing module and characteristic extracting module；Described characteristic extracting module, signal is even respectively It is connected to GEOGRAPHICAL INDICATION detection module, activation primitive and pond module；Described pond module signal respectively is connected to feature extraction mould Block, activation primitive and full link block；Described full link block signal respectively is connected to grader and pond module；Described classification Device signal respectively is connected to full link block and analyzes judge module；Described analysis judge module signal respectively is connected to decompose mould Block, GEOGRAPHICAL INDICATION detection module, grader, result display module and study module；Described study module signal respectively is connected to Analyze judge module, pixel data storehouse and area data storehouse.

2. image geo-positioning system based on convolutional neural networks as claimed in claim 1, it is characterised in that described video Upload module, for uploaded videos information, sends video information to picture frame extraction module；Described picture upload module, uses In uploading pictures information, pictorial information is sent to GEOGRAPHICAL INDICATION detection module；Described picture frame extraction module, for by video Information is decoded, and intercepts the GEOGRAPHICAL INDICATION of video and obtains picture frame complete in video information, by GEOGRAPHICAL INDICATION and image Frame sends to GEOGRAPHICAL INDICATION detection module；Described GEOGRAPHICAL INDICATION detection module, for judging video information and the picture letter uploaded Whether breath there is GEOGRAPHICAL INDICATION, if comprising GEOGRAPHICAL INDICATION, directly pictorial information being sent to analyzing judge module, if do not had Have comprise GEOGRAPHICAL INDICATION then by picture send to decomposing module and characteristic extracting module.

3. image geo-positioning system based on convolutional neural networks as claimed in claim 2, it is characterised in that described decomposition Module, for the pictorial information received is resolved into Pixel Information, the Pixel Information after decomposing sends extremely to analyze and judges mould Block；Described characteristic extracting module, for the picture received is carried out feature extraction, the characteristic information extracted sends to activating Function；Described activation primitive, for being attached characteristic extracting module and pond module；Described pond module, is used for docking The characteristic information received processes, and reduces the characteristic vector of characteristic extracting module output, improves result simultaneously；Described full connection Module, for being attached grader and final characteristic information；Described grader is for classifying according to characteristic information Process.

4. image geo-positioning system based on convolutional neural networks as claimed in claim 3, it is characterised in that described pixel Databases contains sampled pixel information；Described sampled pixel information is by 1.5 hundred million pictures comprising GPS geographical location information The pixel composition of decomposing resources；

In described area data storehouse, the data of storage are small area labelling；Described small area is labeled as: complete by filter out Each regional decompasition in 30000 population density superlatively districts of ball becomes 30000 square areas not of uniform size, and for this The data message after not isolabeling is added in a little square areas.

5. image geo-positioning system based on convolutional neural networks as claimed in claim 3, it is characterised in that described study The learning method of module comprises the following steps:

Step 1: the small area labelling in GPS geographical location information corresponding for sampled pixel information and area data storehouse is carried out Comparison；

Step 2: the sampled pixel labelling being mutually matched and small area labelling are associated；

Step 3: for the sampled pixel information of all small areas labellings in the data base of area cannot be mated, create new small Area labelling carries out interrelated.

6. image geo-positioning system based on convolutional neural networks as claimed in claim 5, it is characterised in that described analysis Judge module, for judging the geographical position of this picture or video capture according to the information received；Described analysis judge module Analyze the method judged to comprise the steps:

Step 1: if receiving the GEOGRAPHICAL INDICATION that GEOGRAPHICAL INDICATION detection module directly transmits, then directly according to GEOGRAPHICAL INDICATION Information judges the shooting ground of picture or video；

Step 2: if receiving the Pixel Information that decomposing module sends over, then send this Pixel Information to study module, Study module, according to the correlation analysis in pixel data storehouse and area data storehouse, draws the pixel geography position of this Pixel Information Put；This geographical position is stored temporarily；

Step 3: the same transmission of the image feature information that sended over by grader is to study module, and study module is according to pixel Correlation analysis in data base and area data storehouse, draws the convolution geographical position of this image feature information；

Step 4: compared in this convolution geographical position and pixel geographical position, if there is difference, then by pixel geography position Put and show to result display module simultaneously as result transmission with convolution geographical position；

Step 5: user may determine that in result display module which result is accurately, feeds back to study module by result；

Step 6: study module can adjust the pass in pixel data storehouse and area data storehouse according to the correct result of feedback before Connection property.

7. the image geo-positioning system based on convolutional neural networks as described in claim 3 or 4 or 5, it is characterised in that institute The sorting technique stating grader comprises the following steps:

Step 1: set hypothesis function as:

$h_{\mathrm{\θ}}\left(x\right)=\frac{1}{1+\exp (-{\mathrm{\θ}}^{T}x)}$

Step 2: draw cost function

$j\left(\mathrm{\θ}\right)=-\frac{1}{n}\[\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}^{\left(i\right)}\log h_{\mathrm{\θ}}\left(x^{(i+1)}\right)+(1-y^{(i+1)})log(1-h_{\mathrm{\θ}}(x^{(i+1)}\left)\right)\]$

Step 3: use and assume that function calculates probit for the classification of each characteristic information and is:

$\frac{1}{{\mathrm{\Σ}}_{j=1}^k{e}^{{\mathrm{\θ}}_j^T}{e}^{x^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\θ}}_1^T}{e}^{x^{\left(i\right)}}\\ e^{{\mathrm{\θ}}_2^T}{e}^{x^{\left(i\right)}}\\ \mathrm{...}\end{array}\right]$

Step 4: carry out classification process according to the probability calculated.