JPS63284974A - Picture compression system - Google Patents

Abstract

PURPOSE:To maximize the code reduction effect of variable length coding totally by selecting a code block of the variable length coding based on the degree of concentration of the probability distribution of picture information or the fixed length coding. CONSTITUTION:A difference circuit 1 subtracts an inter-line prediction value of a prediction circuit 2 from a picture data of a present picture and stores a linear difference into a line buffer 3. A variable sampling density coding circuit 4 applies the variable sampling density coding to obtain the 2nd-order difference. A code distribution calculation circuit 27 obtains the probability distribution of the quantized code, obtains the value of the code probability distribution function and the dispersion of the probability density function, and a fixed/variable length code block selection circuit 28 selects the code block of the variable length coding or the code block of the fixed length coding. A code processed by a fixed length/variable length coding circuit 29 is processed by a zero code compression circuit 5 and a code edition circuit 26 and sent.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field 1] The present invention relates to an image compression method for image monitoring in which image information from a TV camera or the like is transmitted via a telephone line or the like by narrowband transmission. [Background technology] The telephone line currently in use is generally an analog telephone line in the voice band (0.3 to 3.4 KHz), which is a subscriber telephone line.
However, in the case of image monitoring and security systems, as the number of objects to be monitored increases, a low-cost image compression method with a high compression rate is required. Generally, an ITV camera or the like is used as a TV camera for image monitoring, and therefore still images are often obtained using multi-valued image signals. In addition, vector quantization and motion compensation used in TV conference systems and TV'i episodes on 64K bit digital lines (Japanese Patent Application Laid-Open No. 58-101581, Japanese Patent Application Laid-Open No. 58-13
3088) and the like are expensive and difficult to use due to cost constraints. On the other hand, still image transmission methods such as facsimiles are multilevel,
Development of colorization is currently underway, and 64Kbi of l5DN
1 bit/pet (pixel) for t times am color image
International standardization of the compression method is being promoted by ISO and CCITT, and vector quantization (Reference: Dentsu M'??
Technical Report IE84-18 p. 9), Block Coding (References: Electronic Communication T Complete Journal '87/I Vol.
l70-B No. 1 p. 66, Journal of the Institute of Image Electronics Engineers, No. 1
5 Temperature Sensing No. 4 p, 225), sequential reproduction encoding (References:
Journal of the Institute of Electronics and Communication Engineers '87/I Voi, 170-B
No. 1 p. 105), cosine transformation (Reference: Journal of the Institute of Electronics and Communication Engineers '86/10 Vol. 169-B No.
, 10), etc. are being considered. Of these, block encoding and sequential playback encoding are said to be methods with a small amount of calculation, and are hierarchically encoded in order to sequentially change the block size and sampling interval to produce a clear image, and each layer It is compressed and variable-length encoded. The present inventors developed variable sample density coding (IEICE Journal '75/
2 Vol, 58-^No2. Applying p. 97),
We are developing a predictive coding method for second-order differences and a variable pixel compression transmission method combined with zero code compression, as described in Japanese Patent Laid-Open No. 61-296866. Since this conventional method is applied as part of the configuration of the present invention described later, it will be described in detail here. 9 and 10 show a configuration for performing variable sample density compression/expansion of inter-line prediction residuals or DPCM compression/expansion of inter-line prediction residuals, and the image compression unit A shown in FIG. The difference circuit 1 calculates the inter-line prediction residual using the sample image data (pixels of the current line) and the prediction data from the prediction circuit 2 (the restored pixels of the previous line that has been compressed and encoded), and calculates the inter-line prediction residual. The difference is passed through a line buffer 73 to a variable sample density encoding circuit 4, which compresses the quadratic side residual into a quantization characteristic code in a direction perpendicular to the inter-line prediction residual, that is, in the line direction. The quantization characteristics are the combination of the amplitude @ difference value (quantization level) and the time difference value (sample interval) of the secondary side residual as shown in Table 1, and the corresponding 4-bit code. This is a mutual conversion table. As shown in FIG. 11, the points plotted on the coordinate axes of the amplitude difference value and the time difference value are quantization points, which are connected by straight lines in the figure. In Table 1, 15 quantization points are taken as shown in FIG. 12, and these quantization points are represented by quantization characteristic codes encoded into 4 bits. These quantized characteristic codes are combined into two 8-bit codes that are modulated and transmitted by a modem, but there are consecutive zero codes indicating that the secondary side residual of the quantized characteristic code is zero. Therefore, the zero code compression circuit 5 compresses the zero code (
Due to the compression characteristics (zero line), when there are consecutive zero codes, the transmission code (Fl to FE) indicating the loss of zero codes or the number of lines containing only zero codes is converted into a 4-bit quantization characteristic code other than the zero code ( In the case of a change code), it is compressed into a zero compression code with two 8-bit transmission codes (0 to EF) and transmitted. This zero code zero line compression characteristic depends on the number of zero codes and 8
This is a table of mutual conversion with bit code. The quantization characteristic code is decoded by the variable sample density decoding circuit 6, taken in to the interpolation synthesis circuit 8 via the line buffer 7, and synthesized with the value of the line buffer 9 of the feedback loop inside the prediction circuit 2, The signal is input to the prediction circuit 2 for prediction of the next line. Here, when the time difference values (sample intervals) are thinned out at regular intervals, the interpolation synthesis circuit 8
When each pixel is thinned out to 64 x 64 pixels, 256
This is for restoring to an image of ×256 pixels.
Even when only the change area is thinned out every 4 pixels and sent to a background image of 256 x 256 pixels, only the change area is interpolated to restore a coarse image and synthesized with the background image of 256 x 256 pixels. In the reception decompression unit Bfm, the code data is decoded into a quantized characteristic code by the zero code decoding circuit 10 shown in figure m10, and this quantized characteristic code is converted to the variable sample density decoding circuit 1.
1 restores the inter-line prediction residual, takes this inter-line prediction residual into the interpolation synthesis circuit 13 via the line buffer T12, and converts one line of image data into a decoding loop made up of the prediction circuit 15 and line buffer 14. Restore. Here, if all the time difference values of the above quantization characteristics are set to 1, variable sample density encoding becomes the same as DPCM, so the variable sample density encoding method for inter-line prediction residuals is equivalent to DPCM for inter-line prediction residuals. become. In the DCPM of the inter-line prediction residuals or the variable sample density coding method (variable sample density predictive coding method) of the inter-line prediction residuals described above, the total per pixel is W, and most of the amount is D.
Since it uses quantization similar to PCM, the quantization characteristics can be fixed, and it can be said to be simpler than other compression methods. Furthermore, since the prediction method uses orthogonal second-order differences between lines and within lines, the prediction residuals are concentrated to zero compared to the first-order difference prediction method or the planar prediction method, making it possible to reduce the quantization error and reduce the amplitude. Even if the dynamic range of the difference value is small, in practice,
It has the characteristic that overload distortion is not noticeable. In addition, since the second-order differences are concentrated at zero, the effect of zero-code compression becomes greater, and since the prediction residual is calculated for each line with the sample (current image), errors do not propagate in the direction between lines. , which is unique to variable sample density coding, has the characteristics of less edge business and better image quality. Furthermore, the effect of zero code compression is limited to cases where zero codes are consecutive, so the effect is small when compressing the entire image.
It can be said that the effect becomes greater when compressing and encoding only changed pixels, which will be described later. The quantization characteristics of variable sample density encoding are as shown in Figure 11, by making the quantization points overlap in the range of quantization values with different time difference values (sample intervals). By encoding into a code that corresponds to the combination with (interval), the width of the quantization value is expanded in areas with large time difference values, improving tracking of contours and edges, and reducing edge business. be. Now 1jS1
When quantizing with a variable sample density using the quantization characteristics shown in Figures 1 and 12, when the amplitude difference value is less than or equal to a constant value αpn, samples are taken at a position where a constant interval Tpn + is added. The time difference value is increased, and in the case of 0 repetitions, the prediction residual signal for the predicted value by the restored pixel before the time difference value Στpn+1 is quantized into an amplitude difference value. When the time difference value is 2, the quantized amplitude difference value is the constant value ap1. It is encoded into a quantization characteristic code determined by a combination of a time difference value and an amplitude difference value (also called a quantization level) so that a value larger or smaller than that can be selected. (Table 1) M quantification characteristics (Table 2) Zero code compression Now, in variable sample density coding, the compression rate is achieved at quantization points with large time difference values, but in the conventional method described above, other block coding It is difficult to keep the time difference value within a small range of about 1 to 2 and increase it to a large value of 3 or more in order to reduce the deterioration of image quality, as in the case of or cosine transformation. Furthermore, by selecting multiple quantization characteristics optimally depending on the image and adapting the quantization characteristics to optimally set the time difference value, it is expected that image quality will improve and the compression rate will increase. The method of determining quantization characteristics from the dispersion and deviation of the current image does not allow accurate determination of second-order differences. Similarly, it is not possible to accurately determine the M child conversion characteristic to be used next from pixels that have already been transmitted or restored in the case of second-order difference quantization that is processed line by line. Therefore, the present inventors proposed a compressed transmission method for monitoring images when a change has already been detected (Patent Application No. 60-600012).
This method transmits the image information of the part that has changed between frames based on the interframe prediction residual between the 7 frames of the previous image or the 7 frames of the predicted image and the current image frame. shows the circuit configuration of an image compression unit using this method. In this configuration, first, a change detection circuit 19 compares the current image frame imported from the 7-frame memory 16 into the current image 7-frame buffer γ7a 17 and the reference image 7-frame in the reference image frame buffer 18 to detect a change.
If a change is detected, the change pixel detection circuit 20 detects the current image 7.
Large interframe residual error (absolute value of difference) between the current image frame 7 of frame x 7a 17 and the predicted image (previous image) frame of frame x 7a 7a 21 is divided into large change pixels and small zero pixels. Furthermore, if necessary, by propagating and degenerating the changed pixels, one or two zero pixels sandwiched between the changed pixels are changed to changed pixels, and one or two changed pixels sandwiched between zero pixels are changed to zero pixels. Then, a zero pixel isolated among changed pixels and a changed pixel isolated among zero pixels are removed. The second-order difference molecular measurement encoding circuit 22 includes a quantization characteristic selection circuit 23
The difference between the 7 frames or the developed image 7 is calculated, and the value of the pixel that changes between the frames is calculated using the inter-line and intra-line second-order differential molecular measurement quantization method (second-order DCPM, or differential variable sampling density code) as shown in FIG. It is used to compress changing areas or changing pixels efficiently. As shown in Table 2, when moving from a zero pixel to a changed pixel, the zero code compression circuit 24 transmits the value of the changed pixel using a changed pixel code (FD, xx) as an initial value for prediction, and converts the value of the changed pixel to a zero from the changed pixel. When moving to a pixel, a zero pixel code (for example, FE) is transmitted as the end value of prediction, and the number of zero codes with a prediction residual of zero and the number of lines where all pixels are zero codes are encoded and compressed. Transmit. Also, by setting the value of the changed pixel in the 7th frame of the current image to a value other than zero, changed pixel compression can be performed without using changed pixel codes or zero pixel codes, but in this case change ii! Even if the value of I■ element is zero, it is necessary to add +α. In methods for processing images in blocks, there are many known examples of combinations of cosine transformation and various compression methods such as DPCM and vector quantization. Cosine transformation is used to first quantize and restore the coefficients of low-frequency components with strong correlation in the image, and then quantize the high-frequency components, which is transmitted hierarchically, making it difficult for block boundaries to appear.
Good image quality with SNH of 35 dB or more can be obtained with a compression rate of PEL. However, since the frequency distribution of the quantization codes themselves concentrates on codes where the amplitude difference value is zero, the amount of codes can be reduced in multilevel images by variable length coding such as Huffman coding. The frequency distribution of codes changes depending on the target image, so a Huffman code that is optimal for one image may have the opposite effect on another image.Furthermore, in Huffman encoding, the codebook of variable length codes is set optimally. It is necessary to calculate for each pair an image. This increases the cost by increasing the total number of W and 1N, so for example (Kato et al., 'Adaptive orthogonal return coding method using multiple variable length code sets' PSCS-864
, 6, 1986), images are divided into classes and a codebook of representative variable-length codes is prepared for each class in advance, and it is judged based on criteria which class the image to be coded belongs to. There are some that allow you to select a codebook. In addition, cosine transformation has a problem in that microcomputers have a large number of W and S, and variable pixel compression (zero code compression) cannot be used effectively. Here, in order to reduce deterioration in image quality and to perform quantization within the range of changing pixels, it is not possible to increase the time difference value of the quantization characteristic. In addition, for multilevel images, the amount of code can be reduced by variable length coding called Huffman coding, but in variable sampling density coding using quantization characteristics with small time difference values, the frequency distribution of the occurrence of the code itself is such that the amplitude difference value is zero. Since the current method concentrates on the code of In other words, in the conventional method, the frequency distribution of the code changes depending on the target image, so in No.17man coding, the codebook of variable length codes is
There is a problem that the amount of calculation required to set fi increases and that zero-code compression for compressing changed pixels cannot be utilized. The code tJ circuit 26 in FIG. 13 is the code Hegg generating circuit 2.
The number of transmission pixels, compression parameters, and full-screen/
A pufferfish is added to the compressed code every 7 frames as a code to indicate selection of a changed part, and the data is transmitted as communication data. The compression parameters mentioned here include the quantization characteristics and the prediction coefficients of the interframe load prediction circuit, and if the prediction coefficients are optimal, the prediction residuals will concentrate to zero, the quantization error will become small, and the image quality will improve. [Object of the Invention] The present invention has been developed in view of the above-mentioned points, and its purpose is to provide an image compression method that efficiently transmits high-quality images by reducing the amount of code. [Disclosure of the Invention 1] The present invention is an image compression method that divides an image into a plurality of blocks and performs quantization encoding on each block. The code is determined, and based on the quantization code, a statistic representing the degree of concentration of the probability distribution of image information in each block position, or a statistic representing the degree of concentration of the probability distribution of the quantization code is determined, and this statistic and The feature is that a variable length encoding codebook or fixed length encoding is selected by comparing the magnitude with a certain setting value. The present invention will be explained below with reference to Examples. K19 Figure 1 shows the basic circuit configuration of the image compression section A of this embodiment, and FIG. 2 shows the basic circuit configuration of the expansion section B of this embodiment. Structures with the same number perform the same operation. First, compression encoding will be explained based on the configuration of FIG. The image data used here is obtained by separating the video signal of the surveillance TV camera into horizontal and vertical synchronization signals and image signals, and then YC-separating this image signal into a luminance signal (Y) and a color difference signal (C). , each number 46 with an A/D converter (not shown).
~8-bit digital value (converted and written to the current image 7 frame memory, in this example, 1i[
The slope signal (Y) of the lij image is 256×256 pixels, 8 bits (256 floors, 7 m)/pixel. Further, the image data of the embodiment is obtained by dividing an image into a plurality of blocks and reading out each block. The difference circuit 1 subtracts the line-to-line predicted value of the prediction circuit 2 from the image data of the current image and sends the first difference to the line buffer 73.
Write the line. That is, the line buffer 3 stores the first-order difference, which is the prediction residual between lines, between the pixels of the current line to be compressed and encoded next, and the predicted value of the next line based on the restored pixels of the previous line that has been compressed and encoded. is memorized. This first difference is variable sample density encoded in the line direction by the variable sample density encoder 4, and then the next first difference is predicted based on the first difference of the pixel to be compression encoded and the compression encoded restored pixel. Find the second difference, which is the predicted residual between the two
A quantization code corresponding to a combination of a time difference value (sample interval) and an amplitude difference value (quantization level) is determined from the order difference using quantization characteristics. This quantization code has a fixed length (4
b i t ), and the case where the quantization characteristic selection circuit selects from a plurality of quantization characteristics as in the embodiments shown in FIGS. 3 to PIS 8, which will be described later, will also be considered. Here, the method of the present invention is not limited to the quantization of second-order differences, but also includes a method in which only first-order differences are used without taking the line-to-line differences. In other words, the configuration and effects of the present invention apply regardless of whether it is a first-order difference or a second-order difference. Note that the following description of the embodiment will be based on second-order differences. In this embodiment, the code distribution calculation circuit 27 calculates the frequency of occurrence of quantized codes for each quantization characteristic in a preselected standard image, and the correspondence relationship between quantized codes and variable-length code numbers for this standard image. is calculated, and a conversion table between the quantization code, the variable length code number, and the variable length code is recorded in the codebook of the variable length code made up of memory. A plurality of variable length codes are prepared for each quantization characteristic, such as a code for an image that has few contour components and is easy to compress, and a code for an image that has many contour components and is difficult to compress. Examples of variable length codes are shown in Tables 3 to 6. As a method for creating a variable length code, in addition to performing Huffman encoding on a standard image, priority may be given to efficiency of encoding and decoding. Then, the code distribution calculation circuit 27 calculates the probability distribution of the quantized code, calculates the value of the code probability distribution function, the variance of the probability density function, etc., and compares the statistical value with a certain set value. A fixed/variable length codebook selection circuit 28 selects a variable length encoding codebook or a fixed length encoding codebook. Specifically, if the ratio of codes with a code length of 4 bits or less in Table 3 is greater than a certain setting value, variable length encoding is performed in Table 3, and if it is smaller, the code length is 4 bits in Table 4.
It is determined whether the ratio of codes less than or equal to it is larger than a certain setting value, and if it is larger, variable length coding is performed according to Table 4, and if it is smaller, fixed length coding is performed. In Table 3, there are 5 codes shorter than 4 bits, and in Table 4, there are 7 codes. The codes in Table 3 have a high 5 concentration, and when there are many zero codes in changing pixel compression, the probability of occurrence of two 2-bit codes is high, resulting in large compression, and compression is performed with five codes shorter than 4 bits. The degree of zero code is stronger than other codes with 5 bits or more, so it has a compression effect, and the codes in Table 4 have a low 5 concentration degree, and when there are few zero codes due to change pixel compression, 2 bit
This is suitable when it is advantageous to have more 3-bit codes than the t code. Similarly, in the tIS5 table, there are 6 codes shorter than 4 bits, and in the tIS5 table, there are 7 codes. The codes in Table 5 have a high 5 concentration, and are suitable for cases where there are few extension codes such as zero code compression, or for the quantization characteristics of 16-level amplitude difference values, and the codes in the tlS6 table have a 5 concentration. In cases where the degree of compression is low and there are few extension codes such as zero code compression instead of change pixel compression, or
Let us go back to the quantization characteristics of six levels of amplitude difference values. By selecting these codebooks, variable-length encoding is performed for each block of the image, eliminating the need for an increase in the amount of code compared to fixed-length encoding, and since the codebook is predetermined, it is possible to perform variable-length encoding for each block of the image. There is no need to send
Overall, variable length encoding can be performed efficiently and the code length can be shortened. Now, the code encoded by the fixed length/variable length encoding circuit 29 is further processed by the zero code compression circuit 26 to convert it to g? The code is compressed, and the code Ili collection circuit 26 collects m sets according to the selection result of the fixed/variable length codebook selection circuit 28 and transmits them. This transmitted code data is transmitted to the code distribution circuit 3 of the decompression section B.
1 for zero code and fixed/variable length codebook selection circuit 32
The code distributed to selected codes and decoded by the zero code decoding circuit 10 is decoded by the fixed length/variable length decoding circuit 33 based on the selected code, further decoded by the variable sample density decoding circuit 11, Image data is reproduced in the same manner as in the tIS10 diagram configuration. Figures 3 and 4 show the image compression section A and decompression section B of this embodiment. The image compression section A of this embodiment divides the image into i-defective blocks and processes each block separately. It is a statistic that represents the degree of concentration of the probability distribution of image information. For example, as a difference signal distribution, 2
The 5th degree of concentration is determined by the 5th degree of concentration calculating circuit 34 based on the small degree of spread of the distribution such as the deviation of this second degree difference, and the 5th degree of concentration is determined to have a fine gradation when the 5th degree of concentration is stronger than a certain value. The quantization characteristic is selected from a weak one to a coarse quantization characteristic by the function of the quantization characteristic selection circuit 35, and based on this selection, a quantization code is obtained by the variable sample density encoding circuit 4, and this quantization code is established. A code distribution calculation circuit 27 calculates a statistic that indicates the degree of concentration of the distribution, a minimum comparison is made between this statistic and a certain set value, and a fixed/variable length codebook selection circuit 28 selects a value with a concentration of 5. For example, tpJ is stronger.
3. Select the variable length encoding in Table 5, and set the weakest one to the fourth one.
Select the variable length encoding shown in Table 6, and select the weaker fixed length encoding. Therefore, in this embodiment, variable length encoding can be performed simultaneously with quantization encoding. The decompression unit B includes a code distribution circuit 31 and a quantization characteristic selection circuit 3.
The selection food of quantization characteristics for decoding of No. 6 is distributed to correspond to the selection of quantization characteristics of the image compression section A. In this embodiment, the 5-degree intermediate is used to select between variable-length encoding and fixed-length encoding, which simplifies the process. In this case, the quantization characteristics and amplitude difference values are selected for each line, and the quantization characteristics, amplitude difference values, and variable length encoding,
The fixed-length encoding selection code is encoded and transmitted only for lines that have been changed. In addition, since the above five concentrations are based on the first-order difference (:) and the second-order difference for each IR line in the line direction, the quantization characteristics and amplitude difference values can be determined more accurately than based on the current image or restored image. The selection of coding and fixed-length coding becomes more accurate, which has the effect of improving image quality and increasing the compression rate.In addition, when determining the code based on the standard image, as a result of the 5 concentration calculation, variable-length coding is When selected, the extension code of the zero code compression code is added to the Huffman code by using the frequency and probability of occurrence of extended code F in Table 1 as the frequency and probability of occurrence of the zero code compression code in Table 2 after zero code compression. By assigning variable length coding, it is possible to perform variable length coding that is optimal for the statistical properties of each image, but due to this Huffman coding (variable length coding), the frequency of occurrence of quantization codes in variable sample density coding is At the same time as calculating the probability distribution and probability of zero codes in the quantization code, the concentration on zero codes (proportion of short code lengths) is calculated. Perform variable length encoding of the codebook according to the concentration level,
By performing fixed-length encoding when the deviation is large, it is possible to prevent the average code length from becoming longer than the fixed-length code. However, including FO and FF in Table 2, the code length of the Huffman code with the highest probability of zero code occurrence is determined to be 2biL, and Fl has 5 code zeros, F2 has 6 code zeros, etc.
・F9 means 12 code zero. Huffman encoding requires a large amount of calculation and may not be suitable for microcomputer processing, but it is recommended that high-speed DSPs dedicated to image processing
etc., it is possible to achieve the effect of reducing the amount of code. However, in the case of changed pixel compression as in Example 3, which will be described later, the proportion of changed pixels whose change is larger than a certain value in the transmitted image as determined by the changed pixel detection circuit is small.
Since there is a large proportion of zero pixels whose change is smaller than a certain value and is zero, if we do not assign an extension code of the zero-code compression code in the Huffman code and perform equi-code compression after /)7man encoding. The code length of the extended code becomes long and becomes ineffective. Furthermore, when performing Huffman encoding for each image, as a result of Huff7 encoding, the code length of the zero code becomes 3 bits or 4 bits.
As a result, the code length of the zero code becomes large, and the efficiency of zero code compression may deteriorate. Even in such a case, calculate the Huffman code for the entire image or large block, and select variable-length coding for lines whose concentration is higher than a certain value in the above 5, and select fixed-length coding for lines with weak concentration. Therefore, if there is a possibility that the effect of variable-length coding may be low, by using fixed-length coding, you can select between variable-length coding and fixed-length coding on a line-by-line basis, and the overall amount of code can be reduced. can. If the code length of the quantization code instead of the Huffman code is combined with variable-length encoding so that the code length becomes shorter in order of the frequency of occurrence, it is not optimal for each image, but it is more suitable for microcomputer processing. Easy encoding is possible. This means that a specific code in the variable-length code hood is converted into a zero-code code, so that the part of the amplitude difference value where the zero codes of zeros are equal to or greater than a certain value is compressed and encoded into a zero-code corresponding to the number of consecutive zero-codes. This corresponds to variable-length zero code compression encoding, which is characterized by assignment. Furthermore, in order to predetermine the variable-length zero code compression code in accordance with the quantization characteristics regardless of the target image, a standard image is compressed and a codebook is created based on the frequency and probability of occurrence of the quantization code. By transmitting selection codes for variable-length encoding and fixed-length encoding on a line-by-line basis according to the above 5 concentration levels, it is possible to avoid transmitting a correspondence table between variable-length codes and quantization codes on a line-by-line basis, and to The amount of code can be reduced. In this case, in order to apply Huffman encoding, zero code compression is performed in advance, the probability of occurrence of a zero code code is determined, and then code assignment is performed together with other quantized codes. In Table 3.4, when assigning quantization codes in order of frequency of occurrence, the zero code is set to 1111. In this case, although it is not necessarily optimal like Huffman encoding, it is not much different from Huffman encoding and is suitable for microcomputer processing. This zero code code is 4 bits, and this makes it possible to add the effects of zero code compression regarding secondary differences and change pixel compression to variable length encoding. The zero code code 1111 corresponds to the code F (hexadecimal number) in Table 2, and the code following 1111 < 4 bits follows F (corresponds to 4 bits. Since the zero code code is 2 bits, the second
The meanings of the zero signs in the table are: Fl is 5 sign zeros, F2 is 6 sign zeros,
..., F8 uses 12 code zeros and modified ones for variable length coding, and arranges zero codes up to four zero codes. Note that the tJS7 table is an example of the frequency distribution of codes for quadratic difference DPCM encoding and variable length zero code compression encoding. Table 8 shows an example of variable sampling density encoding with quantization characteristics for a time difference value of 2. In the case of variable sampling density encoding, the code length is variable in order of decreasing code length from the most frequently occurring code for simplicity. The correspondence between the sample density code and the variable length code in Table 3 is
Target image and quantization This embodiment shown in FIGS. 5 and 6 uses image data of 7 frames of the current image to be compressed and encoded,
Predicted image 7 based on the restored previous image or current image read out from the 7 frame patch 7737 through the prediction circuit 38
The difference circuit 39 calculates the inter-frame difference with respect to the frame, and converts this inter-frame difference into second-order difference molecular measurement quantization (second-order difference DPCM, or difference variable sampling) as in the second embodiment shown in FIGS. Density encoding) and variable length zero code compression encoding, and prediction based on the current image 7 frames to be compressed and encoded next and the restored previous image or current 11ji image stored in the 7 frame cross 7T37. A changed pixel detection circuit 40 determines a changed pixel whose inter-frame difference with an image frame is larger than a certain threshold value, and a changed pixel of the current image is detected as a secondary difference numerator? II
I quantization (secondary differential DPCM or differential variable sample density encoding) is performed and variable length zero code compression encoding is performed by the variable length encoding circuit 29' and the zero code compression circuit 5. 41 in FIG. 5 is an interpolation circuit, 33 in FIG. 6 is a variable length decoding circuit for decoding variable length codes, 42 is an interpolation synthesis circuit, 43 is a 7 frame x 7y, 44 is a prediction circuit, These circuits 42 to 44 restore the current image frame. X11-1 This embodiment consists of an image compression section shown in FIG. ff17 and an expansion section B shown in FIG. A unit 45 divides the blocks into blocks, each block is treated as a unit pixel of a unit block, the average value of each block is separated as a unit pixel value by an average value separation unit 46, and a coarse image compression unit 47 converts it into a quantization code. This code is then restored by a coarse image restoring section 48. Next, a difference circuit 49 obtains a frame-to-frame error signal between the original image and this coarse picture, and this inter-frame error signal is divided into a plurality of blocks. , a hierarchical image compression method is adopted in which the difference signal between frames is quantized and encoded by the 5 concentration degree circuit 40, the quantization characteristic selection circuit 35, and the error image compression unit 51 for each block. ? g%
A statistic that represents the degree of concentration of the probability distribution of difference signal information (e.g., the variance of the difference signal distribution, etc.) or a statistic that represents the degree of concentration of the probability distribution of the quantization code (e.g., the value of the probability distribution function, the variance of the probability density function, etc.) etc.) is determined by the sign distribution calculation circuit 27,
In this block unit, a fine gradation quantization characteristic and a time difference value are selected for those with a strong concentration, and a coarse gradation quantization characteristic and an amplitude difference value are selected for a weak concentration, and quantization encoding is performed. , the fixed/variable length food book selection circuit 28 selects a codebook for variable length encoding by comparing this statistic or a statistic representing the degree of concentration of the probability distribution of the quantized code (e.g., variance) with a certain set value. Or select fixed length encoding and use fixed/variable length encoding circuit 2! Adaptive variable-length coding is now performed in the Since the frame-to-frame error signal has a small correlation between pixels and a large high-frequency component, it is thought that there are many blocks that will have the opposite effect if variable-length encoding is applied, but such blocks can be fixed-length encoded and other Foodbook variable-length coding, which is highly effective, can be applied to blocks, reducing the amount of code as a whole. Various methods can be considered for quantizing and encoding a coarse image, but as described above, in this embodiment, blocks are divided hierarchically by the block dividing unit 45, and for the L-order block, each block one step below is treated as a unit pixel. Then, the average value for each block is separated by the average value separation unit 46, the separated average value is used as the value of the unit pixel, and the unit pixel of the highest block is first quantized and encoded by the coarse image compression unit 47, Furthermore, a coarse image is restored from this code by a coarse image restoring unit 48, and then a frame difference error code between the original image and this coarse image is obtained by a difference circuit 49, and this frame difference signal is sent to a lower block. The frame is subdivided, and the frame error signal is quantized and encoded for each of the lower 10 blocks. In this case, since the rough image is simple, the amount of code is small, but there is a risk that the boundaries between blocks will be visible. Note that instead of the block division and average value separation in Figure 7, the entire image is first quantized and encoded with a coarse sample interval, or quantized and encoded with a quantization characteristic that has a large range of time difference values, and then the coarse image is generated. Then, find the frame error signal between the original image and this coarse image, and calculate this frame error signal! If a hierarchical image compression method of sequential sharpening is adopted in which blocks are subdivided into blocks with one loss and the interframe error signal is quantized and encoded for each block, the configuration will be the same as that of Embodiment 3, and the average of the mean value separation will be Calculations are no longer required, rough images can be obtained more easily, and in particular, the correspondence between the amplitude difference value and the time difference value of the sampling interval, quantum, and negative characteristics is brought closer to a monotonically decreasing function where the amplitude difference value x the time difference value is a constant value. As a result, the block boundaries of the coarse image do not appear, and the variance of the inter-frame error signal can be reduced. As a result, the effect of variable-length coding is increased, the amount of code can be reduced, and the image quality can be improved. In addition, the low frequency components of the image are quantized and encoded by averaging the entire image or by discrete cosine transformation in block units, and a coarse image is restored from this, and then the frames between the original original image and this coarse image are Embodiment 3 employs a hierarchical image compression method of sequential sharpening in which an inter-frame error signal is obtained, this inter-frame error signal is subdivided into a plurality of blocks, and the inter-frame error signal is quantized and encoded for each lower block. Similar effects can be obtained by applying this configuration, but the amount of calculation for averaging or discrete cosine transformation increases. In addition, 50 in FIG. 7 is an interpolation that adds the error image obtained by restoring the quantization code of the error image by the error image restoring section 52 and the coarse image obtained by restoring the above-mentioned coarse image restoring section 48. The difference circuit 39 calculates the difference between the image interpolated and synthesized by the interpolation synthesis circuit 50 and the image data of the current image. Further, 53 in the fjS8 figure is a circuit for separating code data into a selection code of the quantization characteristic selection circuit 36, a selection code of the fixed/variable length food book selection circuit 32, and a compression code. 54 is a coarse image restoration unit for restoring a coarse image from the code decoded by the fixed/variable length decoding circuit 33; In the error image restoring section that restores the image, the restored images obtained by the original parts 54 and 55 are added together in the interpolation and synthesis circuit 13, and the interpolated and synthesized image is further added to the current image in the changed pixel synthesis block synthesis section 56. (3rd
Table) (Table 4) (Table 5) (Table 6) (Table 7) (Table 8) [Advantageous effect 1 of the invention As described above, in the predictive quantization method, the present invention ?, which represents the degree of concentration of probability distribution of information?
1" amount, or a statistic that represents the degree of concentration of the probability distribution of quantization codes, and then selects a codebook for variable-length encoding or fixed-length encoding by comparing this statistic with a certain setting value. -, the code reduction effect of variable length coding can be maximized overall by selecting the codebook for variable length coding optimally for each block, and by using fixed length coding for blocks that have the opposite effect. As with the quantization code of second-order difference DPCM or second-order difference variable sample density coding, for the first-order difference between lines, the second-order difference is obtained in the line direction, and this second-order difference is quantized. In general, compared to the first-order difference, the degree of spread of the distribution such as the deviation/dispersion of the second-order difference is smaller, or the five-concentration degree of the amplitude difference value of the quantization code is stronger, so the overall effect of variable-length coding is The effect is large.Also, for blocks where the zero concentration of the amplitude difference value of this quantization code is stronger than a certain value, select those classified by five concentrations from the food book of variable length coding and By selecting fixed-length encoding, lines that would have a long average code length with variable-length codes can be converted into fixed-length codes, thereby reducing the amount of code for the entire image. By selecting fine gradation quantization characteristics and time difference values, and selecting coarse gradation quantization characteristics and amplitude difference values for weaker ones, the quantization characteristics are more accurate than with the current image or restored image. It is possible to adaptively determine the amplitude and amplitude difference values, and for this purpose, it is possible to accurately select variable-length coding food book or fixed-length coding, and in combination with the adaptation of quantization characteristics and amplitude difference values, it is possible to In addition, in order to compress consecutive zero codes of the quantization code by zero codes, a specific code of the variable length code can be added to the short code length part of the variable length code. By assigning it to the code, when using a quantization characteristic with a small time difference value, it is possible to use a fixed-length code for a line where the average code length is longer than a variable-length code for changing pixels, resulting in good image quality even in changing pixel compression. It is also possible to further increase the effect of increasing the ratio.

[Brief explanation of the drawing]

FIG. 1 is a circuit configuration diagram of an image compression section according to Embodiment 1 of the present invention;
Fig. 2 is a circuit diagram of the decompression unit same as above, Fig. 3 is a circuit diagram of the image compression unit of Embodiment 2 of the present invention, Pt54 is a circuit diagram of the decompression unit same as above, and Fig. 5 is a circuit diagram of the invention 6 is a circuit diagram of the decompression unit of the third embodiment of the present invention, FIG. 7 is a circuit diagram of the image compression unit of the fourth embodiment of the present invention, and FIG. 8 is a circuit diagram of the decompression unit of the same as above. tjS9 is a circuit diagram of the conventional image compression unit which is the basis of the present invention, FIG. 10 is a circuit diagram of the decompression unit of the same as above, and FIGS. 11 and 12 are of the same as above. While explaining the operation, FIG. 13 is a circuit diagram of another conventional image compression section. A... Image compression unit, 1, 9.7... Line buffer, 2... Prediction circuit, 4... Variable sample density encoding circuit, 5... Zero code compression circuit, 6... Variable sampling density decoding circuit, 7... Line buffer, 8... Interpolation synthesis circuit, 27... Code distribution calculation circuit, 28... Fixed/variable length codebook selection circuit, 29... Fixed length /Variable length encoding circuit.

Claims (6)

[Claims] (1) In an image compression method that divides an image into multiple blocks and performs quantization coding on each block, a quantization code corresponding to a combination of a time difference value and an amplitude difference value is determined based on quantization characteristics, and the quantization code corresponding to the combination of a time difference value and an amplitude difference value is calculate a statistic representing the degree of concentration of the probability distribution of image information for each block based on the quantization code, or a statistic representing the degree of concentration of the probability distribution of the quantization code,
An image compression method characterized in that a codebook of variable length encoding or fixed length encoding is selected by comparing the magnitude of this statistic with a certain setting value. (2) Calculate the statistics representing the degree of concentration of the probability distribution of image information for each block, select fine gradation quantization characteristics and time difference values for those with a strong degree of concentration, and Select a coarse gradation quantization characteristic or an amplitude difference value, quantize it and encode it, find a statistic that represents the degree of concentration of the probability distribution of this quantization code, and compare the magnitude of this statistic with a certain setting value. 2. The image compression method according to claim 1, wherein a variable-length encoding codebook or fixed-length encoding is selected according to the following. (3) Quantize and encode the coarse image to restore this coarse image, then obtain an interframe error signal between the original image and this coarse image, and divide this interframe error signal into multiple blocks. , the inter-frame error signal is quantized and encoded for each block, and a statistic representing the degree of concentration of the probability distribution of the inter-frame error signal information or a statistic representing the degree of concentration of the probability distribution of the quantization code is calculated. Select fine gradation quantization characteristics and time difference values for the units with strong concentration, select coarse gradation quantization characteristics and amplitude difference values for weak concentration, perform quantization encoding, and calculate the above statistics. Alternatively, a variable-length encoding codebook or fixed-length encoding is selected by comparing the magnitude of a statistic representing the degree of concentration of the probability distribution of the quantization code with a certain setting value. The image compression method described in Section 1. (4) A pixel whose inter-frame error signal is larger than a certain value is a changed pixel, a pixel whose value is smaller is a zero pixel, and an inter-frame effective error signal is calculated by setting the inter-frame error signal of this zero pixel to zero, and this inter-frame effective error The signal is subdivided into multiple blocks, and the inter-frame effective error signal is quantized and encoded for each lower-order block, and the statistics representing the degree of concentration of the probability distribution of the inter-frame effective error signal information or the probability of the quantization code are calculated. Find the statistical value that represents the degree of concentration of the distribution, select fine gradation quantization characteristics and time difference values for those with strong concentration, and choose coarse quantization characteristics and amplitude differences for those with weak concentration. Along with selecting and quantizing the values,
A patent claim characterized in that a variable-length encoding codebook or fixed-length encoding is selected by comparing the magnitude of the above-mentioned statistic or the statistic representing the degree of concentration of the probability distribution of the quantization code with a certain setting value. The image compression method described in item 3. (5) Next, find the first difference, which is the prediction residual between lines, between the pixels of the current line to be compressed and encoded and the predicted value of the next line based on the restored pixels of the previous line that has been compressed and encoded. Regarding the first-order difference, the next one based on the first-order difference of the pixel to be compressed and encoded next in the line direction and the restored pixel that has already been compressed and encoded.
A second-order difference, which is a prediction residual between the predicted value of the first-order difference, is obtained, and a quantization code corresponding to the combination of the time difference value and the amplitude difference value is obtained from this second-order difference using the quantization characteristic, and the image is divided into multiple blocks. divided into blocks and quantized and encoded in units of blocks, and obtain statistics representing the degree of concentration of the probability distribution of image information for each block, or statistics representing the degree of concentration of the probability distribution of the quantization code, and calculate this statistic. 2. The image compression method according to claim 1, wherein a codebook of variable length encoding or fixed length encoding is selected by comparing the amount with a certain setting value. (6) The inter-frame difference between the current image frame to be compressed and encoded next and the restored previous image stored in the frame buffer or the predicted image frame based on the current image is quantized using second-order difference measurement, or Find pixels that have a difference between frames that is larger than a certain threshold and zero pixels that are small.
In addition, an effective component image is obtained with the current image of zero pixels or the pixel value of the inter-frame difference set to zero, and this effective component image is divided into multiple blocks and quantized and encoded for each block, and the effective component image is calculated for each block. A code book for variable length encoding is obtained by calculating a statistic representing the degree of concentration of the probability distribution of image information or a statistic representing the degree of concentration of the probability distribution of quantization codes, and comparing the magnitude of this statistic with a certain setting value. The image compression method according to claim 5, characterized in that alternatively, fixed length encoding is selected.