TW490617B - Process for compressing digitized images, in which a coding quality factor is ties to each zone - Google Patents

Abstract

The invention relates to a process for coding digitized images with a view to their compression, in which the image is separated into individually coded zones, such as macroblocks. This process is characterized in that a filtering (10) is applied to the source image and a filtering is applied to the same image emanating from the decoding of the coded image, so as to determine a quality factor for each zone, representing the deviation (20) between the decoded image and the source image, and in that this quality factor is used to modify a coding parameter of each zone so as to minimize the deviations between the quality factors of the various zones. Preferably, the filtering is psycho-visual application to MPEG coding.

Description

490617 V. Description of the invention (1) The present invention relates to a compression method for digitized images. A digitized image is equivalent to a lot of information. In order to store and / or transmit relevant image information, especially related to moving images, it is necessary to rely on a compression method that enables it to record or transmit a reduced amount of digital data without loss of the quality of the reproduced image. Port compression involves using various techniques to avoid transmitting or recording large amounts of information, and Huffman coding specifies the number of bits (bits) for the most frequently occurring information. Compression also depends on the nature of human vision, especially the fact that the human eye is sensitive to fine details (· high spatial frequency), coarser details (low spatial frequency) are low, and the visibility of fine details increases with movement. Descending. Therefore, the compression method relies on a transformation that assigns greater weighting to low-level spatial frequencies than to high-level spatial frequencies. In the common compression method that forms the MPEG standard, a separate cosine transform (DCT) is used. Then for simplicity, reference is made in particular to the MPEG2 standard, although the invention is not limited to this standard. In this method, the image is divided into large blocks of 16x16 image elements (pixels), each of which is formed by 4 blocks each having 8x8 pixels.

According to the MPEG2 standard, each huge block of 16x16 pixels is transformed into a huge block of matrix coefficients, whose rank is equivalent to the horizontal spatial frequency, and its rank is equivalent to the vertical spatial frequency. These coefficients are encoded according to the accuracy defined by the quantization interval. The larger the value, the worse the accuracy. According to the MPEG2 standard, the quantization interval one can be selected for each giant block. Therefore, for each giant block, you can choose to obtain excellent image quality (thus implying a small quantization interval) and limit the number of bits to be transmitted (this implied amount

i. Explanation of the invention (2) ------------ Interval is large) Although this item has a quantization interval / quantization interval; the same quantization interval, = possibility of modulation ', actually, each image is spaced. And the quantification of changes in the use of giant blocks of various levels._ In addition, the coding cost determined on the transmission wheel; green standard; the storage standard often assigns a fixed bit rate per image or the proposed coding. The coding cost is generally determined by Control module duties. The selection method lies in the quantization interval displayed by each giant block or each block, which is like the overall cost of coding, and it is obtained on the block! Excellent quality '^ It is consistent with the quantization interval of each block of the shadow. The block or block should be selected so as to optimize the quality of each giant block or block, a predetermined value. That is, the total number of bits (encoding cost) of the image must be equal. However, there is actually a defect in not using this encoding. Obvious results are obtained through oral experience. The present invention is based on observations, and the human eye is sensitive to sample quality, which means that different quality is combined with as much quality as c-quality to give the image a flawed impression. -In order to make the eyes feel, the method of the present invention is characterized by each image area, especially the $: block '. The coding method makes all the areas display a quality factor that is substantially the same or better, which is spiritually visual. . Prime Factor 'In order to obtain uniform quality in the regions, it is best to change the quantization interval. < In the specific example ', the figure of merit is the video signal to noise ratio. The reason is that the difference between the pixel intensity of the root image obtained by each pixel and the corresponding pixel brightness of the decoded image is measured in a known way.

Page 6 490617 V. Description of the invention (3) The signal-to-noise ratio is determined by, for example, the sum of the squares of the differences in brightness of all pixels in the area. In one example, the signal-to-noise ratio has a PSNR or SNR value defined by the following first and second formulas: 255 PSNR = lOlog Σ, dif f:

N max ^ SNR = lOlog

In N formula, d i f f refers to the brightness difference between the pixels of the root image and the decoded image, max is the maximum value of the pixels considered, and N is the total number of pixels. Although this figure of merit provides satisfactory results, the best result is obtained by a figure of merit related to the mental and visual performance of human sight. In this situation,

丄 / V. Description of the invention (4) Each area: Product: The factor depends on the brightness, spatial frequency, color and motion. This kind of figure of merit is defined as follows: point-picked images are subjected to mental visual filtering, and the difference between each image point by @ prime represents the figure of merit after searching. Psycho-visual filtering involves applying proof brightness on the one hand and visibility on the human eye on the other. Because of this, the image contains a reduced part of the visibility of the image, spatial, also; r due to the uneven sentence of the jade clothing environment. The non-uniformity can be gated or equivalent to movement or time. Inter-occlusion is defined as: a function that expresses the relative, wide, and spatial frequency of the image relative to the environment. Fine details may not be felt, but the spectrum and surrounding spectrum are extremely sharp: defects are more visual. Gu Wu said that this kind of spatial occlusion of Feng Feng's different flaws is particularly referred to the following paper: "Image Quality Prediction Based on Visual Model COM-30 Volume 7, Issue m · J · Lukas and ZL · Budrikis", I EEE Communication Conference Record July 1982. 2 · N · β · Nill, "Weighted Cosine Variations of Visual Models for Shrinkage and Quality Assurance", £ 1: £ 5; Proceedings of the Communications, Volume com 33_ June 1985. Peng ',, 3 * K: N. Ngan, KS · Leong, and H · Singh, "Cosine Variation Coding of Images in a Receptive Field", IEEE Acoustics. Speech and Noise Processing Conference Records (37:11) Issue), November 1989. ° ^ 4. L. Carrioli and M.G. Albanesi, "Based on human eye shovel-shift work

490617 V. Description of the invention (5) Compression of the grayscale image as much as possible>, Alta FreQuenza (Vol. 57 No. 5), 273-284 pages, June 1988. 5. SJ Daly, "Noise Adaptation for Image Data Compression, Visual Contrast Sensitivity," Eastman Kodak Company. Motion masking is a kind of filtering, which proves that the sensitivity to fine details actually decreases with motion, and in terms of average speed, the low-yield #gate frequency is higher and the spatial frequency is more visual. For further details on motion occlusion, see the following publications. 6. DH Kelly, "Motion and Vision, Temporally Critical Surfaces of Stable Space", J. Opt. Soc. Human 111. (Volume 69, 10 Issue), October 1979. 7. · S. Daly, "Engineering Observation from Space Velocity and Temporal Temporal Visual Modes" 'Human Eyes and Electronic Imaging Π' SPIE Conference Records (32L, pages 180-191, January 1998. Finally, 'temporary masking is equivalent to 20ms The temporary total of Bloch's law for the following short-term image parts. According to Bloch's law: the brightness deviation (or error) dL! Is felt during the image ^, which is the same as the production deviation dh experienced during h ms, but these values meet the following Relationship: Eryou Ershan = (^ 21: 2, (t " t2) < 20ms, so 'according to these waves, and the difference between the wave passing through the source image and the wave-out decoding image', each pixel of the image Or if the image ^ establishes a quality factor, it is a double number for each area (especially a block or a giant block), especially a quantization interval, so that the dispersion asymmetry of the quality is minimized.-In general, the f factor Between the values of the parameters

丄 / V. Explanation of the invention (6) Good results have not been known in the relationship. For all the images in the process, the root parameter parameters have high quality, for example, the quality of Tienan. Here this change phase is in the second determined encoding 0 '. Because of this, it is best to specify the modification based on this product, repeat the average withdrawal amount as an example step, pass the parameter variable, and continue the inverse change one step encoding. The difference between the parameter values The parameter quality is determined by the value of the quality code parameter in the step. If a specific example of the image neutralizes the performance, it is selected based on the connotation of each image and image. This relies on iterative procedures, including for example ::: should; determine: encoding parameter variables, and establish a quality image, determine the average quality factor is expensive quality. In the step of reducing the quality and reducing the quantization interval, it is best to increase or decrease the coding parameter to minimize the value, such as gradual. In the process, the first step is applied in this way. This second step is sufficient to obtain the correct uniformity of quality. The best result can be obtained by repeating the following steps: The initial step is assumed to be proportional to the change of the encoding parameter variable to determine the number of times. Therefore, the proportionality coefficient is the difference between the first one on the one hand, and the n-th step, the ratio between the other. The new value available for this proportionality factor is the result of the newly determined change in the quality factor. The coding cost must show a certain value to meet this condition. In, select a number of regions and change its coding parameter. Number of selected zones The variation range of the coding parameter is the result of Zhong Yu's achievement that the coding cost has a certain value. For example, during a step,

490617 V. Description of the invention (7) If the coding cost is within the required coding cost (for example, 10%) and t d & is not the lowest quality (in terms of uniformity), the other two: Low ', and thus-aspects improve Pinjie it female} — 1 plus coding costs. On the other hand, if the encoding quality is uniformized, the k% (for example, 5%) area of the best quality is displayed, and the quantization interval is determined to reduce the encoding cost. good. =: In the process, the lowest quality region and the best time quality & of y% are selected to change the encoding parameter variables. When this operation is performed using Yanshi ##, the coding cost changes, for example, the selected determination factor is used, and the cost tends to the correct value. For example, if inscription == and / or y ′ makes the encoding b ′ rain code too expensive, you can filter the number of regions in the team's supply interval, or increase the number of regions to increase the quantization interval. -Reverse $ ΐ! Each modification operation ’is performed repeatedly using calculations. This is a minor and complex process that cannot be repeated during the actual change of the 1-interval. In this major iterative process, each time you change the quantization interval ^ /, for example, gradually. Take the minimum value Generally speaking, it is not possible to achieve the entire image quality factor of a =. Between this iterative step, if the interval is not asymmetric, the full repetition will be interrupted. Low time 'Therefore, the present invention generally regards digitized images in view of compression = decision, in which the image is divided into individual coding regions, such as giant blocks, basin = code root image filtering, and the encoded image is decoded as a wave to determine The quality factor of each area represents the deviation between the two images of the decoded image: C, and this quality factor is used to change each of the :: 翏 variables to minimize the deviation between the quality factors of the areas. < coding tits 0617

During the first step, the editors of each district changed. The coding parameter variable is in the least increments. During the second or subsequent steps, the π variable change system should be used. Temporarily due to the convergent changes h, t / According to the expected pattern of the coding parameter steps of each area, a coding parameter variable to be used in the next step is determined. According to a specific example, the image coding cost is predetermined or supervised. In the two steps or the subsequent steps in the prismatic system, the change of the encoding cost should be ^ f Q. The encoding cost is a pre-model that is proportional to the variation of the encoding parameter variable. In the latter case, the coding parameters of each zone should be determined to minimize the change in the 'coding cost' in the next step compared to the ongoing step. In the second step or subsequent steps, it is best to: repeatedly determine the coding parameter variables of each area, and gradually change the coding parameter variables of the first determined portion of the σ prime factor area to gradually reduce the quality of the mouth by 2 2 By gradually changing the second certain part of the poor quality area, the encoding parameter variable is improved to improve the quality; it is determined that the item (or these) of the encoding parameter variable is changed to obtain the encoding

490617 V. Description of the invention (9) Changes to the book, and change the number of regions that change the encoding parameters, so as to minimize the deviation from the encoding cost required for the image. In this case, the steps of each iterative process of gradually changing the coding parameters that show the best and / or number regions are to average the deviation of the whole prime factor from the average figure of merit, and when this ^ tends to decrease, repeat That's an interruption.卞 g is no longer according to a specific example, the coding parameter variable is the quantization interval for each picture element assigned to the area. The value of the relationship number and this mental visual filtering takes into account the effects of spatial occlusion and / or occlusion. A linkage and / or temporary Ϊ = filtering determines the image signal to noise ratio. In a specific example, the coding is related to the object-oriented degree of benefit. The main σσ prime factor and region When the code is object-oriented, the uniformity of the quality factors is the best. The various objects of the present invention can make the method of the present invention particularly applicable to the image. Other features and advantages of the present invention can be obtained from Shenzhao ^ / or pass. In the drawings, some specific examples of the drawings are described. The first diagram is the method and equipment of the present invention. The second diagram is a simplified diagram of the ancient road and the soil. The steps of the method of the present invention are illustrated. ; The figure is a summary of the method obtained by the method of the present invention (the method described with reference to the drawings, including borrowing: a graph.), While maintaining the coding cost of each image-fixed, quantization of each S block / or supervised, and the reverse page 13 490617 No. 1 _ 89103326 V. Description of the invention (10) In line ii, the subjective quality of the MP £ G2 block is uniformized. —__ Code 1 9 As shown in Figure 1, the root image is wailed by mental vision 1 η < Calculate the height of the two daggers ;; Use the unit "Communication to the code: The dagger image also looks at ί; = ;: Decoded (block ⑻, and this code ⑽it; j root: the figure of merit and the solution to the image, For each block image ::; the deviation between the average quality factor of the Tf image quality horse shirt image ^ the 1000 average quality factor and the resolution of the lens 20 is 3: 2 miles J :: 'Block 14 used data, including mouth ⑯ 八 Λ 彳 block 22 ^ encoding cost of the block. The mouth and shellfish scattered scattered in the specific example, the quality of each giant block, the relationship. Decrease the quality dispersion, and use the helmet mode between the political and internal intervals, and repeat it. Use the following steps based on the expected steps in the initial steps, either by always or by "control (not shown in the figure)"; Cost,…. The coding cost of each image using the data bit rate that determines the bit rate of the image must be continued. ^, ,, and not all images are the same, or are supervised, which means that each is determined. In addition, it is assumed that quantification The interval is expressed as an integer between the quantization and 31 during the initial steps. 疋 Established at a constant value and also provided by the competitor on page 14 2002 · 〇3 3〇.〇14 V. Description of the invention (11) Photo provided. IF 12 provides coding costs on the one hand (block 22), and the image of _, 1 10 6 ^ code, at the same time as the root image, is mentally astigmatized, and also reflects 20. The provided by block 20 It is not only the average dispersion or error of the order of the gold department to the image. 舛 Zengxi During this first step, block 14 prepares the first step for each giant block: ... in the second step of the process, the quantized interval block is used to enter the image. ΐίϊίdecision ;; scattered or error If it is greater than or less than the entire shadow (progressive), and the error is less than the average, the quantization interval is reduced by one unit.) For large blocks, the quantization interval is increased by one unit. If there is a slight decrease or a large increase, the increase is changed so that :: The error is equal to the average, for example, the following value A ^ Bk is started gradually. For this second step, the determined quantization interval (the original code) The device uses each giant block and then determines each giant block (increase and decrease the unit by one unit). For the number of steps, use the interval QSCnk, k represents the cost of the wall code, and the quantization interval determines the connection between the summerization interval and quality. The factor and the coding of this predictive quality factor £ «^ ~ and the full / item pre-4 mode are simple straight-line functions: such as editing the cost C 0 S tnk. (2) 490617 V. Description of the invention (12) £ fflapnk = Akx QSCnk + Cst C0stnk = Bkx QSCnk + Cst ’In this equation, Cst and Cst represent constant values. Therefore, at each step, the coefficients Ak and Bk of each giant block can be determined or updated according to the following formula:

Ak

Cos tnk-Cos t (n-l) k △ Cost!

Qscnk-Qsc (n · l) k Δ Qsc, (n-l) k

Emapnk_Emap (3) scnk ~ Qsc (n.1) k

Bk keeps the quantization interval between Γ unchanged, which means △ "~ = 〇, then" ", the amount of ^ ^. However, in the second step, since the interval given in the first step does not have to be changed from the previous step. Since ^ this u figure of merit and the formula 2, from the second step, we can foresee the large blocks). 'Ma cost (quantization interval has been determined in the previous step of the entire shadow Jun μ The following relational table, the estimated change in the cost of the mother ΛΡί (: 1 :( :: 〇3ΐ: ΡΓ6ί1') can be borrowed on page 490617 V. Description of the invention (13) APictCostPred = EBkx AQsck (4) k Assuming that the RefPictCost value can be used to determine the coding cost, you can seek to minimize the amount: ACostErr = APictCostPred- (PictCostn-Re / PictCost) * (5) In this formula, PictCostn Represents the cost of image coding at the moment. This amount ACostErr represents the difference between the estimated change in coding cost and the previously determined excessive cost. Unless otherwise stated, where the quantization interval or its change △ QsCk is sought, the full cost of image coding can be maintained Try to be as close as possible to the reference coding cost of this image. At the same time, of course, seek to equalize the quality factor & 'meaning to reduce the product: (the difference (2 phase map pair) in the first group of flat, specify [△ _] +, including = interval increase ) Of all giant blocks' and the error of the quality factor is low-the second group indicates all the giant blocks of [A Qsr 1--the quantization interval is reduced], and because the rescue contains Δjsck is negative (error. Gong Er The error of the quality factor is greater than the average

490617

Contains quantization interval. Case No. 89103326 V. Description of the invention (14) A third group, which indicates all the large blocks changed by [△ Qsck]. Therefore, the coding cost of the giant block of the first packet is equivalent to a decrease, and the coding cost of the giant block of the second packet is equivalent to an increase. Based on this item, it is divided into groups. In the second step (or n step), another iteration is performed to determine the quantization interval, provide the maximum uniformity of the quality factor for the next step, and minimize the error of the cost relative to the reference cost . For this purpose 'select from the -th group to display the smallest grouping error, block (eg' select from the second group to display the largest grouping error of f% huge blocks (eg 15%). For the group [AQSCk] + all k +% For large blocks, the quantization interval is increased = the difference is asymptotic, thus reducing the coding cost, ... ^ 丄 ^^ All k% giant blocks of the knife group [△ "~], and the quantization interval ^ minus f determines the weight dQ-, such as gradually increasing, The average error approaches the disease of the factor gate and increases the cost of editing. Based on the improvement of these quantization intervals, I can predict the products of each giant block; Taking factors such as four as one, increase the groupings in parallel [eight], k, giant

Page 18 2002 · 03. 30. 018 490617

490617 V. The area of invention description (16)) The coding cost has a monotonous effect. Furthermore, this method has nothing to do with the filtering, mental vision or other types used. Due to the repetitive nature, the method of the present invention is well-suited for impact compression to be stored, such as producing DVD discs, or storing on tape. However, if the processing speed is high enough, this method can also be used for instant storage or instant transmission. The encoding method of the present invention is not limited to the MPEG method. It can also be applied to other compression methods, such as JPEG (Compression of Still Images) or ETSI. The DCT of individual cosine functions can also be changed to rely on fast Fourier transform, or the spectrum can be split into bands as a workaround. The compression method of the present invention should also be applied to object-oriented coding, such as MPEG4. In this case, the figure of merit includes parameters that give favorable image areas; as a variation (or combination), the diverse objects (multimedia) of the image are individually processed to uniformize the figure of merit of each object. Regardless of the application, the method and apparatus of the present invention can improve the subject quality of the encoded audio sequence. Ϊ · Π · Page 20 · 490617 Brief description of the diagrams Figure 1 is a schematic diagram of the method and device of the invention; Figure 2 is a diagram illustrating the steps of the method of the invention; Figures 3 and 4 are the results of the method of the invention Graph Φ

1BI Page 21

Claims (1)

490617 VI. Scope of patent application Costs. 8. If the variable is applied, it will be determined in progress. 9: If it is in the application process, or as the best quality of each area changes, it will show the decision to improve this quality, and the coding required for the change will be 1 0. / In the process, the average of pairs, or. 11. Rushen is assigned to the district 12. Rushen is aware. 13. For the method of applying for item 7 of the patent scope, wherein the coding method of each area is that in the next step, the change in coding cost is minimized relative to step t. The method according to item 8 of the patent, wherein in the second step and subsequent steps, the coding parameter variables are repeatedly determined, that is, the coding parameter variables of the first determined part of the display number are gradually changed, and / or gradually worsened The coding parameter variable of the second determining part of the quality is used to change the coding cost obtained by the change of the coding parameter variable; the number of zones of the variable coding parameter variable is changed so as to minimize the change relative to the shadow cost. The method according to item 9 of the patent, wherein in each iterative step of gradually changing the coding parameter variables of the apparent or worst quality factor area, all the areas are determined to change the figure of merit relative to the average figure of merit and when this average no longer tends to decrease, The method of item 1 of the patent scope is ceased repeatedly, and the coding parameter is the quantization interval of the value of the correlation coefficient of each pixel. The method of claiming the first item in the scope of patents, wherein the filtering is the method of claim 12 in the scope of the mental patents, in which the mental visual filtering is performed. Ββγί · Page 23 490617 6. The scope of the patent application takes into account space shielding and / or motion and / Or the effect of temporary cover. 14. The method according to item 1 of the scope of patent application, wherein filtering determines the ratio of the image signal to the noise. 15. The method according to item 1 of the scope of patent application is applied to MPEG2 type encoders. 1 6. The method according to item 1 of the scope of patent application, wherein the coding is object-oriented, and the quality factor is related to the degree of regional benefit. 17. The method of item 1 of the scope of patent application, in which the coding is object-oriented, and the uniformity of the quality factor is the best, exceeding the individual objects obtained individually. 18. The method described in item 1 of the scope of patent application shall be applied to the storage and / or transmission of the image. Page 24