RU2420021C2 - Method to compress images and video sequences - Google Patents

Abstract

FIELD: information technologies.

SUBSTANCE: it is proposed to represent each colour pixel with three colour components, every of which is initially coded with ten bits. Coding is carried out by breaking the initial colour video frame into non-covering space units, and subsequent separation of a bit representation of each colour component of a pixel into a senior part, made of more than one senior bit, and a junior part made of at least one junior bit, then separate coding of the senior and junior parts, besides, coding of the senior part is carried out by application of more than one coding method, every of which takes into account pixel-to-pixel connections only within the limits of the processed space unit, estimation of a coding error, selection of a coding method, giving the smallest error, sending data on the method of coding by means of a prefix code transfer, coding of the junior part, which is carried out by averaging of more than one value, included into the junior part, besides, dimensions of the averaging areas within the limits of the junior part, a number of bits, fixed in advance, is set as required for a compact representation of initial space colour unit.

EFFECT: efficient compression of a colour high-quality image without visible visual distortions.

25 cl, 29 dwg, 13 tbl

Description

The invention relates to methods for processing video data and, in particular, to a method for compressing video sequences.

As you know, the efficiency of information exchange is improved if it is possible to reduce the redundancy of the transmitted message, for example, by removing non-essential parts, or rather, such elements that are quite easy to recover on the receiving side. If, when transmitting written messages, whole phrases can be replaced by conventional symbols, for example, mathematical, or abbreviations, then when transmitting images, and even more so videos, you have to use more complex methods, removing from the transmitted images irrelevant information about points (pixels) according to a specially developed technique and then restoring these points when playing back images. At the same time, the degree of compression and the correctness of image recovery depend on the algorithm used, and complex, detailed algorithms require significant computing resources.

Various approaches and algorithms are known in the art for compressing images and video sequences, in particular, the solution described in US Patent Application Laid-Open No. 2008/0019598 [1] should be mentioned. In the described method and device, rather large blocks of 4x4 pixels are subjected to processing, which implies buffering of four lines (lines) in the input stream. In this case, four colors and the corresponding raster display (bitmap) are formed. This approach provides compression without visual loss of only signals with a high degree of stationarity, i.e. Suitable for compressing a limited set of natural images.

In US patent No. 7239743 [2] and No. 7082217 [3] also an attempt is made to solve the compression problem with a fixed compression ratio based on the encoding of the image blocks of 4 × 4 pixels. Both patents provide for the formation of a color palette with the construction and encoding of a bitmap. The main disadvantage of these solutions is that they are focused on working with a relatively large block (4 × 4) and are unable to efficiently process non-standard areas, for example, sharp unpredictable contours (transitions) inside the image. The method described in "Graphics for the Masses - A Hardware Rasterization Architechture for Moible Phones" by Tomas Akenine-Moller and Jacob Strom ACM Transactions on Graphics, vol.22, no.3, Proceedings of the ACM SIGGRAPH 2003, p works slightly better. .801-808, July [4]. This method uses a texture compression algorithm based on processing the image with blocks of a fixed size, for example 4 × 4, 2 × 4, 8 × 8, followed by calculating the average values of these blocks and calculating in each pixel the difference between the individual value of its brightness and the average value and encoding these differences using a set of fixed quantizers. This method also proved to be ineffective for processing sharp color and brightness contours.

Closest to the claimed invention is the solution described in the laid out application for US patent No. 2007/0253483 [5]. This solution also uses 4 × 4 blocks and uses spatial prediction of block values based on various spatial directions (vertical, horizontal, diagonals) and in addition introduces the so-called spatial filtering of the processed pixel (meaning the elimination of correlation in one of the spatial directions) , determine the prediction mode and coding. In addition, the filtered pixel values undergo various block transforms and coding. This approach uses several encoding modes, but it does not guarantee a fixed bit rate, and the processing unit has a relatively large size (4 × 4). In addition, this approach uses a buffer string to store previously encoded data, which does not allow independent coding of image blocks, i.e. the degree of stability of the algorithm to interference is low.

The problem to which the claimed invention is directed is to develop an effective method for compressing high-quality color images without visible visual distortion and transmitting encoded video data with a fixed compression ratio for a fixed small portion of the image in advance using limited computing resources.

The technical result is achieved by developing an improved method for encoding a high-quality video sequence based on the encoding principles for blocks of 2 × 4 pixels and encoding frames of photo or video images with a resolution of 10 bits per color component while maintaining a fixed bit rate and a fixed compression ratio, moreover, the two lower bits and the eight upper bits are encoded separately, grouping the bits thus separated into blocks, using d For each similar block, there is more than one encoding method, with the subsequent estimation of the encoding error and with the choice of the method that ensures the minimum error, while the message about the chosen encoding method of the block is transmitted in the form of a variable length encoded prefix.

Moreover, for the effective application of the method, it is preferable that the video sequence was presented as a set of frames, each frame being recommended to be processed as an image, and each image should be represented as a set of non-overlapping blocks of 2 × 4 pixels, provided that the compression ratio is 2, four.

It is also advisable to consider the upper eight bits of the image in the form of blocks of 2 × 4 pixels, each such block being treated as a single block of 2 × 4 in size (hereinafter, methods of such processing will be discussed, in particular, methods identified as F-, O -, D-, L-, U-, H- and M-methods) or as two blocks of 2 × 2 pixels in size (N, P, S, C and E methods are used here).

As the main advantages of the proposed method over the known analogues can be called the following:

- compression of a color digital image and video sequence without visually distinguishable distortions at a fixed data rate (fixed bitrate);

- encoding and decoding with a fixed compression ratio of 2.4 times for color images with 10-bit depth for each color component with small computing resources while maintaining high quality;

- the use of only one buffer line, which guarantees a low signal delay and an economical mode of memory use.

The inventive method is oriented to work with small blocks of 2 × 4 pixels and provides for their independent coding. It ensures the stability of the compressed bit stream to interference and random access to any part of the processed image without decoding pixels from previously encoded blocks.

The method includes many compression methods, each of which corresponds to certain characteristics of the processed image; the choice of coding method is based on the estimation of the coding error of a given block by all possible methods and the choice of a method that provides the smallest error; when decoding, to determine the method used in encoding this block, a special feature is analyzed, transmitted by variable-length prefixes.

Effective coding of the original color image with a color depth of 10 bits per component is ensured by dividing it into two parts; visually important signal (eight upper bits that form the image) and non-essential signal (lower two bits). This separation allows you to enter various degrees (levels) of coding accuracy while maintaining a reliable representation of the main visual information.

In more detail, the claimed method can be represented as follows.

The method is based on block coding, while it is focused on processing input blocks of 2 × 4 pixels in three color channels at the same time. This block initially requires 240 bits to store. After compression is completed, 100 bits are required to store such a block, which together comprise the following four elements:

- A heading indicating which compression method is used.

As a result of the experiments, it was found that the best option would be to use a Variable Length Encoding Header (VLEH). The size of the header depends on the method used and can vary from 1 bit to 10 bits.

- Data on the compact representation of the least significant bits of the block (Least Significant Bits Data - LSBD). In a compressed 2 × 4 block, this data, as a rule, depends on the chosen encoding method and takes from three to six bits per block.

- Compressed bitstream. It represents in a compact form the top eight most significant bits for all pixels within a 2 × 4 block.

- Additional part. These are complementary bits within a packet that are introduced to provide a fixed bit rate. By its nature, this part is optional and is used only if necessary.

Due to the limitations associated with the possibilities of redistributing the number of bits between different parts of the image, the efficient use of traditional statistical coding combined with decorrelation conversion is very difficult, as the results of relevant studies show. In addition, due to the lack of the possibility of effective decorrelation in the vertical direction (limited amounts of internal memory), universal and relatively effective traditional methods of encoding micro images in these conditions are unsuitable.

That is why, in order to compress ten bits of the image, it is proposed to apply two separate compression methods for each color channel: one for the upper eight bits, and the other for the lower two bits.

In this case, the senior eight bits of the source block of 2 × 4 pixels in size, which together also comprise a block of 2 × 4 8-bit values, are encoded in all possible ways, and then the method that provides the smallest error is selected. Proposed in the invention methods based on the size of the processed block can be divided into two groups. The methods of the first group are designed to process the entire 2 × 4 block. The methods (methods) of the second group are intended for processing a 2 × 4 block in the form of two 2 × 2 subunits, and for each such 2 × 2 subunit the optimal method (method) is selected. The present invention proposes the use of seven methods of the first group, namely:

D-Way. Designed for coding blocks, which include complex sections of the image and sections with pronounced diagonal boundaries.

F-Way. Optimized for efficient transfer of images that contain areas with a pronounced gradient. The method can be applied both in relation to one color channel, and for simultaneous processing of all three color channels.

M-Way. Designed for processing control images, with which they check the resolution of the output device, set up television receivers, etc. Similar images consist mainly of curves with different amplitudes, directions and changing frequency characteristics.

O-Way. Use when one color channel in its energy characteristics as a whole significantly surpasses other channels inside the processed 2 × 4 block. In this case, four possible situations are considered when the luminance channel, the channels R, G or B dominates

L-Way. Applied to a 2 × 4 block in all three channels simultaneously. To ensure an effective representation of the internal structure of the image, for each 2 × 2 subblock an own palette is formed. Then one of three modes is used: the mode of inter-block differential coding of palettes for two adjacent 2 × 2 sub-blocks, the mode of differential coding of colors inside each sub-block and the mode of direct quantization of colors of the palette.

U-Way. Designed for such areas of the image, which are a combination of structurally complex and uniform areas; in this case, uniform regions are encoded by an average value, while structurally complex regions are transmitted by more efficient methods.

H-Way. The method is intended for encoding portions of natural images in which there are smooth transition regions, without pronounced diagonal-oriented borders. The method processes a 2 × 4 block, considering it as a combination of two 2 × 2 blocks, uses a discrete 2 × 2 cosine transform. The coding of coefficients is based on the use of a set of fixed quantizers.

Five other methods were developed for 2 × 2 subunits (to distinguish them from the previously described methods, we will call them “methods” below) included in the second group. These methods are independently applied to both 2 × 2 subblocks, and then for each source block a combination of methods is selected that provides the minimum error. The methods of the second group include the following:

N-method. The method is intended for encoding regions of natural images, based on a new transformation of a block of pixels called the NSW transform and the procedure for fixed quantization of difference components. NSW transformation is a simple two-dimensional transformation based on a combination of predicative coding, one-dimensional Haar wavelet transform, lifting scheme. As a result of applying this transformation to the pixel values, difference values are calculated, which are then subjected to fixed quantization based on a set of nonlinear fixed quantizers.

P-method. The method is a modification of the N method and provides coding of the 2 × 2 subblock using vertical / horizontal / diagonal preferred directions, as well as coding based on the average value.

C method. The method is based on the NSW transformation and consists in its application to all color channels simultaneously with subsequent group coding of difference values from three color channels together. Due to this, if the difference values of the NSW transform over the three color channels are sufficiently similar, an increase in the accuracy of coding (quantization) is achieved by increasing the dynamic range of the quantizer and eliminating the independence of the choice of the quantizer from the color channel.

E-method. The method is used to increase the accuracy of color transitions occurring at the border of two pure colors.

S method. This method is very similar to the N method and has a modified encoding procedure, which is adapted to transmit portions of images in which difference values take relatively small values for all three color channels within the 2 × 2 subunit.

The choice of coding method and method is determined by which of the applied methods and compression methods of the 2 × 4 color block provides the smallest error value.

The choice of method is carried out in stages.

Stage 1. The left and right 2 × 2 subblocks included in the processed 2 × 4 block are subjected to compression by five methods: N, P, C, S, and E. As a result, a set of errors (errors) and intermediate bit streams are formed.

Step 2. For each subblock, select the best method. The choice is based on the search for the smallest error among all calculated.

Step 3. Determine the best combination of methods for the 2 × 4 block based on the N, P, C, S, and E methods applied to the left and right subblocks, and also calculate the combination error.

Step 4. The 2 × 4 block is processed in seven ways: D, O, M, F, L, U, H. In this case, the coding errors of the current block are calculated by each of the seven methods indicated.

Step 5. The minimum error is determined by comparing the combination error and the errors caused by the application of methods D, O, M, F, L, U, H. As a result, the best method is selected.

Step 6. VLEH corresponding to the selected compression method is entered into the output buffer.

Step 7. A compressed bit stream is also entered into the output buffer.

Step 8. Based on the selected compression method, the corresponding LSB coding is performed and the coding results are also included in the outgoing stream.

Step 9. If necessary, add complementary bits to ensure a fixed transmission rate of the processed block 2 × 4.

Figure 1 - the main steps of the encoding method.

Figure 2 - view 2.1 - the original subunit 2 × 2; view 2.2 — directions of differences of the NSW transform; view 2.3 is a block diagram of the N-method.

Figure 3 - views 3.1-3.4 - the preferred methods of encoding the P-method; view 3.5 is a block diagram of the P-method.

4 is a view 4.1 - a scheme for evaluating additional free bits for the P-method; view 4.2 is a sequence of bit stream with additional free bits.

Figure 5 - view 5.1 is a simplified sequence of actions of the E-method; view 5.2 is a graphical representation of the structure of a 2 × 4 block and its interpretation in the E-method.

6 is a detailed diagram of a method for encoding a single color transition.

7 is a view 7.1 - illustration of the separation of the block 2 × 4 into the left and right sub-blocks 2 × 2; view 7.2 is a block diagram of the D-method.

8 is a view 8.1 is an illustration of the position of reference points and individual derivatives in the F-method; view 8.2 is a block diagram of the F-method.

Fig.9 is a block diagram of an O-method.

Figure 10 is a block diagram of the L-method.

11 is a flowchart of a method for determining the colors of a palette of a 2 × 2 block.

12 is a general block diagram of a U-method.

13 is a view 13.1 is a flowchart of a first encoding method included in a U-method; view 13.2 is a logical coding scheme for the average value of the “smooth” color in the first coding method included in the U-method.

Fig. 14 is a general block diagram of a second encoding method included in the U-method.

FIG. 15 is a detailed block diagram of a single color encoding within a 2x4 block in accordance with a second encoding method included in the U-method.

Fig - conventions of the color components of the pixels within the 2 × 4 block used in the calculation of the encoding error.

17 is a view 17.1 - a method of dividing bits within a spatial block 2 × 4 into essential and non-essential parts; - view 17.2 - a method of partial grouping of the least significant bits when encoding the least significant bits.

The main stages of the developed coding methodology are shown in Fig.1. Each frame of a video sequence is considered as an independent color image. Each image is represented as a set of non-overlapping 2 × 4 blocks. Step 100 is to enter an image, present it in blocks and encode them in possible ways. In this case, initially 10-bit data is divided into essential and non-essential parts. Further, the essential part forming a 2 × 4 block of 8-bit values is processed in steps 101 ... 121, while the minor part forming a 2 × 4 block of 2-bit values is processed in step 123. Each non-intersecting 2 × block 4 of the original image is divided into a left 2 × 2 subblock (three channels), a 2 × 2 right subblock (three channels), and is also considered as a whole 2 × 4 block (three channels). The left sub-block is processed in steps 101, 103, 105, 107, 109. The right sub-block is processed in steps 102, 104, 106, 108, 110. The whole block is processed in steps 111-117.

After each block is subjected to coding by all methods available in a particular case, coding errors are determined and a method is selected that gives the smallest error for each block (steps 118-120). At 121, a coding error caused by separate coding of the left and right subblocks at steps 101, 103, 105, 107, 109 and 102, 104, 106, 108, 110, respectively, is determined. At step 122, the optimal method is selected (among the methods selected at steps 118-120).

At step 123, the LSBD Least Significant Bits Data is encoded. The number of bits in LSBD encoding depends on the encoding method selected.

At step 124, an output bitstream is formed, which consists of a header indicating which method is used. The size of the header depends on the particular method and can take from one to ten bits. Then the data on the least significant bits (LSBD) of the 2 × 4 block in compressed form is entered into the stream. Depending on the method selected, this data may take from three to six bits per 2 × 4 block. If necessary, the stream is replenished with complementary bits, which are designed to provide a fixed bit rate. A 2 × 2 subblock is encoded by a set of methods. The first of these methods is based on the NSW transform and is called the N-method. This method is applied at steps 101, 102. The NSW transform is a new two-dimensional transform developed based on a combination of predicative coding, a one-dimensional Haar wavelet transform, and a lifting scheme.

We will consider a 2 × 2 block in any color channel as a set of four values A, B, C, D (Figure 2.1). The set of four initial values A, B, C, D is converted into four values: {A, B, C, D} → {s, h, v, d} according to the following formulas:

h = A-B

ν = A-C B = A-h d = A-D C = A-ν

D = A-d

where s is the average value of four pixels, h, v, d are the simplest gradients in the directions: horizontal, vertical, diagonal (Figure 2.2). The block diagram of the N-method is illustrated in Figure 2.3. At step 200, the NSW transform is performed to obtain the values s, h, v, d. Initially, storing s requires eight bits, h, v, d occupy nine bits for each value: 1-bit for character storage + 8-bit value. Certain procedures are used to present this data compactly. The N-method is applied to the 2 × 2 subunit, which is part of a 2 × 4 block for three color channels. Obviously, each 2 × 2 sub-block should be encoded on average no more than 90: 6 = 15 bits. These fifteen bits should represent s, h, v, d with sufficient accuracy. In order to ensure reliability and ease of coding, the coding of the quantity s is carried out using simple uniform quantization. An accuracy of six bits is sufficient to represent the average value for the natural parts of the images - step 201. The quantization of the set of difference values {h, v, d} is based on the construction of a set of fixed quantizers similar to Lloyd-Max optimal fixed quantization - steps 202 ... 208 (see . "Quantizers for symmetric gamma distributions" in Proc. IEEE Globecom Conf, p.214-118, 1983; "The enhanced LBG algorithm" // Giuseppe Patane, Marco Russo) - [6]. Moreover, each quantizer refers to an image with a certain structure and consists of more than one value, which is the recovery level of the quantized value. The quantization process consists in identifying a suitable quantizer among the set of quantizers for {h, v, d} (step 209 after steps 202 ... 208) for one color, and then finding the corresponding recovery level for each difference (step 210). Due to the limited bit budget, the following restrictions are imposed: the set of quantizers consists of eight separate quantizers, each of which has four values: two positive and symmetrically two negative (these values are hereinafter referred to as recovery levels). Typically, recovery levels are in the range [-255 ... 255]. Moreover, the moduli of the values of the recovery levels of one quantizer are a monotonically increasing discrete function. A table with preferred recovery levels is shown below:

Table 1 Quantization table Quantizer number Id × 0 Id × 1 Id × 2 Id × 3 0 -220 -80 80 220 one -23 -four four 23 2 -200 -10 10 200 3 -80 -40 40 80 four -138 -fifty fifty 138 5 -63 -eighteen eighteen 63 6 -42 -12 12 42 7 -104 -28 28 104

The coding process of difference values {h, v, d} includes the following:

- determination of the quantizer number N from the set of quantizers, three values of which are the best approximation for {h, v, d};

- determination of recovery level indices for {h, v, d} in quantizer number N.

The described process of encoding difference values is hereinafter denoted by the process of fixed quantization according to the table (FQvT). For more generalization, it makes sense to formulate FQvT in general mathematical terms. A set of quantizers can be described as a two-dimensional matrix Q (quantization tables), including QS rows (quantizers) and IX columns (recovery or quantization levels). Suppose that you want to quantize a set of difference values {d ₁ ... d _K }. We define a general expression for calculating the error E of the approximation, which is caused by quantization of the initial set of difference values {d ₁ ... d _K }. Without loss of generality, we assume that it can depend on the set of differences {d ₁ ... d _K }, the selected recovery levels {Id ₁ ... Id _K }, as well as the selected quantizer QI. E = E ({d ₁ ... d _K }, (Id ₁ ... Id _K }, QI). Finding the optimal quantizer from the set of quantizers consists in determining the argument of the expression:

This expression mathematically describes steps 202 ... 208 and the conditional jump in Fig.2.3. From a technical point of view, it means that for each quantizer (QI) the set of difference values {d ₁ ... d _K } should be approximated by the best recovery levels that are uniquely identified by the indices {Id ^Q1 ₁ ... Id ^QI _K ). Then, among all the quantizers from the set of quantizers, only one should be selected, and this should ensure a minimum approximation error. For a given quantizer QI and a given difference quantity d _{j, the} best recovery level in the general case is determined by the expression:

It should be noted that another way to determine the best level of recovery is to compare d _j with the midpoints between adjacent levels of recovery. In this case, it is only necessary to perform the operations of "comparison". In addition, it is easy to notice that if the table is symmetric (each quantizer includes positive and negative values), then during quantization it is only necessary to consider the amplitudes of the difference values.

The function E ({d ₁ ... d _K }, {Id ₁ ... Id _K ), QI) may also include additional weights that allow us to calculate the weighted error (if, for example, difference values from color channels with different visibility are used). Obviously, the quantization table presented cannot ideally convey an arbitrary combination {h, v, d}. However, the results of the experiments indicate that for natural images, the set of values (h, v, d) is localized in certain subdomains. Based on this, the quantization table is optimized specifically for these cases.

The choice of the optimal quantizer (or the row in Table 1) is based on the calculation of the error of approximating the subblock with a certain set of {h ', v', d '}. Consider the process of quantization of difference values h, v, d as adding additive noise to the initial values:

Then the error in the intensities of pixels A, B, C, D can be estimated as follows:

If the average value of s is quantized, then the corresponding Δs must be added to ΔA in order to correctly estimate the remaining pixel errors.

Based on the studies, acceptable results can be achieved if, as a measure of the error E, use the sum of the squares of the values ΔA, ΔB, ΔC, ΔD (step 207). For each row, this value is calculated and then a minimum is selected. It should be noted that this value is essentially the standard deviation of the coding error of the subblock (the quantization error caused by the quantization of the value of s is not considered, but can also be taken into account).

For one purpose or another, simpler functional dependencies can be used to evaluate the measure of error. For example, the maximum deviation value MAX (| Δh |, | Δv |, | Δd |) is also recommended. The experimental results also confirm its effectiveness.

The second method for processing a 2 × 2 subblock is the P method, which is implemented in steps 103 and 104. The NSW transform is effectively used in encoding the natural parts of images. But some of them may consist of one vertical / horizontal element. In this case, one of the differences will be close to zero and the quantization table of the N-method will not be so effective. To overcome this problem, the quantization table is optimized for these cases. In addition, additional coding sub-modes have been added that provide efficient transmission of sub-blocks with vertical / horizontal or diagonal predominant directions, as well as homogeneous sections within a 2 × 2 sub-block. The encoding method is determined by a variable length sub-prefix (VL), which indicates what type of encoding is used. The P method uses seven submodes. They are presented in Table 2. A generalized block diagram of the P-method is shown in Figure 3.5.

table 2 VL Subprefixes for NSWPZ Method Subprefix (binary representation) Submode Name 00 Horizontal preferred direction (H structure) 01 Vertical preferred direction (V structure) one hundred h the value is close to 0 101 v value is close to 0 110 d value close to 0 1110 Diagonal preferential direction (D structure) 1111 Average value

The H structure means that two "reference points" A, C are used to encode the 2 × 2 sub-block (Figure 3.1). Each of them is encoded in seven bits (simple truncation of the least significant bit). The truncated low order is restored to zero.

V structure means that for the coding of the 2 × 2 subblock, two “reference points” A, B are used (Fig. 3.2). Each of them is encoded in seven bits (simple truncation of the least significant bit). The truncated low order is restored to zero.

D structure means that two "reference points" A, B are used to encode the 2 × 2 sub-block (Figure 3.3). Each of them is encoded with six bits (a simple truncation of the two least significant bits). Truncated low order bits are restored with a binary value of 10b. Compared with the mode of H / V structures, in this case, reference points A, B are used to restore points C, D in the following way: D = A, C = B.

The average value mode is used to encode all pixels A, B, C, D through the average value (homogeneous block), and the average value takes eight bits.

One of the modes “h / v / d value is close to zero” is used if, after the NSW transform, one of the values (h, v, d) is close to zero. Prefixes are selected based on the following matches:

100: | AB | ~ 0

101: | AC | ~ 0

110: | A-D | ~ 0

In each case, only two difference values are quantized:

h≡0; {v, d} → {Quantizer N, Iv, Id},

v≡0; {h, d} → {Quantizer N, Ih, Id},

d≡0; {h, v} → {Quantizer N, Ih, Iv}.

A preferred quantization table describing a set of quantizers for the NS WPZ method is presented below:

Table 3 Quantization table. P-method Quantizer number Id × 0 Id × 1 Id × 2 Id × 3 0 -250 -240 240 250 one 0 0 -7 7 2 -224 -9 9 224 3 -118 -66 66 118 four -94 -19 19 94 5 -220 -125 125 220 6 -185 -52 52 185 7 -39 -12 12 39

The value of s is initially encoded with six bits by simple truncation of the two least significant bits, however, in some cases, the accuracy of the representation can be improved due to more efficient coding of difference values. The encoding process includes the following steps: at step 300, a 2 × 2 subblock is encoded using structure H; in step 301, a 2 × 2 subblock is encoded using structure V; in steps 302 and 303, a 2 × 2 subunit is encoded using structure D and the average value, respectively. After that, in step 304, the errors caused by the corresponding sub block coding methods are calculated. Then, NSW conversion is performed (step 305). Among the modules of difference values (h, v, d), the smallest value is determined - step 306. Then, two nonzero difference values are encoded using the FQvT method using Table 3 and the quantization process described above (step 307). At step 308, additional free bits are determined, and based on this, the value s is encoded with an accuracy of six, seven, or eight bits. The error caused by the coding of the 2 × 2 block using the FQvT method is estimated — step 309. Among the five errors obtained in step 310, the smallest one is selected. The corresponding subprefix from Table 2 is injected into the bitstream before the compressed data - step 311.

As a result of experiments, it was found that the 6-bit accuracy of the representation of the average value in some cases is insufficient to ensure the visual identity of some parts of the image. In this regard, it is proposed to use one of the quantizers (Number 1) as an indicator. It contains both zero values and non-zero values, which are encoded by prefix codes of a special form: Id × 0 → 0b; Id × 1 → forbidden value; Id × 2 → 10b; Id × 3 → 11b. A table with recovery levels for this quantizer is shown below:

Quantizer number Id × 0 Id × 1 Id × 2 Id × 3 one 0 0 prohibited value -7 7

An algorithm for detecting the number of "Free Bits" is shown in Figure 4.1. The decoder can determine the presence of free bits in a similar way, whereby the coherence of the encoder / decoder is not violated. The correspondence of the number of free bits and the accuracy of the representation of the average value is shown in Table 4:

Table 4 Interpretation of free bits Number of free bits Representation accuracy of the "s" value 0 6 one 7 9 8

For efficient decoding, indexes encoding difference values must be included in the compressed stream before the encoded mean value information. The recommended structure is shown in Figure 4.2.

The third method of this group is called the S-method. This method is used at steps 105, 106. It is specifically designed to reproduce areas with a relatively low intensity (and, therefore, with small difference values) in all three color channels within the 2 × 2 subunit and is a modification of the N-method due to the use of a different quantization tables and other mean coding procedures.

The value of s is encoded using five bits as follows:

- take the lower six bits of the eight-bit representation of s;

- of these, the high five bits are considered as a code representation of s. When decoding, the least significant bit of the decoded value s is restored to zero.

The quantization table provides the transmission of minor changes in intensities:

Table 5 Quantization table. Small value coding Quantizer number Id × 0 Id × 1 Id × 2 Id × 3 0 -one -2 one 2 one -2 -6 2 6 2 -one -four one four 3 -four -10 four 10 four -3 -8 3 8 5 -four -twenty four twenty 6 -8 -40 8 40 7 -3 -fifteen 3 fifteen

The procedure for quantizing difference values is similar to the corresponding procedure of the N-method.

The fourth method is called the C-method. It is applied at steps 107 and 108. The method is based on the application of the NSW transform and consists in its application to all color channels simultaneously with subsequent group coding of difference values from three color channels together. The effectiveness of this method for a certain class of images is determined by the high similarity of the amplitudes of the difference values after the NSW conversion of all three channels and their combined wide dynamic range, which is a distinctive feature of individual regions of the specified class of images. Due to group coding, increasing the quantizer dynamic range (up to eight recovery levels), as well as eliminating the independence of the quantizer from the color channel, coding (quantization) accuracy is improved if the difference values of the NSW transform over the three color channels are sufficiently similar. These features can be considered as distinctive features of the claimed invention. Table 6 presents the recommended optimized set of quantizers.

Table 6 Quantization table. C method Quantizer number Id × 0 Id × 1 Id × 2 Id × 3 Id × 4 Id × 5 Id × 6 Id × 7 0 -128 -180 -213 -245 128 180 213 245 one -3 -fifteen -thirty -fifty 3 fifteen thirty fifty 2 -twenty -86 -160 -215 twenty 86 160 215 3 -fifty -82 -104 -130 fifty 82 104 130 four -90 -145 -194 -240 90 145 194 240 5 -9 -40 -73 -106 9 40 73 106 6 -163 -122 -66 -fifteen fifteen 66 122 163 7 -180 -139 -63 -25 25 63 139 180

The quantization process is very similar to the quantization process used in the N-method. The main difference is that for the whole set of difference values from the 2 × 2 subunit: {hR, hG, hB, vR, vG, vB, dR, dG, dB} one common quantizer is chosen. Each of the difference values is encoded by a 3-bit value. A total of nine differences occupy: 9 × 3 = 27 bits. The quantizer index is encoded in three bits, which ultimately gives (27 + 3) = 30 bits. Each value {sR, sG, sB} is encoded with five bits by truncating the three least significant bits. Shot discharges are restored through a binary value of 100b. In general, a 2 × 2 sub-block in three color channels is encoded using forty-five bits. To determine the optimal quantizer, a procedure similar to that used in the N-method is used. To ensure optimal visual quality, the total error of approximation of a set of difference values in all three color channels should take into account the different color sensitivity of the human visual system. In particular, the present invention proposes the use of specially defined weights that reflect the observer’s different perceptions of red, green, and blue. In this case, the error E generated by the FQvT procedure for {hR, hG, hB, vR, vG, vB, dR, dG, dB} can be expressed as follows:

Here ΔhC, ΔvC, ΔdC denote the error of approximation of h / v / d difference values after the FQvT procedure in color channel C. W _c are weight coefficients reflecting the relative visibility of each of color channels C to humans. A simpler way of estimating the error E is to calculate the sum of the weighted maximum coding errors of h / v / d difference values:

This method of estimating the error provides adequate results in terms of the visual quality of the encoded image and is recommended for use if the implementation complexity is critical.

The fifth method is referred to as the E-method. It is performed in steps 109 and 110. It is known that at the boundaries between two clear colors, standard encoding methods can cause visible distortion. In view of this, it is proposed to split the initial 2 × 4 block into four smaller microblocks with sizes of 1 × 2 pixels — step 500 in FIG. 5.1. Moreover, each 1 × 2 microblock consists of two spatial pixels, designated A, B - Fig. 5.2. Each spatial pixel consists of three color components R, G, B. For each microblock, the dominant channel is determined and then the pixel values in the dominant channel are transmitted by encoding the corresponding values in six bits; minor (non-basic) channels are encoded by transmitting average values between two values for each spatial position, using the quantization and limitation procedure - step 501. A block diagram of the coding of the microblock 1 × 2 is shown in Fig.6. At step 602, the sum of the pixel intensities for each color channel is calculated. Then a sequential comparison of the calculated amounts is carried out in order to determine the main (dominant) channel. If the sum of intensities within the red channel is greater than the corresponding sums within the green and blue channels - step 603, then a conclusion is made about the dominance of the red channel. Otherwise, if the sum of the intensities within the green channel is greater than the sum of the intensities of red and blue - step 604, a conclusion is made about the dominance of the green channel. Otherwise, if the sum of the intensities within the blue channel is greater than the sum of the intensities of red and green - step 605, a conclusion is made about the dominance of the blue channel. Otherwise, it is assumed that the dominant channel is the violent channel, and the color difference channels are minor - step 606. The pixel intensities of the dominant channel are stored in temporary variables D1 / D2. Depending on which channels are recognized as minor, certain average values are calculated - steps 608, or 609, or 610. A key feature is the method of calculating the average: averaging is performed not within each minor channel, but between two minor channels in each spatial position. For example, consider step 608. The first average value is calculated in the spatial position A between the red and blue channels. The second average value is calculated by averaging the intensities in the spatial position B in the red and blue channels. After that, temporary variables that store the values of the dominant channel are encoded with 6-bit values by truncating the two least significant bits - step 607. The truncated bits are restored with the binary value 10b. The averaged values of the minor channels are initially limited to an upper threshold of sixty-three (if the value is greater than 63, it is set to 63) - step 611. Then, the bounded values are quantized by truncating the three least significant bits - step 612. The truncated bits are restored using the binary value 100b. Additional information indicating the number of the dominant channel (two bits) is transmitted with each 1 × 2 microblock.

In addition to the methods described above, the authors developed seven encoding methods that are designed to encode a 2 × 4 block. These methods complement the set of methods described above, and in certain cases provide a more efficient transfer of the internal structure of the image.

The first method of this group will be further called the D-method. The method is applied at step 113. It is intended for those areas of the image in which there is a combination of small areas with a clearly defined structure (cross-diagonal borders) and small natural areas. The D-way method is to determine 2 × 2 sub-blocks within a 2 × 4 block that have a regular (cross-diagonal) and natural structure. A sub-block with a regular structure is encoded using a predefined pattern and template values. The template defines spatial positions for each of the template values. A sub-block with a natural structure is encoded by quantizing the pixel values, which is expressed in the truncation of the lower two bits.

The encoding process of a 2 × 4 block in a D-way is disclosed in Figs. 7.1, 7.2. At step 700, a subblock with a regular (patterned) structure is determined. For this, the template values of the left and right subblocks 2 × 2 are determined, then the errors of approximation of the left and right subblocks 2 × 2 by template coding are determined. Based on a comparison of the calculated errors, a subblock with the most pronounced regular structure is determined. The described actions are performed at steps 701, 702.

Template values are determined in accordance with the following formulas:

Here: C ₀ , C ₁ - template values, k - subblock index. k = 0 is the left sub-block, k = 1 is the right sub-block.

Error B1E (k) for approximating the subblock by template coding is calculated by the formula:

After that, the number of the sub-block to be encoded using template coding is placed in the outgoing compressed stream.

The second method will be referred to below as the F-method. It is implemented at step 111. The method is intended for the effective representation of blocks consisting of 2 × 4 pixels with a gradient structure. The method can be applied to each color channel individually. To ensure a given flow rate, as well as to maintain high accuracy of the representation of gradient transitions, the average coding points, the average value of the modules of horizontal inter-pixel differences, and also the signs of each difference are used as coding elements. The reference points, as well as the directions of the individual inter-pixel differences are shown in Fig. 8.1. A flowchart is shown in FIG. 8.2. At step 800, 6 difference values are calculated: D1, D2, D3, D4, D5, D6. At step 801, the average value of the moduli of the quantities D1 ... D6 is calculated. At step 802, the signs of each of the difference values are calculated. The average value of the difference modules, the signs of the differences, and also two reference points - A2, B2 - step 803 are transmitted as code information. Reference points are transmitted without loss. Thus, the F-method can be formulated as follows: the coding of blocks containing gradient sections is performed simultaneously or separately on three channels, performing the following operations:

- calculate six difference values D1, D2, D3, D4, D5, D6 according to the formulas: D1 = A1-A2; D2 = A2-A3; D3 = A3-A4; D4 = B1-B2; D5 = B2-B3; D6 = B3-B4;

- calculate the average value of the modulus of the difference values D1 ... D6;

- calculate the signs of each of the difference values D1 ... D6;

- as the code information transmit the average value of the modules of differences, signs of differences, as well as two reference points - A2, B2.

The third method is intended for efficient coding of high-frequency test images and will hereinafter be referred to as the M-method. It is applied at step 117. The method includes three coding techniques.

The first coding method is to apply asymmetric uniform quantization of the pixels of the source block. At the same time, each pixel of the block, or rather, the intensity of each color channel for each pixel of the block is encoded with three or four bits. The 3/4-bit codes describing the encoded pixel values are calculated using the following formulas (here, the index i denotes the ordinate axis, j denotes the abscissa axis):

where I is the pixel value of the source block; C is the bit code that encodes the corresponding pixel. The procedure for decoding pixels encoded in this way is described by the following formulas (here D is a decoded pixel):

The second coding method is based on averaging with partial coarsening of two 2 × 2 subunits within a 2 × 4 block. In this case, the code values are calculated according to the following formulas:

For this capability, the decoding procedure is performed in accordance with the following formulas:

The third coding method is based on a partial reconstruction of one of the color channels along the other two, the pixels of which are encoded by truncating several least significant bits. In particular, the red and green channels are encoded according to the formulas:

while the two extreme pixels of the blue channel, which cannot be reconstructed from the red and green channels, are encoded according to the following formulas:

,

,

The decoding procedure in this case is implemented in accordance with the expressions:

After applying to the processed block 2 × 4 of the above three sub-methods, the sub-method is determined that provides the minimum error. Its number (one of three) is transmitted in the first two bits of the compressed stream.

The fourth method used when encoding a 2 × 4 block will be defined as the dominant color encoding method or the O-method. It is applied at step 112. The method is effective if one color channel in its energy characteristics substantially exceeds other color channels within the block. In this method, four situations are analyzed: the violent channel is dominant, red, green or blue. A 2-bit index is used to indicate the dominant color number. The pixel values of the dominant color are transmitted without distortion by 8-bit values. The values of the two remaining channels are encoded by averaging within each channel and transmitting each such average value with an 8-bit value. A detailed block diagram of this method is presented in Fig.9. At step 900, the sums of pixel intensities for each of the red, green, or blue channels are calculated. Then, in steps 901, 902, 903, the values of the three calculated sums are sequentially compared with the predetermined threshold value. The decision on the number of the dominant channel is made in steps 904, 905, 906, depending on the results of the comparison in steps 901, 902, 903. If no comparisons were successful, it is assumed that the violent channel is dominant and the color difference channels are minor - step 907. Then, at step 908, the average values of the minor channels are calculated. Then the pixel intensities of the dominant channel, as well as the average values of the minor channels, are losslessly encoded - steps 909, 910. Thus, the compressed stream consists of a 2-bit index of the dominant channel, eight 8-bit pixel values of the dominant channel, and two 8-bit values of average values of minor channels.

The fifth method we called the L-mode. The method is applied in step 114 to a 2 × 4 block for all three color channels simultaneously. In this case, color channels are processed independently. Within the 2 × 4 block, for each 2 × 2 sub-block, its own two-color palette is formed. Then one of the three coding modes of the two palettes is used: the differential coding mode of the palettes, the differential coding mode of colors within the palette, or the explicit quantization of colors. The use of differential coding of palettes or differential coding of colors of palettes provides in some cases increased transmission accuracy. In addition, often one of the palettes turns out to be composed of equal colors.

In view of this, additional color processing is proposed that can increase the likelihood of using differential encoding mode. The flowchart of the method is presented in FIG. 10. At steps 1000 and 1001, two-color palettes are determined for the left and right 2 × 2 sub-blocks. At step 1003, to evaluate the possibility of using the differential encoding mode of the palettes, the logical condition is checked:

InterCondition≡ (-2≤C00-C10 <2) ∧ (-1≤C01-C11≤2),

Here: (C00, C01) the colors of the first palette (left 2 × 2 sub-block), (C10, C11) the colors of the second palette (right 2 × 2 sub-block). If this condition is true, then the colors of the first palette are transmitted without loss, while the colors of the second palette are encoded as corrections to the colors from the first - step 1011. The following actions are performed:

- two indexes are calculated, the values of which are converted to numbers greater than or equal to zero according to the expressions:

TIdx_y = C00-C10 + 2. Index range: [0 ... 3];

TIdx_x = C01-C11 + 1. Index range: [0 ... 3];

- the calculated indices are converted to a 4-bit CDVal (Cumulative Difference Value) using a table that describes the joint distribution of the TIdx_x / TIdx_y indices:

TIdx_x = 0 TIdx_x = 1 TIdx_x = 2 TIdx_x = 3 TIdx_y = 0 0 one 2 3 TIdx_y = 1 four 5 6 7 TIdx_y = 2 8 9 10 eleven TIdx_y = 3 12 13 fourteen fifteen

The use of the index distribution table is an additional distinctive coding method, since the set of index values TIdx_x / TIdx_y by itself provides the ability to encode difference information. However, the use of the table allows you to expand the range of possible mutual values of the colors of the palettes due to the potential localization and removal of rarely encountered combinations, as well as coding of others. Two additional tables are used to decode the TIdx_x / TIdx_y indices:

Cdval 0 one 2 3 four 5 6 7 8 9 10 eleven 12 13 fourteen fifteen TIdx_y -2 -2 -2 -2 -one -one -one -one 0 0 0 0 one one one one Cdval 0 one 2 3 four 5 6 7 8 9 10 eleven 12 13 fourteen fifteen TIdx_x -one 0 one 2 -one 0 one 2 -one 0 one 2 -one 0 one 2

They allow you to calculate the corresponding indices based on the four-bit value of CDVal. The restoration of the difference values is performed based on the formulas:

C00-C10 = TIdx_y-2. New range: [-2 ... 2]

C01-C11 = TIdx_x-1. New range: [-1 ... 2]

Additional processing of color palettes, increasing the likelihood of using differential coding palettes, is as follows.

- Detecting the fact of equal colors in each palette and setting the flag of equality (FP) to 1 in case of equality:

if (C00 = C01) → FP1 = 1

else → FP1 = 0

if (C10 = C11) → FP2 = 1

else → FP2 = 0

- Checking that only one palette has equal colors:

FP1 + FP2 = 1

- Suppose the colors of the first palette are equal (C00 = C01). Then the color from the second palette, which is the farthest from the original colors of the first palette, should be "inserted" into the first palette. To do this, check the condition | C00-C11 | <| C00-C10 | and the fictitious color is inserted:

If | C00-C11 | <| C00-C10 | → C00 = C10

Else C01 = C11

- If the insertion of a fictitious color has been carried out, information on this should be stored in a logical time variable so that if the differential encoding mode of the palettes is nevertheless impossible to use, restore the changed color.

The final step in the differential palette encoding mode is to generate a color map consisting of four records - step 1010. If the condition turned out to be false at step 1003, each palette is encoded independently in one of two possible modes: differential palette color encoding mode or explicit quantization mode. At 1004, InterDifbit is set to zero to provide the decoder with information about the impossibility of using the differential encoding mode of the palettes. Then the colors in each palette are reordered in such a way that the first color in each palette corresponds to the first (left - top) pixel within the corresponding 2 × 2 subunit. Then, at step 1005, a condition is checked for differential color coding within each of the palettes. For simplicity, the coding steps for the palette corresponding to the left block are shown in the flowchart only. If the condition is true, IntraDifBit is set to a single value, the first color is lossless encoded, the second color is encoded as a 2-bit correction to the first color - step 1007. If the condition turned out to be false in step 1006, each color is encoded by truncating the least three bits - step 1008 The recovery of the truncated bits is carried out with a binary value of 100b. Similar actions are performed for the second (right) palette. In conclusion, at step 1009, a color map is formed consisting of three entries. The L-method uses two types of color map: a color map, consisting of three entries, and a color map, consisting of four colors. A color map consisting of three entries means that the first pixel in the block is always encoded with the first color in the palette. Each entry in the color map is a color bit code used to encode a given pixel. The determination of the color by which the current pixel will be encoded is based on iterating over all the colors of the palette and choosing the one whose absolute difference between the intensity and the original pixel value is minimal. The determination of the colors of the palette can be implemented using various mathematical and algorithmic approaches, for example, based on the method of least squares. The flowchart of the method for determining two colors proposed in this invention is presented in Fig.11. This method is a compromise between complexity and efficiency and consists in dividing the initial values of four pixels into two groups and calculating the average values within each group, which are then considered as the colors of the palette.

The sixth method is called the U-method and is applied at step 115. The method is intended for areas of a color image in which there is a combination of complex and relatively uniform sections within a single color channel or within a single processed subunit. The method consists of two sub-methods: U1 and U2, which are used for compact representation of the initial 2 × 4 block and are optimized for possible combinations of complex and relatively uniform image sections. Based on the calculation of errors in the approximation of the source block by both methods, the optimal one is selected, that is, that provides a smaller error. The general block diagram of the U-method is presented in Fig. 12. Consider the sub-path U1. It is used for the 2 × 2 subblock, which is part of the original 2 × 4 block, and is effective if one of the three color channels of the 2 × 2 subblock can be encoded with an average value, while the remaining two channels (the so-called “active colors ”) are encoded using the NSW transform and the FQvT procedure. A block diagram of the method is shown in FIG. 13.1 (for simplicity, the block diagram shows the steps for only one sub block). At step 1300, the average values of mR, mG, mB are calculated for each color channel within the processed subunit. Then, for each color channel, the sum of the absolute differences (SAD) is calculated - step 1301. As an example, below is the formula for calculating the SAD for the red channel:

SAD _R = | A _R -mR | + | B _R -mR | + | C _R -mR | + | D _R -mR |

Then a color channel is determined, which can be optimally transmitted by an average value: SAD values of the color channels are compared with the threshold value set earlier (SMOOTH_THRESHOLD) in a fixed order: green channel, then red channel, then blue. If all SADs are greater than the threshold value, the color channel that has the minimum SAD is selected - steps 1302, 1303. It is determined whether the average value of the “uniform” color channel is small or not by comparison with another thirty-two threshold set in advance. If the average value of the “uniform” color channel is small, instead of explicitly specifying the number of this color color with a 2-bit prefix (the color channel number can be 0, 1 or 2), this 2-bit prefix is set to three and an additional 2-bit value is then transmitted explicitly indicating the number of the "uniform" color channel. Then, the average value of the “uniform” color channel is encoded with a 3-bit value by truncating the lower two bits. Otherwise, the average value is encoded with a 6-bit value by truncating the least significant two bits - steps 1304, 1305, 1306. The remaining two colors (“active colors”) within the 2 × 2 sub-block, designated below as the first and second colors, are encoded using NSW conversions and FQvT as follows (steps 1307, 1308):

- for pixels within the first and second colors within this sub-block, an NSW transform is performed. As a result, two sets of values are calculated: {s1, h1, v1, d1}, {s2, h2, v2, d2};

- {h1, v1, d1}, {h2, v2, d2} are encoded using the quantizer table - Table 8. Moreover, each set of difference values {h1, v1, d1}, {h2, v2, d2} is encoded using its own quantizer. The quantization procedure is similar to the FQvT quantization described above. The optimal quantizer is calculated based on the definition of a quantizer among eight possible ones so that a minimum approximation error is provided for each of the sets of difference values {h1, v1, d1}, {h2, v2, d2};

- The values s1, s2 are encoded in one of two modes:

- 6/6 bits: the mode is used if the average value of the “uniform” color channel is small and the bit budget allows you to encode s1, s2 by truncating the lower two bits;

- differential mode is used if the 6/6 bit mode cannot be applied, and the following steps are taken:

- The error caused by the coding s1, s2 of 5/5 bits is evaluated (truncation of the least three bits):

ErrQuant = | s1- (s1 & 0 × F8) | 0 × 04 | + | s2- (s2 & 0 × F8) | 0 × 04 |

The error caused by quantizing s1, s2 into 6/6 bits is estimated, followed by the differential representation of s2 as a 3-bit signed correction to s1:

- temp_s1 = s1 & 0 × FC; temp_s2 = s21 & 0 × FC;

- temp_s2 = temp_s2-temp_s0;

- clamp (temp_s2, -31, 31) - “restriction on two thresholds”

- temp_s2 = temp_s2 & 0 × FC

- temp_s2 = temp_s2 + temp_sl

ErrDifQuant = | s1-temp_s1 + s2-temp_s2;

- If (ErrQuant <ErrDifQuant), s1, s2 are encoded in 5/5 bit mode by truncating the least three bits; otherwise s1, s2 are encoded as: 6 bits (s1) +3 bits + sign (s2-s1).

The compressed stream format is an important aspect of the U1 sub-method:

- Nineteen bits → compressed information about the average values (explained in Fig.13.2);

- Three bits → quantizer number: for the first active color channel;

- (3 + 3 + 3) bits → indices of recovery levels of difference values {h, v, d} of the first active color channel;

- Three bits → quantizer number for the second active color channel;

- (3 + 3 + 3) bits → indices of recovery levels of difference values {h, v, d} of the second active color channel.

Thus, forty-three bits are required to encode one 2 × 2 subblock. The steps described above are performed for both 2 × 2 subunits that are part of the processed 2 × 4 block.

Table 8 Quantization table. U1 way Quantize - Set N Id × 0 Id × 1 Id × 2 Id × 3 Id × 4 Id × 5 Id × 6 Id × 7 0 0 -eighteen -34 -45 0 eighteen 34 45 one -fourteen -60 -115 -160 fourteen 60 115 160 2 -6 -38 -62 -95 6 38 62 95 3 -56 -77 -one hundred -112 56 77 one hundred 112 four -70 -107 -142 -160 70 107 142 160 5 0 -17 -32 -130 0 17 32 130 6 -53 -125 -196 -232 53 125 196 232 7 -16 -65 -one hundred -120 16 65 one hundred 120

If in the same color channel within the 2 × 4 block, the pixels of one 2 × 2 subunit form a uniform structure, while the pixels of the second 2 × 2 subunit form an uneven region, an effective coding method is the U2 sub-method, the block diagram of which is shown in Fig; a detailed flowchart of the coding process of a single color channel within a 2x4 block is shown in FIG. Each 2 × 4 color block is considered as three independent blocks in three color channels: a red 2 × 4 block, a green 2 × 4 block, and a blue 2 × 4 block. Each of them is split into two adjacent 2 × 2 subunits - step 1500. For each of them, the average approximation error is calculated - steps 1501, 1502. As an error, without loss of generality, the SAD described above was chosen. Then, at step 1503, a uniform subblock is determined with its subsequent encoding — step 1504 or 1506. The second subblock is encoded using the NSW transform and the FQvT procedure — steps 1505, 1507. The corresponding quantization table is presented in Table 9. Each difference value from {h, v , d} is encoded using a 3-bit index indicating the best recovery level within the quantizer, the number of which is also transmitted by a 3-bit value, the s-value is encoded by a 7-bit value by truncating the least significant bit, which restores etsya zero. A 1-bit index is used to indicate a subblock that is uniform. Thus, each color channel within a 2 × 4 block is encoded using 7+ (3 + 3 + 3 + 3) + 1 + 8 = 28 bits, which is eighty-four bits for the entire 2 × 4 block.

Table 9 Quantization table. U2 way Quantizer number Id × 0 Id × 1 Id × 2 Id × 3 Id × 4 Id × 5 Id × 6 Id × 7 0 -60 -141 -178 -233 60 141 178 233 one -2 -16 -28 -fifty 2 16 28 fifty 2 -10 -76 -150 -190 10 76 150 190 3 -44 -80 -108 -129 44 80 108 129 four -one hundred -150 -201 -227 one hundred 150 201 227 5 -one -2 -3 -5 one 2 3 5 6 -8 -38 -65 -88 8 38 65 88 7 -60 -110 -160 -170 60 110 160 170

The last method of this group is called the Hadamard transform coding method (H-method). The method is applied at step 116. In the H method, the original 2 × 4 block is considered as a combination of two 2 × 2 subunits, and the well-known Hadamard transform (otherwise called the integer Discrete Cosine Transform 2 × 2 - see "Multiresolution progressive image transmission using a 2 × 2 DCT ", Seock Hoon Woo, Chee Sun Won // IEEE 07803-5123-1 / 99) [7] to each of them with the subsequent specially developed FQvT encoding procedure. The conversion is described by the following formulas:

Direct conversion:

Inverse Transformation:

It is known that this transformation carries out the separation of the source signal into components with different visual significance for the observer. Based on this, the diagonal component of the transformation, dD, can be set to zero without a significant deterioration in quality. The remaining components S, dH, dV are encoded as follows:

- The S component is encoded with a 6-bit value by truncating the two least significant bits. Recovery is carried out with a binary value of 10b;

- {dH, dV} components are quantized by searching for the best quantizer and the best recovery level within the quantizer.

For each quantizer included in the set of eight quantizers, the best recovery level of the components {dH, dV} is determined based on minimizing the absolute value of the difference between the recovery level and the value of the corresponding component. The coding error of {dH, dV} components by any quantizer is calculated by summing the modulus of the difference of the corresponding recovery level and the value of each component, which is the smallest among all possible within the given quantizer. Then a quantizer is determined that provides the minimum coding error. A quantization table optimized for quantizing the {dH, dV} components is presented in Table 10.

Table 10 Quantization Table {dH, dV} Component Quantizer number Id × 0 Id × 1 Id × 2 Id × 3 Id × 4 Id × 5 Id × 6 Id × 7 0 -6 -47 -86 -106 6 47 86 106 one -twenty -thirty -41 -47 twenty thirty 41 47 2 -one -twenty -38 -48 one twenty 38 48 3 -3 -10 -16 -twenty 3 10 16 twenty four -eighteen -40 -64 -75 eighteen 40 64 75 5 -6 -16 -26 -32 6 16 26 32 6 -8 -40 -73 -89 8 40 73 89 7 -9 -thirty -52 -63 9 thirty 52 63

Each 2 × 2 subblock within one color channel is encoded using six bits to encode the average, 3-bit quantizer number and two 3-bit indices to encode the optimal recovery levels of the dH, dV components.

The choice of the encoding method for the current 2 × 4 block being processed is based on the determination of the encoding method among the entire set of methods used in this invention, which minimizes the encoding error. Based on general considerations, it follows that the effectiveness and adequacy of the measure of distortion (coding errors) can have a significant impact on the visual quality of a compressed image. Based on the analysis of various methods for assessing distortion, without limiting generality, in this invention, two methods for assessing distortion are used, a detailed description of which is given below.

It must be emphasized that the two methods described do not limit the use of other methods for assessing visual quality, but are only options for implementation. Consider a block of pixels 2 × 4 for all three color channels - Fig.16. The first way to estimate distortion is to calculate the Weighted Root Mean Square Error (B-MSE, WMSE)

Here, the symbols A, B, C, D denote pixels located in the source block, as shown in FIG. 16. Symbols A, B, C, D with an underscore indicate pixels located in a distorted (encoded) block. Wc are weights reflecting the relative importance of the color channel. It was experimentally established that acceptable results from the point of view of the visual perception of a compressed image can be obtained with the following values of the weight coefficients:

Red channel Green channel Blue channel W _c value 6 16 four

The second distortion estimation method proposed in this invention is called the Weighted Quadratic Max-Max Criterion (VKMMK, WSMMC). Consider a 2 × 2 subblock included in a 2 × 4 block — FIG. 16. For definiteness, consider the first sub-block (A1, B1, C1, D1 pixels). The WSMMC calculation consists of the following steps:

- Calculate the weighted square of the maximum pixel-by-pixel deviation between the original and encoded block over all three color channels:

- Calculate the sum of the values determined in the first step:

- Calculate the maximum among the values determined in the first step:

MMax = MAX [MaxSq _R , MaxSq _G , MaxSq _B ]

- Calculate the sum of the values obtained in the second and third steps:

WSMMC = SMax + Mmax

To calculate the WSMMC for the entire 2 × 4 block, you must add the WSMMC values calculated for both 2 × 2 subunits.

The choice of compression method consists of the following items:

- The left and right subblocks that are part of one 2 × 4 block are encoded separately by five well-known methods: N, P, C, S, E-methods. After compressing each subblock with one of the five known methods N, P, C, S, E, the error calculation is performed according to one of the methods described above, and the formation of intermediate bit streams is also carried out;

- For each of the two subunits, a method is determined that gives the minimum error - steps 118, 119;

- Based on the results of paragraph 2, for the 2 × 4 block consisting of two 2 × 2 subunits, the best combination of N, P, C, S, E methods is determined that provides the minimum combination error (error) of encoding the 2 × 4 block in the form of two subunits 2 × 2 - step 121;

- The entire 2 × 4 source block is processed by each of the seven methods D, O, M, F, L, U, H, which are optimized to compress the 2 × 4 block. In this case, after encoding the block by any of the indicated methods, the encoding error is calculated according to one of the methods described above;

- The minimum error is calculated among the errors caused by the application of seven known methods D, O, M, F, L, U, H to the original 2 × 4 block - step 120;

- The minimum error is calculated. As a result, the optimal encoding method is determined that provides the minimum error - step 122.

Each 2 × 4 block being processed consists of pixels, each of which is represented by three color components, each of which is initially represented by a 10-bit value, as a result of compression, it is encoded with exactly 100 bits. They are the so-called elementary compressed packet, which in turn consists of VLEH, compressed bits, bits representing information about low-order bits in a compact form, and complementary bits (optional).

The set of information about the possible encoding methods (based on all possible combinations within the framework of this invention), the number of compressed bits, the number of bits required for encoding the least significant bits, the VLEH length and the values of the bits included in it, as well as the number of complementary bits are presented in the table 11. The values and lengths of VLEH and, accordingly, the number of bits required for the coding of the least significant bits, as well as the number of complementary bits, depend on the implementation, however, the general principle of generating the output compressed stream is It is one of the key principles that distinguish this invention. In the table, the method name includes one or two letters. In the latter case, the name is decrypted as follows: the first letter corresponds to the method encoding the left 2 × 2 subunit, the second letter corresponds to the method encoding the 2 × 2 right subunit.

Each header can take from 1 to 10 bits; the bit number indicated in the table means the order in which the corresponding bit is placed / retrieved from the bitstream.

Low-order coding (LSB-coding) refers to an efficient method of representing the two least significant bits of the initial 10-bit pixel values within the processed 2 × 4 block through one or more color channels. The present invention provides four different LSB coding procedures.

3-bit encoding

For each color, the coarsened average value of the least significant bits is calculated within the 2 × 4 block:

Here: Ic (x, y) is the original 10-bit pixel value in color channel C at position (x, y). mLSBc consists of one bit describing the coarser average of the lower digits. The coarse averages over all three color channels together form a 3-bit number, which is encoded information about the least significant bits within a 2x4 block.

4-bit encoding

This low-order coding method is similar to 3-bit LSB coding for red and blue color channels. For the green channel, a more accurate representation of the least significant bits is used due to the separate coding of the least significant bits of the left and right subblocks:

The coarse averages of the red and blue channels, together with two bits describing the coarse averages of the least significant bits of two subunits in the green channel, together make up a 4-bit value, which is encoded information about the least significant bits within a 2 × 4 block.

5 bit encoding

This low-order coding method is similar to 3-bit LSB coding for the blue color channel. For the red and green channels, a more accurate representation of the least significant bits is used due to the separate coding of the least significant bits of the left and right subblocks:

The coarse 1-bit blue channel average, together with two bits describing the coarse average values of the least significant bits of two subunits in the red channel and two bits describing the coarser average values of the least significant bits of two subunits in the green channel, together constitute a 5-bit value, which is coded information about the least significant bits within the block 2 × 4.

6-bit encoding

For each color channel, two bits are used to represent coarse information about the average values of the least significant bits of the subblocks within the original 2 × 4 block:

In this method, a 6-bit value, which is encoded information about the least significant bits within a 2 × 4 block, is formed by combining three 2-bit values corresponding to coarsened information about the average value of the least significant bits.

The restoration of the corresponding parts of the two least significant bit planes is based on the receipt of one or two bits for each color channel and the formation of an average value for two bit planes over the entire 2 × 4 block or the formation of two average values for two bit planes within the left and right subunits of 2 × 2 . The average value (Middle Value) is calculated according to the formula:

For illustrative purposes, figures 17.1 and 17.2 show the splitting of the original 10-bit values within a 2 × 4 block into two parts: 8-bit and 2-bit, as well as the splitting of the two least significant bit planes into two parts: left and right.

The invention can be used in various industrial products that are used for processing high-quality color images of high resolution. The main field of application is TV broadcasting and TV playback of high-quality video stream. The use of this invention in consumer electronics products can lead to a significant reduction in the requirements for internal hardware resources, such as memory, data buses, without reducing the quality characteristics of the products, which in turn will reduce associated costs. In addition, ultra-high-definition television, as well as displays for mobile devices, can be included in the scope. In addition, the claimed invention can be used as part of a complex video coding system for the intermediate storage of processed images without visual loss in order to reduce the amount of memory required.

Claims (25)

1. A method for compact bit representation of a high-quality color video sequence, which consists in block coding of each color video frame of a video sequence, in which each color pixel is represented by three color components, each of which is initially encoded with ten bits, characterized in that the encoding is carried out by breaking the original color video frame into non-overlapping spatial blocks of more than one horizontal pixel and two vertical pixels or, the subsequent division of the bit representation of each of the color components of the pixel into the most significant part, consisting of more than one high-order bit, and the low-order part, consisting of at least one low-order bit, in the sum forming the original ten bits, separate encoding of the high and low parts, and the coding of the older part is carried out by applying more than one coding method, each of which takes into account inter-pixel communications only within the processed spatial block and is used to reduce the number of bits required to represent the processed data, evaluate the encoding error, select the encoding method that gives the smallest error, send data on the encoding method by transmitting the prefix code, encode the younger part, which is done by averaging more than one value included in the younger part, the sizes of the averaging regions within the younger part depend on the selected encoding method of the senior part, and by choosing the encoding method of the senior part, Chiva smallest error, the encoding method minor portion size prefix code indicating the encoding method used, and the amount of complementary non-data bits is set beforehand fixed number of bits required for a compact representation of the original color space block. 2. The method according to claim 1, characterized in that the video sequence is represented as a set of frames, each frame being processed as an image, and each image represented as a set of non-overlapping blocks of 2 × 4 color pixels, each of which consists of three 10- bit values, and each such block as a result of compression is represented by a bit packet with a predetermined fixed number of bits less than the number of bits required to store the original block in uncompressed form. 3. The method according to claim 1, characterized in that the highest eight bits of each component of the color pixels of the processed frame are considered and processed as blocks of 2 × 4 pixels. 4. The method according to claim 1, characterized in that the high eight bits of each component of the color pixels of the processed frame are considered and processed as blocks of 2 × 4 pixels, each of which in turn is represented as two subblocks of 2 × 2 pixels. 5. The method according to claim 4, characterized in that the processing of each of the 2 × 2 pixel subunits is carried out by the optimal method selected independently from the set of processing methods of the 2 × 2 subunit based on the results of evaluating the coding error of this block. 6. The method according to claim 3, characterized in that during the processing of each block with a size of 2 × 4 pixels, encoding is performed by all methods available in a particular case and then a coding method is selected that gives the minimum error, and the choice is made by comparing the encoding error in the optimal block method 2 × 4 pixels in size with the total error of this block represented as two subblocks of 2 × 2 pixels in size. 7. The method according to claim 4, characterized in that in the 2 × 4 block, 2 × 2 subunits are identified that correspond to a relatively regular structure, i.e. belonging mainly to the background and irregular structure, i.e. belonging mainly to a natural image, and the detection operation consists in performing operations in the following steps:
- find a block with a regular structure by determining the template values for the left and right 2 × 2 subunit, determining errors caused by encoding the left and right 2 × 2 subunit with a regular structure, comparing the template values and choosing a block with a predominantly regular structure, while the template values structures in two colors are calculated by the following formulas:

where C ₀ (k), C ₁ (k) are the template values for the left sub-block 2 × 2 (k = 0) and the right sub-block (k = 1);
- BlE (k) errors caused by the coding of the left and right 2 × 2 subunits with a regular structure are calculated by the formula:

then the number of the sub-block to be encoded using template coding is placed in the outgoing compressed stream with the subsequent storage of template values; the pixel values of the unencoded subblock are encoded by truncating the lower two bits. 8. The method according to claim 3, characterized in that the coding of the blocks is carried out in parallel with three sub-methods, namely:
- a method based on a sequential color shift and consisting of the following steps:
- the intensity for each color channel for each pixel of the block is encoded with three or four bits, calculated by the following formulas:

where if denotes the conditional operation IF, else denotes the conditional operation ELSE, which is logically related to the if operation used above, the index i denotes the ordinate axis, j denotes the abscissa axis, I is the pixel value of the original block, C is the bit code encoding the corresponding pixel, when the procedure for decoding pixels encoded in this way is described by the following formulas (here D is a decoded pixel):

- a coding method based on averaging with partial coarsening of two 2 × 2 subunits within a 2 × 4 block, in which the code values are calculated according to the following formulas:

while decoding is performed in accordance with the following formulas:

- an encoding method based on the partial reconstruction of one of the color channels along the other two, the pixels of which are encoded by truncating the few least significant bits; in particular, the red and green channels are encoded according to the formulas:

and the two extreme pixels of the blue channel, which cannot be reconstructed from the red and green channels, are encoded according to the following formulas:

,

,
moreover, the decoding procedure is implemented in accordance with the expressions:

after applying to the block being processed 2 × 4 of the three methods described above, the method providing the minimum error is determined and its number is transmitted in the first two bits of the compressed stream. 9. The method according to claim 3, characterized in that when processing blocks with uneven integrated color channel saturation, the dominant color method is used, which consists in analyzing four situations, namely, checking which of the channels is dominant: luminance channel, red channel, green channel or blue channel, moreover, a 2-bit index is used as the index of the dominant color number, and the pixel values of the dominant color are transmitted without distortion by 8-bit values, while the values of the two remaining channels are encoded by averaging within each of the channels and transmitting each such average value with an 8-bit value. 10. The method according to claim 3, characterized in that when processing 2 × 4 blocks, a color palette is formed, and all three color channels are processed simultaneously and independently, while for each 2 × 2 subunit they form their own two-color palette using any of the known methods, the preferred method consists in dividing the initial values of the four pixels of the 2 × 2 subunit into two groups relative to the average and calculating the average values within each group, which are then considered as the colors of the palette, and one of the three p presses coding, namely the differential encoding mode palette of colors of the differential encoding mode within a palette or explicit quantization of colors, which initially evaluate the possibility of using differential encoding palettes mode by checking the first logical condition:
InterCondition≡ (-2≤C00-С10 <2) ∧ (-1≤C01-С11≤2),
where (C00, C01) are the colors of the first palette related to the left sub-block 2 × 2, (C10, C11) are the colors of the second palette related to the right sub-block 2 × 2, taking into account the fact that it is preferable to use preliminary processing of colors of each palette , which consists in verifying that only one palette of the two has equal colors, in which case replacing one of them with a color from the other palette, which is the most distant in the sense of the absolute value of the values of equal colors, in this case, if the first logical the condition is met, then the colors first palette transmitted lossless, while the second color palette encoded as corrections to the colors of the first panel, proceeding from the following:
- calculate two indexes whose values are converted to numbers greater than or equal to zero according to the expressions:
TIdx-y = С00-С10 + 2, where the range of the index: [0 ... 3],
TIdx-x = C01-C11 + 1, where the range of the index: [0 ... 3];
- convert the calculated indices into a 4-bit generalized differential value (CDVal) using a table describing the joint distribution of the TIdx-x / TIdx-y indices:

further decoding of TIdx-x / TIdx-y indices is performed on the basis of two additional tables:

and the restoration of the value of the differences is performed based on the formulas:
C00-C10 = TIdx-y-2 with a new range: [-2 ... 2],
C01-C11 = TIdx-x-1 with a new range: [-1 ... 2];
if the first logical condition is not met, the color change is canceled if it has been implemented, and then the colors in each palette are reordered in such a way that the first color in each palette corresponds to the first (left-top) pixel within the corresponding 2 × 2 subunit, after which the second logical condition for differential color coding for each palette is checked, which consists in analyzing the absolute difference of the color values in each palette, and if the catch for the palette under consideration is true, the coding of the first color is lossless, and the second color is in the form of a 2-bit correction to the first color; otherwise, the value of each color is encoded by truncating the lower three digits, while the restoration of the truncated bits is carried out by a binary value 100b; in conclusion, a color map is formed, which can consist of 3 entries or 4 entries, while a color map consisting of 3 entries means that the first color in the sequence corresponds to the first pixel in the subblock, while to ensure a given bit budget, a color map of 4 entries is preferable to use for the differential encoding mode of the palettes. 11. The method according to claim 3, characterized in that when processing blocks for areas of a color image in which there is a combination of complex and relatively uniform sections within the same color channel, or within one processed sub-block, two methods are used: U1 and U2, providing the possibility of a compact representation of the initial 2 × 4 block and optimized for possible combinations of complex and relatively uniform sections of the image, while on the basis of calculating the approximation errors of the initial block, the wallpaper my sub-methods select the optimal sub-method that provides the smallest error, and sub-method U1 is applied to each 2 × 2 subblock that is part of the original 2 × 4 block in the case when one of the three color channels of the 2 × 2 subblock is suitable for encoding with an average value, while the two remaining channels are encoded using Non-Separable Wavelet (NSW) transforms and Fixed Quantization via Table (FQvT) procedures; sub-block U2 is applied to a 2 × 4 block within one color channel in the case when one 2 × 2 subblock within the channel in question is suitable for medium coding, while the other 2 × 2 subblock within the channel in question can be effectively encoded with using NSW conversion and FQvT procedures. 12. The method according to claim 3, characterized in that when processing the blocks, an encoding method is used, which means that the original 2 × 4 block is considered as a combination of two 2 × 2 subunits, each of which is considered as a set of four values A, B , C, D, and the Hadamard transform is applied to each of the subunits, followed by
FQvT encoding procedure, while the conversion is described by the following formulas:
direct conversion:

inverse transformation:

while the remaining components S, dH, dV are encoded as follows:
- The S component is encoded with a 6-bit value by truncating the two least significant bits, and the restoration is performed with a binary value of 10b;
- {dH, dV} components are quantized by searching for the best quantizer and the best recovery level within the quantizer,
moreover, for each quantizer included in the set of eight quantizers, determine the best recovery level of the components {dH, dV}, based on minimizing the modulus of the difference between the recovery level and the value of the corresponding component; the coding error {dH, dV} of the components by any quantizer is calculated by summing the modulus of the difference of the corresponding recovery level and the value of each of the components, which is the smallest among all possible within the given quantizer; then determine the quantizer, which provides the minimum encoding error, using the following quantization table, optimized to quantize the {dH, dV} components:

wherein each 2 × 2 subblock within the same color channel is encoded using six bits for encoding the average, 3-bit quantizer number and two 3-bit indices for encoding the optimal recovery levels of the dH, dV components. 13. The method according to claim 4, characterized in that during processing each 2 × 2 subunit is considered as a set of four values A, B, C, D, to which the NSW transform is applied, obtaining four values: {A, B, C , D} → {s, h, v, d} according to the following formulas:
h = AB

v = AC B = ah d = AD C = av

DA-d
where s is the average of four pixels, h, v, d are the simplest gradients in the directions: horizontal, vertical, diagonal, with each 2 × 2 sub-block being encoded on average with 15 bits, and the value of s is encoded using simple uniform quantization with accuracy an average of 6 bits; the quantization of the set of difference values {h, v, d} is carried out by constructing a set of fixed quantizers, each quantizer refers to an image with a certain structure and consists of more than one value, which is the recovery level of the quantized value, and the quantization process is as follows: a suitable quantizer among the set of quantizers for {h, v, d} for one color, and then finding for each difference the corresponding recovery level, the following boundedness: set of quantizers consists of eight quantizers, each having four values, two positive and two negative symmetrical, which operate recovery levels functions in accordance with the table:

the process of encoding difference values {h, v, d} includes the following operations:
- determination of the quantizer number N from the set of quantizers, three values of which are the best approximation for {h, v, d};
- determination of recovery level indices for {h, v, d} in quantizer number N. 14. The method according to claim 4, characterized in that when processing the 2 × 2 subunits, the NSW transform and the optimized quantization table are used

in this case, an additional seven coding sub-modes are used for subblocks with vertical / horizontal or diagonal preferred directions, as well as homogeneous sections within a 2 × 2 block, the coding method being determined by a variable length subprefix (VL), which indicates what type of coding is used

where H structure means that to encode the 2 × 2 subblock, two "reference points" A, C are used, each of which is encoded with seven bits, performing a simple truncation of the least significant bit, the truncated least significant bit being restored to zero; The V structure means that two "reference points" A, B are used to encode the 2 × 2 subunit, each of which is encoded with seven bits, performing a simple truncation of the least significant bit, the truncated least significant bit being restored to zero; The D structure means that for encoding the 2 × 2 subblock, two “reference points” A, B are used, each of which is encoded with six bits, performing a simple truncation of the two least significant bits, and the truncated lower bits are restored with a binary value of 10b, while the lower of the truncated bits is restored with a value of 0, while the oldest truncated bit is restored with a value of 1; reference points A, B are used to restore points C, D in the following way: D = A, C = B; the average value mode is used to encode all pixels A, B, C, D with the average value, obtaining a uniform block, and the average value takes eight bits; the “h / v / d value is close to zero” mode is used if, after the NSW conversion, one of (h, v, d) is close to zero, and the prefixes are selected based on the following correspondences: 100 | AB | ~ 0; 101 | A-C | ~ 0; 110: | AD \ ~ 0,
in this case, in each case, only two difference values are quantized:
h is 0; {v, d} → {Quantizer N, Iv, Id}
v is 0; {h, d} → {Quantizer N, Ih, Id}
d is 0; {h, v} → {Quantizer N, Ih, Iv}, and s is encoded with 6, 7 or 8 bits; the encoding process of the subunit includes the following steps: a subunit of size 2 × 2 is encoded using the structure H; then a 2 × 2 subblock is encoded using structure V; then a 2 × 2 subblock is encoded, respectively, using the D structure and the average value, followed by calculation of errors caused by the corresponding subblock coding methods, after which the NSW conversion is performed and the smallest value is determined among the difference value modules (h, v, d), then two nonzero difference values are encoded using the FQvT method using the table

and the quantization process described above, then determine 20 the number of free bits by initially setting the number of free bits equal to 0, analyzing the number of the selected quantizer, and if it turned out to be 1, sequential comparison with 0 indexes of recovery levels for each of the two non-zero difference values and increase by 1 number of free bits in case of equality of the recovery level index in each of the two comparisons and, based on this, coding the quantity s with an accuracy of six, seven or eight bits, respectively obstacle to the table:

then, the error caused by the coding of the 2 × 2 sub-block is estimated using the FQvT method, and the smallest is selected from the five errors obtained, and the corresponding sub-prefix is introduced into the bitstream before the compressed data. 15. The method according to claim 4, characterized in that when processing the blocks, the NSW transform is applied to all color channels simultaneously with the subsequent group coding of difference values from three color channels together in accordance with the table of an optimized set of quantizers:

moreover, for the whole set of difference values from the 2 × 2 subunit: {hR, hG, hB, vR, vG, vB, dR, dG, dB}, one common quantizer is selected and each of the difference values is encoded using a 3-bit value so that a total of nine differences occupy: 9 × 3 = 27 bits, while the quantizer index is encoded with three bits, which ultimately gives (27 + 3) = 30 bits, and each value from {sR, sG, sB} is encoded into five bits by truncation of the three least significant bits, and the restoration of the truncated bits is carried out through the binary value 100b so that the two least truncated the series are restored with 0 value, and the oldest of the truncated ones is restored with 1 value, while optimal visual quality is ensured by using predetermined weights that reflect different observer perceptions of red, green and blue colors, which are taken into account when calculating the error E generated by the FQvT procedure for {hR , hG, hB, vR, vG, vB, dR, dG, dB}, expressed as follows:

where ΔhC, ΔvC, ΔdC denote the error of approximation of h / v / d difference values after the FQvT procedure in the color channel C; W _c - weighting coefficients reflecting the relative visibility of each of the color channels С∈ {R, G, В} for a person. 16. The method according to clause 15, characterized in that the error E is estimated by calculating the sum of the weighted maximum coding errors h / v / d of the difference values:

17. The method according to claim 3, characterized in that when processing the blocks, the original 2 × 4 block is split into 4 smaller microblocks with sizes of 1 × 2 pixels, each 1 × 2 microblock consisting of two spatial pixels A and B, each spatial pixel consists of three color components R, G, B; for each microblock, the dominant channel is determined and then the pixel values in the dominant channel are transmitted by encoding the corresponding values in six bits; minority channels are encoded by transmitting average values between two values for each spatial position, using the quantization and limitation procedure, namely, that the pixel intensities of the dominant channel are stored in temporary variables D1 / D2 and depending on which channels are recognized as minor, certain average values are calculated, and averaging is performed between two non-main channels in each spatial position, then temporary variables that store the values Nia dominant channel, 6-bit encoded values by truncating the two least significant bits, and subsequent recovery of truncated bits of binary value is performed 10b; the averaged values of non-primary channels are first limited by an upper threshold of 63, then the bounded values are quantized by truncating the three least significant bits, and the truncated bits are restored with a binary value of 100b, and additional information indicating the 2-bit number of the dominant channel is transmitted with each 1 × 2 microblock. 18. The method according to claim 4, characterized in that when processing the 2 × 2 subblock to reproduce sections with small difference values in all three color channels within the 2 × 2 subblock, an alternative medium coding procedure is used, namely, the value s is encoded using five bits as follows:
- take the lower six bits of the eight-bit representation of s;
- of these, the high five bits are considered as a code representation of s;
when decoding, the least significant bit of the decoded value s is restored to zero; while the transmission of minor changes in intensities is ensured by applying the following quantization table:

19. The method according to claim 3, characterized in that each processed 2 × 4 block is considered to be composed of pixels, each of which is represented by three color components, each of which is initially represented by a 10-bit value, and as a result of compression, the block is encoded with 100 bits , which are an elementary compressed packet, which in turn consists of a Variable Length Encoding Header (VLEH), compressed bits, bits representing information about low-order bits in a compact form, and possible complementary bits, while the summary information about possible s coding methods, the number of compressed bits the number of bits required to encode the LSBs, and the length VLEH bit values included in it, and also on the number of complementary bits are in the following table:

where the identification of the method used is carried out in one or two letters, and in the latter case, the method is decrypted as follows: the first letter corresponds to the method encoding the left 2 × 2 subunit, the second letter corresponds to the method encoding the 2 × 2 right subunit, one letter corresponds to the encoding method a 2 × 4 block, with each VLEH header occupying from 1 to 10 bits; the bit number indicated in the table means the order in which the corresponding bit is placed / retrieved from the bitstream. 20. The method according to claim 1, characterized in that when choosing the optimal encoding method for the current 2 × 4 block being processed, the effectiveness of the distortion estimation method is checked based on the Weighted Quadratic Max-Max Criteria (WSMMC) and consisting in the following:
- calculate the weighted square of the maximum pixel-by-pixel deviation between the original and encoded block over all three color channels:

- calculate the sum of the values determined in the first step:

- calculate the maximum among the values determined in the first step:
MMax = Max [MaxSq _R , MaxSq _G , MaxSq _B ]
- calculate the sum of the values obtained in the second and third steps: WSMMC = SMax + Mmax,
Then, the WSMMC values calculated for both 2 × 2 subunits are added, and the choice of the optimal encoding method reduces to the following steps:
- the left and right subblocks, which are part of one 2 × 4 block, are encoded separately by five known methods: N, P, C, S, E-methods and after compressing each subblock with one of the five known methods N, P, C, S, E perform error calculation according to one of the claimed methods, and also carry out the formation of intermediate bit streams;
- for each of the two subunits determine the method that gives the minimum error;
- based on the results of the previous step, for the 2 × 4 block consisting of two 2 × 2 subunits, determine the best combination of N, P, C, S, E methods, which provides the minimum combination coding error of the 2 × 4 block in the form of two 2 × 2 subunits ;
- the entire 2 × 4 source block is processed by each of the seven methods D, O, M, F, L, U, H, which are optimized for compressing the 2 × 4 block, and after encoding the block using any of these methods, an encoding error is calculated according to one of the methods claimed herein;
- calculate the minimum error among the errors caused by the application of the seven methods D, O, M, F, L, U, H stated here to the original 2 × 4 block;
- calculate the minimum error, on the basis of which determine the optimal encoding method that provides the minimum error. 21. The method according to claim 1, characterized in that the least significant bit coding (LSB) is applied to the two least significant bits of the initial ten-bit pixel values within the processed 2 × 4 block through one or more color channels. 22. The method according to item 21, characterized in that perform 3-bit encoding, which consists in the fact that for each color they calculate the coarse average value of the least significant bits within a 2 × 4 block:

where Ic (x, y) is the initial ten-bit pixel value in the color channel C at position (x, y); mLSB _C consists of one bit describing the coarse average of the least significant bits, while the coarse averages of all three color channels together form a 3-bit number, which is encoded information about the least significant bits within a 2 × 4 block. 23. The method according to item 21, characterized in that perform 4-bit encoding, which consists in the fact that for the green channel use a more accurate representation of the least significant bits by separately encoding the least significant bits of the left and right subblocks:

while the coarse averages over the red and blue channels in conjunction with two bits describing the coarse averages of the least significant bits of two subunits, in the green channel together make up a four-bit value, which is encoded information about the least significant bits within a 2 × 4 block. 24. The method according to claim 20, characterized in that 5-bit encoding is performed, namely, that for the red and green channels a more accurate representation of the least significant bits is used due to the separate encoding of the least significant bits of the left and right subblocks:

while the coarse one-bit average over the blue channel in conjunction with two bits describing the coarse average values of the least significant bits of two subunits in the red channel and two bits describing the coarsened average values of the least significant bits of two subunits in the green channel together make up a 5-bit value, which is encoded information about the least significant bits within the 2 × 4 block. 25. The method according to claim 20, characterized in that 6-bit encoding is carried out, consisting in the fact that for each color channel two bits are used to represent coarse information about the average values of the least significant bits of the subunits within the original 2 × 4 block:

wherein a 6-bit value, which is encoded information about the least significant bits within the 2 × 4 block, is formed by combining three two-bit values corresponding to the coarsened information about the average value of the least significant bits, and the restoration of the corresponding parts of the two least significant bit planes is based on the receipt of one or two bits for each color channel and the formation of the average value for two bit planes throughout the block 2 × 4 or the formation of two average values for two bit planes within l second and right subblocks 2 × 2, and the average value (Middle Value) is calculated according to the formula:

Images