RU2295839C2 - Method for compressing digital images and device for realization of said method - Google Patents

Abstract

FIELD: processing of digital images, possible use in systems for capturing and compressing images, for example, photo-video-cameras.

SUBSTANCE: for known method of compression of digital images, including serial usage of operations for dividing unprocessed digital data, received from image capturing device, on a set of channels, with their following direct color transformation, wavelet transformation and quantization, suggested are changed rules of direct color transformation of channels, allowing more complete correlation between digital channels of image, resulting in possible compression of data to lesser size. Also, during processing of Byer's mosaic images suggested compression method allows transition to YCbCr color representation system, for which quantization coefficients are known. Usage of these coefficients results in production of compressed image of lesser size in comparison to quantization of original R, G, B channels. Suggested also is device for realization of method.

EFFECT: increased degree of compression of digital images.

2 cl, 3 dwg

Description

The present invention relates to digital image processing and can be used in devices that compress digital images obtained using cameras and video cameras.

Due to the rapid development and implementation of photo-video shooting systems, devices for high-quality digital image capture are currently undergoing great development. Typically, such devices do not capture full-color, but the so-called Bayer mosaic image (BMI). Increasing the resolution of image capture devices leads to an increase in image size and, accordingly, requires large amounts of memory to store this data. The development of subsystems of information storage devices and an increase in the bandwidth of data transmission channels only partially solve this problem; therefore, the development of BMI compression methods can significantly reduce the requirements for these subsystems and transmission channels.

A description of the invention is known to US patent No. 6154493 "Compression of color images based on a 2-dimensional discrete wavelet transform yielding a perceptually lossless image", IPC H 04 N 7/12, containing a method for compressing digital images, which consists of the separation of raw digital data received from the device for capturing images on a variety of color channels, including color difference. In this case, to eliminate mutual correlation in the formation of differential channels, the direct difference of color channels is used. Each of the channels thus obtained is subjected to two-dimensional discrete wavelet transform and quantization, which allows image compression without visual loss. A device for implementing this method includes an image capture and preprocessing unit, a channel generator, a compression unit, and an information storage unit. The specified method and its implementing device can be used to compress the BMI. The quantization coefficients used depend on the color channel being processed and the sub-band of wavelet coefficients.

However, the indicated compression method and device does not allow efficient data compression, since the correlation between color channels is not completely eliminated, which leads to multiple compression of the same information and, as a result, to a larger size of the compressed data.

In addition, when processing BMI, this compression method does not allow switching to the YCbCr color rendering system, for which quantization coefficients are known, the use of which results in a smaller compressed image.

The technical result of the present invention is data compression, with the effective elimination of cross-correlation between color channels, which gives a smaller size of the compressed data. When applied to the BMI, it allows you to switch to the YCbCr color rendering system, which makes it possible to use the quantization coefficients developed for this color rendering system.

This result is achieved due to the fact that in the well-known method of compressing digital images, which consists in sequentially applying the operations of dividing the raw digital data received from the image capture device into multiple channels, with their subsequent direct color conversion, wavelet conversion and quantization, the following are proposed rules for direct color conversion of channels:

- receiving four full-sized expanded matrices (obtaining matrices from a mosaic image is shown in Fig. 1) from three color channels, the first of which contains the values of the first channel, the second and third contain the values of the second channel, the fourth - the values from the third channel,

- using interpolation of color matrices containing the values of the first and third color channels, data is obtained at points where the values of the second color channel are known,

- convert the values in the second and third matrices (C2A ', C2B') using the interpolated values of the first and fourth matrices (C1i ', C3i') into new values (YA and YB) according to the formula

- where a1, a2, a3 are real coefficients,

- using interpolation of color matrices containing the values of the second color channel, data is obtained at points where the values of the first and third color channels are known,

- convert the values of the first and fourth matrices (C1 ', C3') using the interpolated values of the second channel (Y ') into the new values (DC1, DC2) according to the formula:

DC1 = b1 * (C1'-Y ')

DC3 = b2 * (C3'-Y ')

where b1, b2 are real coefficients,

- from the values of the second matrix, the values of the third matrix are calculated, after which the corrections from the calculated values are recorded in the third matrix, and

- from the converted full-sized extended matrices go back to the color channels, three of which are differential.

The image received from the sensor usually has a mosaic structure, when at each point the value of only one color is known, and the values of the other two colors are unknown. The set of known colors at each point can be different, the most common is the so-called Bayer set of color filters (BSCS). Thus, the color channels that make up the image are sparse. The mosaic image obtained with the help of the BSCC is the Bayer mosaic image. In addition, in the BMI the number of points at which the green values are known is twice as many as the number of points with red and blue, therefore the green channel is divided into two subchannels, receiving four channels - two green, one red and one blue. The values of these channels can be written into special matrices C1 ', C2A', C2B ', C3', as indicated in figure 1. This is done for the convenience of software and hardware implementations of the subsequent color channel conversion algorithm. In the case of the BNCC, C1 is the red channel, C2 is the green channel, C3 is the blue channel. To go to the YCbCr system at each point in the image, you must have all three color components:

α _R = 0.299; α _B = 0.114; α _G = 0.587

Y = α _R · R + α _G · G + α _B · В

It follows from the transformation formulas that it makes sense to calculate the Y component at green points, since the green component of the RGB system (for the case of the BSCS) is the one with the highest weight, so the relative error in calculating this particular component will be the smallest. Similarly, instead of the components R and B, it makes sense to calculate the components Cr and Cb. To calculate red and blue colors at points with known green color, linear interpolation can be applied using neighboring points with known red and blue colors (each green point has two adjacent red and blue ones; half the sum is taken to calculate the red and blue values in the green point values from adjacent red and blue dots). After receiving the missing color components in green dots, the Y component (brightness) is calculated. Since there are two matrices with green color channel values for distinguishing the brightness values belonging to these matrices, we denote them YA for C2A 'and YB for C2B'. The resulting brightness values are interpolated into blue and red points (each such point has four adjacent brightness values diagonally, the interpolated value is a quarter of the sum of neighboring brightnesses). The color values Cr and Cb are calculated from the interpolated brightness values and available red or blue colors. The color values obtained from the matrices C1 ', C3' are denoted by DC1 and DC3.

The values in the matrices C2A 'and C2B' in Fig. 1 (before color conversion, these are the values of the green channel, and after brightness) are strongly correlated with each other. To eliminate the correlation, the values of the matrix C2B 'can be predicted using the values of the matrix C2A' and only corrections from the predicted values can be stored in the matrix C2B '. In the simplest case, it is possible to use bilinear interpolation for prediction, that is, as the interpolated value of the matrix C2B ', use a quarter of the sum of four adjacent values of the matrix C2A', that is, values whose position indices in the matrix C2A 'differ no more than one from the predicted indices values in the matrix C2B '. The resulting correction values are denoted by dY.

To maintain the average signal level in the green channel (after color conversion, the luminance channel) for the matrix C2A 'using the corrections stored in the matrix C2B', the following operation is performed: one eighth of the sum of the nearest neighboring corrections stored in the matrix C2A 'is added matrix C2B '. The operation of preserving the average can be any, subject to the following conditions:

The obtained updated values of the matrix C2A 'will be denoted by YA'.

Thus, we obtain from the matrices C1 ', C2A', C2B ', C3', respectively, a matrix with color values Cr - DC1, a matrix with brightness values YA ', a correction matrix dY, a matrix with color values Cb - DC3.

After the conversions are completed, they pass from the matrices to the channels: C1T, C2AT, C2BT and C3T. In this case, the channel values must be selected from the positions of the matrices corresponding to the positions with nonzero values of the original matrices. Of these channels, three are differential.

The formal description of the inverse color channel conversion is as follows:

1. Obtaining from color channels: C1T, C2AT, C2BT and C3T full-sized extended matrices DC1, YA ', dY, DC3.

2. Cancellation of the operation of storing the average for YA 'by applying to the matrix obtained by the elementwise addition of YA' and dY, a filter with the following impulse response:

followed by decimation of the resulting matrix (all positions, except those in which initially there were nonzero values of YA ', are filled with zeros). Thus, we obtain the matrix YA.

3. The construction of predictors for nonzero values of the matrix YB from nonzero values of the matrix YA is achieved by applying to the matrix YA in positions corresponding to nonzero values of the matrix YB an interpolating filter with the following impulse response:

We denote the resulting matrix by R.

Obtaining the values of YB: YB = dY + P (the operation of elementwise addition, performed only for nonzero values of the matrix dY).

4. Obtaining interpolated values for the matrices YA and YB at the positions of nonzero values of the matrices DC1 and DC3 is achieved by applying a filter with the following impulse response to the sum of the matrices YA and YB:

Denote the resulting matrix as Y '.

5. Obtaining matrices C1 'and C3' from the color difference matrices DC1 and DC3 at the locations of nonzero values of the matrices DC1 and DC3:

where b1, b2 are some real coefficients, which in the case of BMI are as follows: b1 = 0.7133; b2 = 0.5643.

6. Using linear interpolation of the matrices C1 'and C3', data is obtained at the positions where the nonzero values YA and YB are located. The obtained interpolated values of the matrix are denoted by C1'i and C3'i.

7. The conversion of nonzero values of YA and YB to C2A ', C2B' according to the following formulas:

where a1, a2, a3 are some real coefficients, which in the case of BMI are as follows: a1 = 0.299; A2 = 0.587; a3 = 0.114.

8. Obtaining from the full-sized expanded matrices C1 ', C2A', C2B ', C3' of the restored color channels: C1, C2 and C3 (Fig. 1).

The specified compression method is implemented by the device.

Figure 1 shows the structure of full-sized extended color matrices.

Figure 2 presents the General structural diagram of a device for compressing digital images.

Figure 3 presents the structural diagram of the block conversion of color channels.

A device that implements the above-described method for compressing color images, for example, BMI, is shown in FIG. 2 and consists of a digital image capture and preprocessing unit (block 2.1), which is connected to a color channel conversion unit (block 2.2) associated with the compression block (block 2.3) and then the storage and transmission unit of compressed data (block 2.4).

The color channel conversion unit (FIG. 3) includes the following blocks. The extended matrix generation block (block 3.1) converts the input color channels C1, C2, C3 and has three inputs: C1, C2, C3. The output C1 'of the block 3.1 is connected to the input of the block for calculating the interpolated values C1 (block 3.2), and the output C3' to the input of the block for calculating the interpolated values C3 (block 3.3). The output C1'i of block 2 is connected to the block for calculating the values of YA (block 3.4) and the block for calculating the values of YB (block 3.5). The output of block C3'i is connected to block 3.4 and 3.5. The output of block 3.1 C2A 'is connected to block 3.4, and the output of block 3.1 C2B' is connected to block 3.5. The output of block 3.4 YA is connected to the block for calculating the values of Y '(block 3.6), as well as to the block for calculating the corrections dY (block 3.7) and the block for storing the average (block 3.10). The output of block 3.5 YB is connected to the input of block 3.6 and the block for calculating corrections dY (block 3.7). The output of block 3.6 Y 'is connected to the block for calculating the values of DC1 (block 3.8) and the block for calculating the values of DC3 (block 3.9). In addition, the output of block 3.1 C1 'is connected to the input of block 3.8, and the output of block 3.1 C3' is connected to the input of block 3.9. The output of block 3.7 is connected to the medium conservation block (block 3.10). The outputs of block 3.7 dY, block 3.8 DC1, block 3.9 DC3, block 3.10 YA 'are connected to the corresponding inputs of the block forming color channels (block 3.12). The outputs of block 3.12 C1T, C3T, C2AT and C2BT are the outputs of the color channel conversion unit. The control and address generation unit (block 3.11) has an input for receiving configuration information and is connected to all of the above blocks of the color channel conversion unit.

The compression device (figure 2) works as follows. Block 2.1 is a source of receiving color channels from a photosensor, and also carries out the necessary pre-processing. As a result, the input channels of block 2.2 receive the color channels C1, C2, C3, which in the case of BMI correspond to the channels R, G, B. Block 2.2 converts the specified color channels into C1T, C2AT, C2BT, C3T, which are fed to the input of block 2.3. Block 2.3 performs direct discrete wavelet transform and quantization for each of the color channels, statistical coding, for example, arithmetic coding or Huffman coding, forms a compressed stream, which is transmitted to block 2.4, which provides storage and transmission of compressed data.

Consider the operation of the color channel conversion unit (figure 3). In block 3.11, the input IS receives configuration information about the dimension of the processed data, the range of their variation, etc. Then block 3.11 generates a synchronizing pulse, which is transmitted to the input of the SS block 3.1. Block 3.1 generates extended matrices C1 ', C2A', C2B ', C3' in accordance with rule 1.1 of the above direct conversion of color channels. Then block 3.11 generates a synchronization signal, which is fed to the clock inputs of SS blocks 3.2 and 3.3. As a result, block 3.2 calculates С1'i values in accordance with rule 1.2, and block 3.3 calculates С3'i values in accordance with rule 1.2. After that, block 3.11 generates a clock pulse, which is fed to the sync inputs by block 3.4 and 3.5. Block 3.4 calculates YA values in accordance with rule 1.3. Block 3.5 calculates the values of YB in accordance with rule 1.3. After calculating the values of YA and YB, block 3.11 generates a clock pulse, which is fed to the clock inputs SS. Block 3.6 calculates the values of Y 'according to rule 1.6. Block 3.7 calculates the dY correction values according to rule 1.4. Then block 3.11 generates a clock signal, which is fed to the corresponding inputs of blocks 3.8, 3.9 and 3.10. As a result, block 3.8 calculates DC1 values according to rule 1.7, block 3.9 calculates DC3 values according to the same rule, and block 3.10 calculates the average in accordance with rule 1.5 based on the dY and YA values supplied to its inputs. At the output of block 3.10, values YA 'are formed. Block 3.11 generates a clock pulse, which is fed to the input of block 3.12. Block 3.12, based on the values of DC1, DC3, dY, YA ', which are supplied to its respective inputs, generates color channels C1T, C2AT, C2BT, C3T in accordance with rule 1.8.

The present description of the invention, including the composition and operation of the device, including the proposed version of its execution, involves further possible improvement by specialists and does not contain any restrictions in terms of implementation.

Claims (5)

1. A method of compressing digital images by sequentially applying the operations of dividing the raw digital data received from the image capture device into a plurality of color channels, including color channels, each of which is subjected to two-dimensional discrete wavelet transform, quantization, and statistic encoding, characterized in that the separation operations color data is carried out by obtaining four full-sized expanded matrices from three color channels (R, G, B), the first of which (C ') contains um the values of the first channel (R), the second (C2A ') and the third (C2B') contain the values of the second channel (G), the fourth (C3 ') values from the third channel (B), using interpolation of color matrices containing the values of the first and the third color channel, receive data at points where the values of the second color channel are known, convert the values in the second and third matrix (C2A ', C2B') using the interpolated values of the first and fourth matrices (C1i ', C3i') into new values (YA and YB) by the formula YA = a1C1'i + a2C2A '+ a3C3'i YB = a1 · C1'i + a2 · C2A '+ a3 · C3'i' where a1, a2, a3 are real coefficients, by interpolating color matrices containing the values of the second color channel, data is obtained at points where the values of the first and third color channels are known, the values of the first and fourth matrices (C1 ', C3') are converted using the interpolated values of the second channel (Y) into new values (DC1, DC2) according to the formula DC1 = b1 · (C1'-Y ') DC3 = b2 · (C3'-Y ') where b1, b2 are real coefficients, according to the values of the second matrix, the third matrix is obtained by prediction, writing only corrections from the predicted values (dY) into it, and from the converted full-sized matrices they go back to color channels, three of which are differential. 2. The digital image compression method according to claim 1, characterized in that when using the quantization coefficients developed for the YCbCr color rendering system, the color channels are the Bayer tiles: red, green and blue, respectively. 3. A method of compressing digital images according to claims 1 and 2, characterized in that the material coefficients a1, a2, a3 take the values: a1 = 0.299; A2 = 0.587; a3 = 0.114. 4. A method of compressing digital images according to claims 1 and 2, characterized in that the material coefficients b1, b2 take the values: b1 = 0.7133, b2 = 0.5643. 5. A digital image compression device, consisting of a digital image capture and preprocessing unit connected to a color channel conversion unit, coupled to a compressed data storage and transmission unit, characterized in that the color channel conversion unit consists of an expanded matrix generation unit that is connected to blocks for calculating the values of DC1 and DC3, blocks for calculating the values of YA and YB, blocks for calculating the interpolated values of C1'i and C3'i; the interpolated value calculation unit C1'i and the interpolated value calculation unit C3'i are connected to the YA and YB value calculation units, respectively, the YA value calculation unit is connected to the Y ′ value calculation unit, the correction calculation unit dY and the average YA ′ storage unit, the calculation unit the values of YB is connected to the block for calculating the values of Y 'and the block for calculating the corrections dY, which is connected to the block for storing the average YA', the block for calculating the values of Y 'is connected to two blocks for calculating the values DC1 and DC3, the blocks for calculating DC 1 and DC3, as well as the correction calculation unit dY and the average YA ′ storage unit, are connected to the color channel generation unit, and the control and address generation unit is connected to all of the above blocks.

Images