JPH0198379A - Orthogonal transform coding device - Google Patents

Abstract

PURPOSE:To reduce the quantity of arithmetic operation in the orthogonal transformation while keeping the picture quality by processing a picture as a block and decreasing the picture element number of each block depending on the characteristic quantity. CONSTITUTION:An original picture data 1 is split into blocks by a basic block read section 2 and a picture data in the basic block is subjected to a differential data with a predicted value at every picture element at a differential picture generating section 3. The basic block is classified into degenerated block depending on the distribution of the differential data at a degeneration block classification section 4 and subjected to 2-dimensional discrete sign transformation by orthogonal transform sections 51-53. The transform coefficient obtained from the orthogonal transform sections 51-53 is quantized by a quantization section and outputted through a data selection section 7.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to an apparatus for converting and encoding image data.

Prior Art □ Regarding the encoding of image data, orthogonal transform encoding is known to be a highly efficient encoding method that can largely remove redundancy in image data.

However, the amount of calculation required for encoding is enormous, and the processing time has been a problem.

Up to now, various high-speed algorithms have been devised to reduce the amount of calculations required to perform orthogonal transform. For example, for discrete cosine transform, there is a high-speed algorithm proposed by Chen et al. Z? p? 97'L-d l
Kot, zzlr/vf/su1 Transsact+o
ns on Commun+cations, vol.
, cOM-25, No. 9. September 1
977).

However, even if the above-mentioned high-speed algorithm is used, the amount of calculation required is still large, and the time required for encoding is by no means fast from a practical point of view.

In addition, there is, for example, Japanese Unexamined Patent Application Publication No. 61-294585, which aims to further increase speed using a high-speed algorithm. In order to concentrate, these are approximated by "0", and calculations for predetermined coefficients are omitted. However, although it is possible to reduce the amount of calculation, there is a problem in that the quantization bit allocation of the transform coefficients is treated as fixed for any image, resulting in poor image quality.

Problems to be Solved by the Invention As mentioned above, with regard to speeding up processing, most devices in the past have only applied high-speed algorithms, and with the demand for faster processing, further speeding up has not been possible. Not planned.

In addition, in the above conventional example, which aims to reduce the amount of calculation by focusing on transform coefficients, the bit allocation for coefficient quantization is fixed. There was a problem in that no measures were taken to maintain image quality or speed up processing by changing the bit allocation with blocks that are not compatible.

By the way, in orthogonal transform encoding, image data is usually divided into a plurality of blocks, and orthogonal transform encoding processing is performed for each block. Size of this divided block (number of pixels)
is usually determined by the balance between the quality of the decoded image, the time (amount of calculation) required for encoding/decoding processing, and the scale of the hardware. Generally, the larger the block size, the better the quality of the decoded image, and it is considered desirable to process with large blocks in order to maintain the image quality. On the other hand, it is known that the conversion processing time (amount of calculation) can be reduced by dividing into smaller block sizes.

By the way, regarding the relationship between the transformation block size and the amount of calculation (processing time), in the case of discrete cosine transformation, the time required to transform the entire image with 8x8 block division is equal to the processing time with 16x16 size block division. Compared to this, there is also a report on about 178 (VSP-161U)
SPIR8MANUAL ZORAN). Furthermore, when the block size is small, the hardware scale is naturally smaller.

In view of this point, the present invention addresses the demand for faster processing.
It is an object of the present invention to provide an encoding device that improves the above-mentioned problems from the above-mentioned viewpoints.

Means for Solving the Problems The present invention provides means for dividing an input image into blocks of n1×n2 pixels (hereinafter referred to as basic blocks), and converting image data in the basic blocks into difference data with predicted values. and convert the basic block into differential data into m1×m2 (ml≦n1.m2) including the basic block size.
≦n2) means for reducing the size (hereinafter referred to as a reduced block), and a plurality of sets of the basic block reduction means are provided, and the plurality of sets of reduced block size are The present invention includes means for selecting one of the reduced blocks, and means for managing selection information of the reduced blocks, and performs orthogonal transform encoding on data of the selected reduced blocks.

Effects The present invention improves image quality by selecting a degenerate block based on the feature amount of differential image data for each divided basic block and adaptively changing (reducing) the number of data to be encoded. Encoding processing can be performed at high speed without significant loss.

EXAMPLES Hereinafter, examples of the present invention will be described with reference to the drawings.

FIG. 1 shows a schematic block diagram of an orthogonal transform encoding apparatus in an embodiment of the present invention. A discrete sine transform is used as the orthogonal transform.

In FIG. 1, 1 is original image data, which is usually stored in a memory, etc., 2 is a basic block successive unit that divides the image data into blocks, and a is the basic block data read out by 2. 3 is a difference image creation unit that converts the image data a in the basic block into difference data with a predicted value of 2 for each pixel, and b is the basic block data replaced with the transformed difference image data. 4 is a degenerate block classification unit that classifies basic blocks into degenerate blocks according to the distribution of differential image data b within the basic block; 51 to 53 are orthogonal transform units that perform two-dimensional discrete sine transformation; and 61 to 63 are each of the above-mentioned orthogonal blocks. A quantization unit quantizes the transform coefficients obtained from the transform units 51 to 53, 7 is a data selection unit, and 8 is a block mode information management unit.

Original image data 1 has a block size of 16x16, for example.
The basic block sequential output unit 2 sequentially reads out the pixels as pixels. Next, the basic block data a is converted into difference image data b by the difference image creation section 3. FIG. 2(a) is a diagram illustrating a first embodiment of the difference image creation section 3. In the figure, a representative value e representing all data in the basic block is obtained as a predicted value, and each pixel data (A.

B,...) with this predicted value e (A-e,
B-e, ...) in sequence.

As a result, the basic block data a is converted into basic block data consisting of the representative value e and the difference image data b. Possible methods for calculating the representative value include selecting the average value of all pixels, an intermediate value, or the most frequently occurring pixel value. FIG. 2(b) is a diagram illustrating a second embodiment of the difference image creation section 3. In the figure, all data in a basic block is sequentially replaced with difference data (BA, CB, . . . ) between adjacent pixels using adjacent pixel data as predicted values. However, for the leftmost pixel in each row, difference data is obtained using the pixel data adjacent above as a predicted value. Furthermore, the pixel in the first row and first column has its value “A” as the representative value e(
e=A), and O is applied as the difference pixel value of the first row and first column. Therefore, also in the case of FIG. 2(b), the basic block data a is converted into basic block data consisting of the representative value e and the difference image data b.

Next, with respect to the basic block data replaced by the difference image data b, the degenerate block classification unit 4 classifies the basic block into a degenerate block based on the distribution of data within the block. As an evaluation criterion for the distribution of data within a block, for example, the variance of each pixel value (differential image data) within a block is determined and used. In the example shown in the figure, the corresponding block is classified into the following three modes based on the obtained pixel value variance.

(1) Mode 1: Pixel value variance Large → 16x16 pixels (2) Mode 2: Medium → 8x8 pixels (3) Mode 3: Small → 4x4 pixels Figure 3 shows the differences between each basic block in one entire image. This shows the distribution state of the variance of pixel values, and by setting appropriate thresholds Thl and Th2, each basic block is classified into three modes as shown in the figure. FIG. 4 shows this mode classification.

In this case, mode 1 is applied to blocks with large pixel value variance, and mode 2 and mode 3 are applied as the variance becomes smaller. In other words, for blocks with sharp changes in brightness level, or edge parts in images, the original basic block data is selected as degenerate data and all pixels in the block are targeted for encoding, and for parts with smooth brightness changes. is the basic block 16X
Processing is performed by converting 16 pixels into 8×8 or 4×4 degenerate data.

As a result, blocks with sharp luminance changes are encoded as usual using mode 1, and blocks with gradual luminance changes are encoded with a reduced number of pixels using modes 2 and 3. As the number of times Mode 2 and Mode 3 are used for the entire image increases, processing with a small block size increases, and the encoding processing time can be shortened. Furthermore, since the block data in modes 2 and 3 correspond to portions with smooth luminance changes, there is little deterioration in image quality due to conversion to degenerate data.

The mode-classified block data is input to discrete sign conversion units 51 to 53 corresponding to each number of encoded data (16×16.8×8.4×4 pixels), and conversion is executed. The coefficients transformed by the discrete sine transformation units 51 to 53 are quantized by quantization units 61 to 63, respectively. In the data selection section 7, the classification information C from the degenerate block classification section 4 is
The encoded data d of the corresponding reduced block is selected according to the following.

At the same time, the block mode information management section 8
Data f is created to manage which mode is applied to which block (basic block).

As a result of the above operation, basic block data a is converted into a combination of block representative value e, encoded data d1, and block mode management information f. Here, the representative value e and the block mode management information f are redundant compared to the case where only normal block division without mode classification is performed and orthogonal transformation is performed, but for e, the basic block data is transformed into difference image data with the predicted value. Normally, this does not pose a problem because the conversion efficiency (compression rate) of the data within the block can be expected to improve by conversion. Regarding the mode management information for each block, even if it were divided into three modes, each block would only have 2-bit management data for the number of basic blocks at most, and would not have much effect on the overall code amount.
In addition, instead of having two bits for each mode as described above, the mode classification management data is created by giving each mode an optimal code word, such as a Huffman code word, based on the number of times each mode is used. It is also possible to eliminate redundancy in management data.

In the above embodiment, a method may be considered in which the pixels of the degenerate block (a-i-d) of each mode are represented by, for example, their average value. In addition, when replacing a plurality of pixels with a representative value, instead of taking the average value, the intermediate value of the pixel values within each basic block may be used as the representative value.

As for the mode classification method, in this embodiment, the variance of pixel values of each pixel in a block is used, but other methods such as correlation of pixels in a block or selection of a mode with the smallest approximation error are also possible.

At the time of decoding, the representative value e of the block and the encoded data d,
The original image is reproduced for each basic block based on the block mode management information f. For blocks to which mode l is applied, basic block data b can be restored by performing inverse orthogonal transformation as conventionally. For blocks that were encoded by applying mode 2 or mode 3, first perform inverse orthogonal transformation,
8x8 or 4x4 degenerate block data are obtained, respectively, and then the original basic block data b is restored by the reverse procedure of the degenerate block creation method shown in FIGS. 5 and 6. At this time, data lost due to the degeneracy operation is replaced with adjacent pixel data,
Alternatively, a method of interpolating from adjacent pixel data groups may be considered. Since the blocks to which Mode 2 and Mode 3 are applied correspond to smooth parts of the image, the deterioration in image quality due to the above decoding operation is minimal.

In addition to the above mode classification, in this embodiment, a preset threshold value Th1. Although processing is performed with Th2 as a fixed value, the mode classification criteria Th1. Th2 may be changed.

For example, by examining the distribution of data in each basic block of the entire image in advance, the appearance frequency of each mode can be set to T so that the amount of data after encoding is within a specific range.
h1. Control by optimizing Th2. in this way,
If mode classification is performed in consideration of the number of uses of each mode, the amount of code can be controlled to a certain extent regardless of the image. At this time, in reality, (a), (
Since the dispersion characteristics of the entire difference image data are considered to be different from (b), the thresholds Thl and Th2 will be set differently for each image. This classification method is effective when there is an upper limit to the number of data that can be transmitted, such as in transmission over a line.

Regarding the hardware scale of the device, the embodiment shown in FIG. 1 has a discrete sign conversion unit for each mode, but the DSP (1) Digital Signal Process
If a dedicated processor such as r) is used, it is possible to perform FDST calculations of 4×4.8×8.16×16 pixels with one processor, and there is no need to actually have multiple conversion units and the hardware scale does not increase.

Further, in this embodiment, a two-dimensional discrete sine transform is used as the orthogonal transform, but this may be a one-dimensional discrete sine transform, or other Hadamard transform or Fourier transform. However, as in this embodiment, the discrete sine transformation is superior as an orthogonal transformation function applied to difference data from some predicted value. Moreover, it is known that discrete sine transform has superior characteristics compared to discrete cosine transform even with small block sizes. (In the case of one-dimensional encoding) □”, Yamane et al., Theory of IEICE (D
), J69-B, ? , PP, 686-69
7 (Sho 6l-7)] It can be said that discrete sine transformation is a particularly promising method for implementing the present invention.

According to this embodiment, by approximating a smooth block with a small number of pixels by dispersing the differential image data of each block, and performing orthogonal transformation using encoded data, it is possible to efficiently reduce the amount of calculation for orthogonal transformation. becomes. In terms of image quality, since the number of encoded data is adaptively changed for each block,
Portions where the luminance changes sharply can be reproduced with the same image quality as before, while portions where the brightness changes slowly will be hardly noticeable even if there is some deterioration due to the reduction in the number of pixels.

As described in detail, the present invention provides the following effects.

(1) By dividing the image into blocks and further reducing the number of pixels in each block based on its feature value, it is possible to reduce the amount of calculation in orthogonal transformation while maintaining image quality, making the encoding process faster than before. can be executed.

(2) Although the representative value for each basic block and the control data for mode classification become redundant, the introduction of degeneracy 10 can be expected to reduce the overall amount of code, making it possible to improve the compression rate.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of an orthogonal transform encoding device according to an embodiment of the present invention, Fig. 2 is a diagram illustrating an embodiment of the difference image creation section of the same device, and Fig. 3 is a basic block diagram for mode classification. Figure 4 is a diagram showing the distribution of the variance values of the differential image data for the entire image, Figure 4 is a diagram showing the group of degenerate blocks when classifying basic blocks into degenerate blocks, and Figure 5 is the block for creating degenerate blocks. Figure 6 shows the case where the pixels within the block are roughly sampled, Figure 6 shows the case where degenerate block creation is performed by representing the average value of multiple pixels within the block, and Figure 7 shows how the basic FIG. 7 is a diagram showing that the distribution of variance values of differential image data for each block is different over the entire image. 1-Original image, 2. Basic block successive section, 3.-Differential image creation section, 4.. Degenerate block classification section, 51-53
・Orthogonal transformation unit, 61 to 63 quantization unit, 7′ data selection unit, 8 ・°゛Block mode information management unit. Name of agent: Patent attorney Toshio Nakao (1 person) Figure 2 Figure 3 Figure 5 +a) 16X11m! Basic a-)
(c) Mode l (16X16) (C1 mode 2 (8
X8) Remode 3 (4X4
)・:f sampling pixel×:pixel not used for encoding
(bl %-Fl (16X16) (C1 mode 2
(8X8) (b) Mode 3
(4X4)O: average value

Claims (5)

[Claims] (1) A device that converts and encodes image data,
means for dividing an input image into blocks of n1×n2 pixels (hereinafter referred to as basic blocks); means for converting image data in the basic blocks into difference data with a predicted value; and means for converting the basic blocks converted into difference data. means for degenerating the basic block to a size m1×m2 (m1≦n1, m2≦n2) including the basic block size (hereinafter referred to as a degenerate block), and a plurality of sets of the basic block degenerating means are provided, comprising means for selecting one of the plurality of sets of degenerate blocks based on feature amounts of the differential image data, and means for managing selection information of the degenerate blocks,
An orthogonal transform encoding device that performs orthogonal transform encoding on data of the selected reduced block. (2) The orthogonal transform encoding device according to claim 1, wherein the means for degenerating the basic block to m1×m2 size is performed by coarsely sampling the basic block. (3) The means for degenerating the basic block to m1×m2 size approximates a plurality of pixels among the differential pixels in the basic block to one. Orthogonal transform encoding device. (4) A patent characterized in that the means for selecting one of the plurality of sets of degenerate blocks based on the feature amount of the differential image data within the basic block is performed by determining the variance of the differential image data within the basic block. An orthogonal transform encoding device according to claim 1. (5) means for selecting one from the plurality of sets of degenerate blocks based on the feature amount of the differential image data in the basic block, the means for selecting one from the plurality of sets of degenerate blocks;
The orthogonal transform encoding device according to claim 1, wherein the orthogonal transform encoding device controls the amount of data after encoding so that it falls within a certain range by adaptively operating each image to be encoded. .