JP2811304B2 - Method of compressing gradation image data - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a method of compressing gradation image data for compressing gradation image data such as medical images such as X-ray photographs and CT images. BACKGROUND OF THE INVENTION With the advancement of digital technology, digitized gradation images are frequently stored, transmitted, and subjected to various digital image processings. However, a tone image has a larger amount of information than a binary image, and thus has a problem of a large amount of data when the tone image is digitized. In particular, in the case of medical images, the number of pixels and the number of bits required for each pixel when digitizing are enormous, for example, 4 million pixels and 8 to 10 bits in a chest radiograph, which is an efficient method for storing and transferring data. Is bad. Therefore, today, a data compression technique for compressing enormous data of a gradation image including a medical image to make it compact has been spotlighted. Data compression technologies can be broadly classified into lossless compression and irreversible compression.However, lossless compression can only achieve a low compression ratio of about 1/2 to 1/3, so a high compression rate of 1/5 or more can be obtained. Irreversible compression schemes, especially transform coding schemes, are attracting attention. Transform coding is an irreversible compression method that divides the entire image into small blocks, performs orthogonal transform on a block basis, quantizes the resulting transform coefficients, and encodes them. This is the most suitable compression method for compression. In transform coding, the distribution of the AC component of the transform coefficient is
It is known that it is approximated to a Gaussian distribution having a peak at zero, and the quantization of the transform coefficient having such a distribution is performed by Mid-Riser type quantization including zero in the quantization determination level, and the quantization output level. It has been reported that when classified into Mid-trace type quantization that includes zero, the Mid-trace type quantization has less random noise in the block and is more preferable. However, when a Mid-trace type quantizer is used, there is a drawback that when the compression ratio is increased, a component-specific pattern image appears, which is unnatural when performing image processing such as enlargement, gradation, and frequency processing. There is a problem that the image appears and the fidelity of the restored image is reduced. By the way, it is understood that the image quality of the restored image in the case of performing the Mid-trace type quantization has a strong correlation with the number of the transform coefficient AC components quantized to zero in the quantization stage at the time of compression. That is, since the image of each block of the restored image is a weighted sum of a pattern unique to the component of the transform coefficient AC component quantized to a value other than zero at the time of quantization, the transform coefficient AC component quantized to a value other than zero is calculated. Is small, in other words, when the number quantized to zero is large, a component-specific pattern image appears in the restored image. Therefore, in order to suppress the appearance of the component-specific pattern image, the number of transform coefficient AC components quantized to zero is reduced by narrowing the quantization width of the quantization determination level at which the quantized output level is zero. In other words, the number to be quantized may be increased to a value other than zero, but in such a case, the average code length at the time of encoding increases, so that the compression ratio decreases. Therefore, the frequency of occurrence of the transform coefficient AC component quantized to zero is checked for each block. Since the total number of conversion coefficient AC components in the block is fixed, it is the same as checking the number other than zero. As a result, it has been found that it is possible to obtain a restored image in which the appearance of a pattern image is suppressed by detecting blocks in which a component-specific pattern image is likely to appear and narrowing the quantization width of only the detected blocks. (Object and Configuration of the Invention) The present invention has been made in view of the above points, and
-When compressing the data of a gradation image including a medical image using a transform coding method that performs trace-type quantization, the generation of a component-specific pattern image is prevented, and a high-quality restored image with a high compression ratio is obtained. In order to achieve this object, digitized grayscale image data is divided into a plurality of blocks, and a transform coefficient obtained by performing orthogonal transform for each block is quantized and encoded. In the method of compressing tonal image data, a range of values of the transform coefficient is divided into a plurality of predetermined sections, and a frequency of occurrence of the transform coefficient in at least one of the divided sections is determined.
The quantization width of the transform coefficient is determined for each block based on the determined occurrence frequency, and compression is performed. (Example) Hereinafter, the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a gradation image data compression method according to the present invention. In the figure, a frame memory 1 stores gradation image data to be compressed (in this example, the number of bits per pixel is 8 bits). Are read in block units (in this example, the block size is 16 pixels in the line direction and the column direction, respectively). The read block data is converted into a two-dimensional discrete cosine transform (2D-DC
T) The cosine transform is performed by the device 3 to obtain 256 transform coefficients for one block. Next, of the 256 transform coefficients thus obtained, 255 transform coefficient AC components excluding one DC component are uniformly quantized by the quantizer 4 with a basic quantization width w _o , and each transform is performed. A pair of a coefficient number and a polarity number, which are fixed-length codes, is obtained for each coefficient. The pair of the coefficient number and the polarity number is temporarily stored in the block buffer memory 6, and the coefficient number is sent to the classification device 5. The classifying device 5 performs a classifying process based on the coefficient number of one block sent from the quantization device 4 and the compression ratio parameter input from the terminal 52, and outputs the class number to the encoding device 7. When reading the class number sent from the classifying device 5, the coding device 7 reads a pair of the coefficient number and the polarity number for one block from the block buffer memory 6, and converts the fixed-length code (number) of the transform coefficient AC component. The data is converted into a variable length code and output to the terminal 8. The code data is transmitted from the terminal 8 or stored in the memory. FIG. 2 shows an example of the quantization device 4. First, 255 transform coefficient AC components excluding the DC component among the transform coefficients for one block are input from the terminal 41. The absolute values of the 255 conversion coefficients are determined by the absolute value circuit 42, and at the same time, the polarity determination circuit 43 determines whether the conversion coefficients are positive or negative. The polarity determination circuit 43 outputs the determination result to the terminal 47 as a polarity number. Here, the polarity number is represented by a 1-bit code, and is “1” when the conversion coefficient is negative and “0” when the conversion coefficient is positive. On the other hand, the absolute value of the transform coefficient output from the absolute value circuit 42 is divided by the basic quantization width w _o by the division circuit 44. The division result is truncated to the decimal point by a truncation circuit 45 (this result is referred to as a coefficient number) and output to terminals 46 and 47. The terminal 46 is connected to the classification device 5, and the terminal 47 is connected to the block buffer memory 6. The polarity number and coefficient number obtained for one transform coefficient input from the terminal 41 form a pair of two, and are temporarily stored in the block buffer memory 6 from the terminal 47. On the other hand, the DC component of each block has a sufficiently small quantization width w.
It is uniformly quantized by _dc , or is temporarily stored in the block buffer memory 6 in the same manner as the other 255 transform coefficient AC components without performing quantization. FIG. 3 shows the state of quantization of the AC component of the transform coefficient excluding the DC component. The horizontal axis is the value of the conversion coefficient, and the vertical axis is the frequency of occurrence of the conversion coefficient. The number sequence shown in FIG. 3 (a) is the sequence of coefficient numbers (k) output from the truncation circuit 45.
The number sequence shown by (b) is the polarity number (j) sequence output from the polarity determination circuit 43. Next, FIG. 4 shows an example of the classification device 5 when the number of classes is four. The terminal 51 is connected to the terminal 46 in FIG. 2, from which the coefficient number output from the quantization device 4 is input. When the coefficient number is represented by k (k = 0, 1, 2, 3, 4,...), The coefficient number k input from the terminal 51 is converted into a 4-bit primary counter control code C _i ( i = 0,1,
Is converted to 2,3). FIG. 5 (a) shows an example of the counter control LUT 53. On the other hand, a compression ratio parameter P input by the user from a terminal (not shown) is input from the terminal 52. In the case of this example, two types of compression ratios are prepared (P = 0, 1), and the larger the P, the higher the compression ratio. Compression rate parameter P input from the terminal 52, the compression ratio control LUT 54, is converted into 4-bit compression rate control code e _p. FIG. 5 (b) shows an example of the compression control LUT 54. Next OR circuit 55 takes in the compression ratio control code e _p a primary counter control code c _i, obtaining the second-order counter control code d _i of 4 bits is a logical sum of both. d _i ← c _i ∪ e _p Next, the counter control circuit 56 executes the second counter control code.
d _i is fetched, and four counters from counter 0 to counter 3 in the counter circuit 57 are controlled by the on / off state of each bit of d _i . Counter 0, the counter 1 from each bit lower bits of the second-order counter control code d _i in this order,
When the bit is on (= 1), the counter control circuit 56 counts up the corresponding counter by +1 and the bit is off (= 0).
In this case, the corresponding counter is not incremented. In FIG. 5 (c) shows the correspondence between the bits and the counters d _i. Incidentally, c _i, is a similar relationship with each counter holds for e _p. 6 is a diagram showing a relationship between operating range of the compression rate parameter P or compression ratio control code e _p and each counter.
For each compression rate parameter P (each e _p), each counter operates only for the coefficients number k of the range indicated by arrows in FIG. 6. However, in this example, the counter 3 operates for all coefficient numbers k regardless of the value of P.
The end of the classification work for one block can be known from the counter 3. Therefore, if another counter similar to the counter 3 is possessed, the counter 3 may be omitted. Next, the comparison processing procedure of the comparison circuit 58 is shown in the flowchart of FIG. The comparison circuit 58 always checks the count value of the counter 3 (F-1). When the count value is less than 255, the comparison circuit 58 is in a standby state. When the count value of the counter 3 is 255, the comparison circuit 58 reads the count threshold S _p which is output from the count threshold LUT 59. FIG. 5 (d) shows an example of the count threshold LUT59. A count threshold S _p is emitting the compression rate parameter P which is input from the terminal 52, a threshold value for the count value. Comparison circuit
When the 58 reads the count threshold S _p, the S _p and counter 2
(F-2). as a result,
(1) If the count threshold S _p is set to class number m equal to or greater than the count value of the counter 2 to 3 (F-3),
S _p is smaller than the count value of the counter 2 is compared with the count value of the S _p and counter 1 (F-4). (2) In comparison between the count value of the count threshold S _p and counter 1, if S _p is more than the count value of the counter 1, and set the class number m to 2 (F-5), S _p counter 1 smaller than the count value comparing the count value of the S _p and the counter 0 (F-6). (3) In comparison between the count value of the count threshold S _p and the counter 0, if S _p is more than the count value, and sets a class number m to 1 (F-7), the count value of S _p is counter 0 If smaller, the class number m is set to 0 (F-8). (4) When the class number m is set, the class number m is output from the terminal 510 (F-9), the classification of one block is completed, and the counters 0 to 3 are counted. Clear the value to zero (F-1
0). The class number m output from the comparison circuit 58 is a parameter indicating the range of the transform coefficient to be run-length encoded at the time of encoding. That is, the class number is relative to 255 of the AC component of the block is m, the transform coefficient whose absolute value is present within the following ranges 0 or A _m of transform coefficients means that the run-length encoding ( (See FIG. 6). FIG. 8 is an example of the encoding device 7. When the classifying device 5 completes the classifying operation for one block, the encoding device 7 reads the class number m from the terminal 79 (the terminal 79 is connected to the terminal 510 in FIG. 4) and stores the class number m in the quantization LUT 73. One of the tables used for quantization is selected from the four stored quantization tables, and at the same time, a data read command from the block buffer memory 6 is sent to the data read circuit 72. The data reading circuit 72 receives the DC component data from the terminal 71 connected to the block buffer memory 6 and the 255 sets of coefficient numbers k and polarity numbers j.
(J = 0 or 1) is read. If the read data is a DC component, the data is output from the terminal 78. If the read data is a pair of the coefficient number k and the polarity number j, the quantization in the quantization LUT 73 selected by the class number m is performed. The coefficient number k and the polarity number j are converted into the code number 1 by the table. FIG. 9 shows an example of the quantization table in the quantization LUT 73. This example is an example in which the coefficient number k is uniformly divided so that the final quantization state is the uniform quantization shown in FIG. 10, but the coefficient number k may be unequally divided. It is possible, and non-uniform quantization can be performed with a final quantization width that is an integral multiple of the basic quantization width _Wo . Next, an example of the coding procedure of the AC component is shown in the flowchart of FIG. First, the code number 1 output from the quantization LUT 73 is sent to a zero determination circuit 74 having a code passage counter 74a therein (P-1). When the code passage count 74a receives the code number l, the code passage count value y is incremented by one (P-2), and the zero determination circuit 74 performs the zero determination of the code number l (P-3). If 1 is zero, after incrementing the zero count value z of the zero counter 75 by +1 (P-4), it is determined whether or not the code passage count value y has reached 255 (P-11). On the other hand, if 1 is not zero, it is determined whether the zero count value z is zero (P−
5) If zero, the code number 1 is converted into a variable length code hl by the code LUT 76 (P-6) and hl is output from the terminal 78 (P-7) Is the code passage count value y equal to 255? It is determined whether it is (P-11). If the zero count value z is not zero the zero count value z is converted by the code LUT77 the variable length code r _z (P-8), was cleared to zero zero count value z (P-9), terminals r _z It is output from 78 (P-10), and it is determined whether or not the code passage count value y has reached 255 (P-11). In the determination of the code passage count value y, if the count value y is not 255, the process returns to the step (P-1) of reading the code number 1 and repeats the above processing. If the code passing count value y is 255 (P-11), it is determined whether the zero count value z is zero (P-1).
2) If it is zero, the code passage count value y is cleared to zero (P-13), and the encoding operation of the transform coefficient for one block is completed. If zero count value z is not zero and code conversion step (P-8), and zero clearing step of zero count value z (P-9), a variable length code r _z output step to the terminal 78 (P-10) After this, the code passing count value y is cleared to zero (P-13), and the encoding operation of the transform coefficient for one block is completed. FIG. 12 (a) shows an example of the code LUT76. This example is an example using a common Huffman code hl for each class. In the case of performing uniform quantization as in the present embodiment, there is no significant difference in the compression ratio between when the optimal Huffman code hl is used for each class and when the common Huffman code hl is used for each class. Huffman code common to each class
By using hl, a simpler device can be realized. FIG. 12 (b) shows an example of the code LUT77. This example is an example in which a B1 code common to each class is used as a run length code assigned to the zero count value z. When the gradation image data is compressed by the present method, the frequency distribution of each zero run length becomes an exponential function suitable distribution suitable for the B1 code. Therefore, it is possible to obtain a high compression ratio by using the B1 code. In FIGS. 12 (A) and 12 (B), C and 1 take a value of 0 or 1, and are 1-bit codes satisfying ΔC. FIG. 13 shows the data reading order of the data reading circuit 72 in FIG. Since the conversion coefficient of the gradation image tends to have a smaller amplitude as the frequency component increases, the probability of generation of a component quantized to zero also increases as the frequency component increases. Therefore, Fig. 13 (a)
Alternatively, as shown in (b), if encoding is performed from the low frequency side toward the high frequency side, the probability of occurrence of a long zero run increases, and effective run-length encoding can be performed. FIG. 14 shows an example of a code configuration. FIG. 14 (a) shows a code example of the result of compressing the entire image of one frame, which is composed of a header part, a code of each subsequent block, and a data end code. FIG. 14 (b) shows an example of the header section configuration, which includes the block size (16 in this example), the number of blocks in the image line direction,
Number of blocks in image column direction and compression ratio parameter P
Is stored. FIG. 14 (c) shows an example of the code configuration of each block.
It is composed of a class number m (in this example, a 2-bit fixed-length code), a DC component code (fixed-length code), and a subsequent AC component code sequence (variable-length code sequence). The AC component code sequence is a coefficient code sequence (variable length code sequences hl ₁ , hl ₂ ,... Hl _q ) composed of q (1 ≦ q ≦ 255) codes as shown in FIG. ) And run length codes (variable length codes r _z ) are alternately arranged. Hereinafter, examples of various parameters used in this embodiment will be described. Basic quantization width w ₀ = 0.5 A ₀ = 0.5, A ₁ = 1.0 A ₂ = 1.5, A ₃ = 2.0 S ₀ = 200, S ₁ = 225 Note that the information amount per pixel of the original image is n.
The block size is N × N with bits (n = 8 in this example)
When (N = 16 in this example), the block image is represented by f (x,
y), a value obtained by calculating a two-dimensional discrete cosine transform equation (shown in the following equation) representing a transform coefficient matrix by F (u, v) at a floating point. In the present invention, instead of the two-dimensional discrete cosine transform device 3, another orthogonal transform device, for example, 2
A dimensional discrete Hadamard transform device, slant transform device, or the like may be used. Also, in the present invention, the block size is not limited to the size of 16 × 16, but is 8 ×.
Sizes such as 8, 32 x 32, 64 x 64 are also applicable,
Further, the number of bits per pixel of the original image is not limited to 8 bits. When counting the coefficient numbers k in a specific section, the classifying apparatus 5 may classify based on the number of coefficient numbers k present in one specific section without providing a large number of specific sections. It is possible. Further, the classification can be performed by combining the classification parameter in the present invention with other classification parameters. (Effects of the Invention) As described above, in the present invention, the range of values that the transform coefficient can take is divided into a plurality of predetermined sections, and the frequency of occurrence of the transform coefficient in at least one of the divided sections is determined. By using, the quality of the restored image, especially
Quantization based on the degree of occurrence of a component-specific pattern image, which is a drawback of the mid-trace type quantization, becomes possible. For example, by determining a quantization width so that a high-frequency block has a low frequency, a pattern image of the pattern image can be obtained. Generation can be suppressed and a high-quality restored image can be obtained.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram of an embodiment of a gradation image data compression method according to the present invention, FIG. 2 is a block diagram of an example of a quantization device shown in FIG. 3 is a diagram showing a coefficient number and a polarity number obtained by quantizing a transform coefficient AC component in relation to the frequency of occurrence of a transform coefficient, and FIG. 4 is a block line of an example of the classification device shown in FIG. FIG. 5A shows a counter control LUT, FIG. 5B shows a compression ratio control LUT, FIG. 5C shows a correspondence relationship between a secondary counter control code and each counter, and FIG. 5D shows a count threshold LUT. 6, (a) and (b) are diagrams showing the operating range of each counter in the counter circuit, FIG. 7 is a flowchart showing the comparison procedure of the comparison circuit, and FIG. 8 is the encoding device shown in FIG. FIG. 9 is a block diagram of an example of the quantization LUT shown in FIG.
FIG. 0 is an example of a final quantization state, FIG. 11 is a flowchart showing an AC component encoding procedure by the encoding apparatus shown in FIG. 1, and FIGS. FIGS. 13 (a) and (b) show an example of two code LUTs showing the data reading order in the data reading circuit of FIG. 8, and FIGS. 14 (a), (b), (c) and (d). ), (E)
Is a diagram illustrating an example of the contents of a code configuration. 1 ... frame memory, 2 ... readout device, 3 ... 2
Dimensional discrete cosine transform device, 4... Quantizer, 5... Classifier, 6... Block buffer memory, 7.

   ────────────────────────────────────────────────── ─── Continuation of front page       (56) References IEEE TRANSACTIONS                 ON COMMUNICATIONS                 VOL. COM-25 [11] (1977)               P. 1285-1292 (Figure 6)

Claims (1)

(57) [Claims] In a method of compressing gradation image data, which divides digitized gradation image data into a plurality of blocks, and quantizes and encodes transformation coefficients obtained by performing orthogonal transformation for each block, The range of values is divided into a plurality of predetermined sections, the frequency of occurrence of the transform coefficient in at least one of the divided sections is determined, and the quantum of the transform coefficient for each block is determined based on the determined frequency of occurrence. A method of compressing gradation image data, characterized in that a gradation width is determined. 2. 2. The method according to claim 1, wherein the frequency of occurrence of the transform coefficient quantized to zero is obtained as the frequency of occurrence of the transform coefficient. 3. 3. The floor according to claim 1, wherein, when calculating the frequency of occurrence of the transform coefficient in the divided section, the frequency of occurrence of the transform coefficient with respect to the absolute value is obtained. A method for compressing tonal image data. 4. 2. The method according to claim 1, wherein the orthogonal transform is a two-dimensional discrete cosine transform.
Item 4. The method for compressing gradation image data according to item 3 or 3.