JP3952116B2 - Image compression apparatus and method - Google Patents

Description

[0001]
[Technical field to which the invention belongs]
The present invention relates to a digital image compression method, and more particularly to an improvement of a JPEG baseline method.
[0002]
[Prior art]
JPEG is known as a digital still image compression method, and a baseline method using DCT (discrete cosine transform) is widely used among a plurality of JPEG compression methods. The procedure of the JPEG baseline method is described in detail in, for example, Japanese Patent No. 2711665.
[0003]
In this compression method, first, an input bitmap image is divided into blocks of a predetermined size, for example, blocks composed of 8 × 8 pixel values. Next, DCT (discrete cosine transform) is performed on each pixel value block. The DCT newly generates a coefficient block composed of 8 × 8 transform coefficients from each 8 × 8 pixel value block. The first conversion coefficient of this coefficient block indicates the size of the DC (direct current) component of the pixel value block, and the other 63 conversion coefficients indicate the sizes of AC (alternating current) components having different frequencies. Next, the coefficients in each coefficient block are quantized. Next, zigzag scanning is performed on each coefficient block, and the coefficients of each coefficient block are arranged in a sequence in a zigzag order. The coefficient sequence is subjected to two-dimensional Huffman coding (run-length Huffman coding) to reduce the number of bits.
[0004]
In the above compression method, in order to obtain a high compression rate, the characteristics of a general digital still image are successfully used. That is, when DCT is applied to a general digital still image, most of the high-frequency AC conversion coefficients are zero or close to zero in the resulting coefficient block. Therefore, most of the high-frequency AC conversion coefficient becomes zero by the subsequent quantization. Therefore, the coefficient sequence obtained by the next zigzag scan includes many consecutive zero columns. In the subsequent two-dimensional Huffman coding, the number of bits is effectively reduced by performing Huffman coding in consideration of the length (zero run length) of a series of zeros.
[0005]
That is, in this two-dimensional Huffman coding, a combination of a sequence of consecutive zeros appearing in a coefficient sequence and one non-zero coefficient (non-zero coefficient) preceding (or following) is encoded. Treat as a single event. Then, according to the principle of Huffman coding, a shorter code word is assigned to an event having a higher statistical appearance frequency. For example, an event in which the zero run length is zero and the bit length of the subsequent non-zero coefficient is 1 has the highest filing frequency, so the shortest code word is assigned. Also, the event that zero run length is 1 and the bit length of the subsequent non-zero coefficient is 1 is the next most frequently filed, so the next codeword is assigned. Similarly, code words corresponding to the appearance frequency are assigned to other events. In this way, the total number of bits is effectively reduced by the two-dimensional Huffman coding that adopts the two dimensions of zero run length and subsequent non-zero coefficient.
[0006]
It is often the case that all the coefficients behind a certain position in the coefficient sequence are all zero. The best one is a block that is completely filled with one color, and all AC coefficients are zero. For such a case, a code word called EOB (end of block) (specifically, a Huffman code word of coordinates X, Y = 0, 0 in a two-dimensional Huffman table) is provided. EOB means that all subsequent AC coefficients are zero. By this EOB, the total number of bits is further reduced.
[0007]
In an actual apparatus, a two-dimensional Huffman table 100 as illustrated in FIG. 1 in which events and codewords for two-dimensional Huffman coding are defined is prepared on a ROM or RAM, and the two-dimensional Huffman table 100 is referred to. However, each event appearing in the coefficient sequence is converted into a corresponding code word. As shown in FIG. 1, in this two-dimensional Huffman table 100, each event is represented by a combination (X, Y) of a zero run length X and a bit length Y of a subsequent non-zero coefficient. A code word assigned to the event is described in a frame corresponding to (X, Y) (specific code words are not shown).
[0008]
[Problems to be solved by the invention]
As described above, the conventional JPEG baseline method is a two-dimensional Huffman coding that adopts a zero-run length element on the assumption that a sequence of transform coefficients after zigzag scanning includes a lot of consecutive zero columns. Therefore, a high compression effect is achieved.
[0009]
However, the above premise is not always true for extremely high-quality digital images. For example, in a digital image taken with a digital camera whose image quality has been improved in recent years, the frequency of occurrence of zero before EOB in the coefficient sequence after zigzag scanning is very low. Therefore, for the output image of the digital camera, the compression effect by adopting the zero run length element in the Huffman compression after DCT is small. On the contrary, the resource burden of the ROM and RAM increases as much as the two-dimensional Huffman table is prepared, which becomes a problem particularly when the compression device is mounted on a device having a severe ROM restriction or an ASIC having a small number of gates.
[0010]
Accordingly, it is an object of the present invention to provide an image compression method improved from the JPEG baseline method, which can obtain a high compression effect on a high-quality digital image with as little resource burden as possible.
[0011]
Another object of the present invention is to provide an image compression method improved by the JPEG baseline method, which achieves the above-mentioned object and can develop an image with an image developing apparatus of the conventional JPEG baseline method which has already been widely used. Is to provide.
[0012]
[Means for Solving the Problems]
According to the improved JPEG baseline compression method of the present invention, when AC coefficients generated by DCT encoding of a digital image are Huffman encoded, for each one-dimensional event group related only to the value of the AC coefficient, respectively. The assigned Huffman codeword group is used. The Huffman codeword group corresponding only to this coefficient value has a smaller number of codewords than the two-dimensional Huffman table corresponding to the combination of the conventional coefficient value and zero run length, so the resource burden when developing the Huffman table is reduced. light. Also, even if zero-run length elements are not included in Huffman coding, high-quality digital images such as digital camera output images have a low occurrence frequency of zero. Compression ratio.
[0013]
In a preferred embodiment, the above Huffman codeword group does not include codewords assigned to zero AC coefficients. Therefore, when the zero AC coefficient is Huffman-coded, the zero AC coefficient is replaced with another predetermined value (for example, value 1) and converted to a Huffman codeword corresponding to the predetermined value. The reason why the Huffman codeword group does not include the Huffman codeword for the coefficient value zero is that the Huffman codeword group has a run length of zero from the two-dimensional Huffman table used in the conventional JPEG baseline method. This is because only the codeword group in the case is extracted. In this manner, by using the Huffman codeword group included in the conventional two-dimensional Huffman table, it is possible to use a conventional JPEG expansion device when developing an image. Note that even if the AC coefficient value of zero is replaced with the value of 1, it is not perceived as image quality degradation by human vision.
[0014]
In a preferred embodiment, in order to make it possible to use a conventional JPEG decompression apparatus during image decompression, a two-dimensional Huffman table used in the conventional method is used to generate an output bitstream of a compression result. Thus, information for generating a two-dimensional Huffman table in which the portion whose run length is zero matches the codeword group used in the Huffman coding is included in the output bitstream.
[0015]
DETAILED DESCRIPTION OF THE INVENTION
FIG. 1 shows a flow of image compression processing by the improved JPEG baseline method according to an embodiment of the present invention.
[0016]
As shown in FIG. 2, this image compression processing can be roughly divided into four stages: preprocessing 1, DCT encoding 3, entropy encoding 5, and bitstream generation 7. In this embodiment, the same method as the conventional JPEG baseline method is used for the preprocessing 1, the DCT encoding 3 and the bitstream generation 7. In the entropy encoding 5, the same method as the conventional method is used for the processing 51 and 53 for the DC coefficient. Further, among the processes for entropy coding 5 AC coefficients, the zigzag scan 55 uses the same method as the conventional one. Therefore, in this embodiment, a new method is used only for the Huffman coding 57 for the AC coefficient.
[0017]
Hereinafter, since the conventional method is well known, only a simple description will be given, and the new method (Huffman coding 57 for AC coefficients) will be described in detail.
[0018]
In the preprocessing 1, color conversion for converting the color space of the input bitmap image from RGB to YUV (YCbCr) and block processing for subdividing the image of each of the three color components into, for example, blocks of size 8 × 8 are performed.
[0019]
In the next DCT coding 3, first, each block is subjected to two-dimensional DCT, and converted into, for example, a transform coefficient block of size 8 × 8. Next, each coefficient block is quantized using a different quantization step for each transform coefficient. The quantization step for each transform coefficient is defined in a quantization table 9 prepared in advance.
[0020]
In the next entropy coding 5, first, the 64 quantized transform coefficients of each 8 × 8 coefficient block are divided into one DC coefficient and 63 AC coefficients. Then, for the DC coefficient, after performing differential encoding 51, Huffman encoding 53 is performed. In the differential encoding 51, since the DC coefficients of the previous block and the current block often take similar values, the next Huffman encoding 53 is made more effective. The value after the differential encoding 51 and the Huffman codeword assigned to it are defined in the DC Huffman table 11 on the memory.
[0021]
In the process for the AC component of the entropy encoding 5, first, zigzag scanning 55 is performed on the AC coefficient of each coefficient block, and 63 AC coefficients in each block are processed in the order of low frequency first and high frequency after. Arrange them in one sequence. Next, Huffman coding 57 is applied to the coefficient sequence. The Huffman encoding 57 is a one-dimensional Huffman encoding based on only one-dimensional elements such as coefficient bit lengths, without taking in zero-run length elements. The correspondence between the bit length of the coefficient and the Huffman codeword is defined in the AC Huffman table 13 prepared on the memory. This AC Huffman table 13 is the same as the zero run length X = 0 portion, that is, only the vertical column at the left end, extracted from the two-dimensional Huffman table 100 illustrated in FIG. 1 used in the conventional JPEG baseline method. The one-dimensional Huffman table based only on the bit length of the coefficient absolute value.
[0022]
In the final bitstream generation 7, the codeword sequence generated by the entropy encoding 5 is used to generate image information such as the image size and the number of quantization bits, and the color space, quantization table, and DC Huffman table 11 generation. Compression information such as an appearance frequency table and an appearance frequency table for generating an AC Huffman table 13 are added and output as a bit stream of a predetermined format. The contents of the appearance frequency table for generating the AC Huffman table 13 included in this output bitstream are not only those that can be generated only by the one-dimensional Huffman table 13 illustrated in FIG. 3 based on this, but are illustrated in FIG. What can generate the conventional 2D Huffman table 100 of the JPEG baseline method, that is, the appearance frequency table for generating the AC Huffman table included in the output bitstream compressed by the conventional JPEG baseline method Content. Therefore, the image compressed in this embodiment can be expanded using a conventional JPEG image expansion apparatus.
[0023]
Hereinafter, the process of the Huffman compression 57 for the AC coefficient in the entropy encoding 5 will be described in detail. FIG. 3 shows the AC one-dimensional Huffman table 13 used in this processing.
[0024]
In the Huffman compression 57, the bit length of the absolute value of each coefficient appearing in the coefficient sequence obtained by the zigzag scan 55 is handled as one event to be encoded. In the one-dimensional Huffman table 13 illustrated in FIG. 3, the appearance frequency of the total 16 types of events when the bit length Y of the absolute value of the coefficient is each value from 1 to 15 and when the bit length Y is 16 or more is determined. Corresponding codewords are assigned. Among these 16 types of events, the event with Y = 1 has the highest appearance frequency, so the shortest (for example, 2 bits) codeword is assigned, and the event with Y ≧ 16 has the lowest appearance frequency, so it is the longest. Code words are assigned (in FIG. 3, the code words are not shown). In addition, as in the prior art, an EOB (end of block) code word meaning “all subsequent AC coefficients are zero” is also prepared (specifically, coordinates X, Y in the two-dimensional Huffman table) = 0, 0 Huffman codeword), which is also a relatively short (eg, 4 bit) codeword.
[0025]
Note that the one-dimensional Huffman table 13 does not define an event that the coefficient bit length Y = 0 (that is, the coefficient value is zero). The reason is that only the portion of run length = 0 is extracted from the two-dimensional Huffman table 100 of the conventional JPEG baseline method shown in FIG. 1 (there is no Y = 0 on the principle of considering zero run length). This is because the one-dimensional Huffman table 13 is used. In this way, by using the one-dimensional Huffman table 13 having a configuration that matches the conventional two-dimensional Huffman table 100, as described above, the image compressed in the present embodiment is a widely used conventional JPEG image. It is possible to deploy with a deployment device. As will be described later, when zero that is not defined in the one-dimensional Huffman table 13 appears before EOB, the coefficient value zero is replaced with a value “1”.
[0026]
FIG. 4 shows the overall processing flow of Huffman coding 57 for the AC coefficient sequence for each block.
[0027]
First, the end of valid data in the AC coefficient sequence obtained by the zigzag scan 55 is determined (S1). Here, the last valid data indicates a coefficient whose coefficients after the coefficient are all 0 (that is, EOB is placed immediately after that). Details of the process for determining the end of the valid data are shown in FIG. As shown in FIG. 5, after initially setting 64 to the value Index indicating the position (order from the top) of each coefficient in the coefficient sequence (S21), the AC coefficient at the position pointed to by Index while decreasing Index by 1 It is checked whether the value of 0 is 0 (S22, S23). That is, whether or not each AC coefficient is zero is checked in order from the end of the coefficient sequence to the beginning. When the first non-zero AC coefficient is found, the position index of the AC coefficient is returned to the main process in FIG. 4 as the last valid data (S24).
[0028]
Referring again to FIG. 4, next, each AC coefficient is read in order from the beginning of the coefficient sequence (S2), the absolute value of the read coefficient is obtained (S3), and whether or not the absolute value is zero is checked. (S4). As a result, when the absolute value is zero, the absolute value is changed to “1” (S5). Since the zero coefficient is not defined in the one-dimensional Huffman table 13 as described above, the zero is replaced with the closest value “1”. In this way, even if zero is replaced with “1”, there is almost no difference in human vision between zero and “1”, and especially in a high-quality image such as an output image of a digital camera, zero is included in valid data. Since the frequency of occurrence is small, there is almost no substantial image quality degradation.
[0029]
Next, the bit length of the absolute value of the AC coefficient is obtained (S6). Details of this process are shown in FIG. As shown in FIG. 6, first, after initializing the bit length Y to Y = 0 (S31), the byte data indicating the absolute value of the coefficient is shifted to the right by 1 bit (that is, the absolute value is 2). The bit length Y is incremented by 1 (S33). Next, it is checked whether or not the absolute value data right shifted by 1 bit is zero (S33). If it is not zero, step S32 is repeated. When it becomes zero in step S33, the bit length Y obtained by the increment in step S32 is returned to the main processing of FIG. 4 as the bit length of the absolute value of the coefficient (S34).
[0030]
Referring to FIG. 4 again, next, the Huffman code word corresponding to the bit length Y of the coefficient absolute value returned from the processing of FIG. 6 is read from the one-dimensional Huffman table 13 shown in FIG. 3 (S7). The Huffma codeword is written in the bit stream to be output (S8). Since the codeword data read from the Huffman code table 13 in the memory in step S7 has a length in bytes, in step S8, only the bit length portion of the net codeword is extracted from the read bytes. Write to the bitstream.
[0031]
Next, the sign of the coefficient is checked (S9). If the coefficient is positive, the value of the coefficient itself is written into the bitstream (S11). If the coefficient is negative, the coefficient is subtracted from 1 (S10). The difference value is written into the bit stream (S11). Also at this time, only the bit length Y portion of the net coefficient is extracted from the bytes of the original coefficient data and written to the bit stream. Thereby, the number of bits of coefficient data is reduced. For example, in the case of a coefficient having the highest appearance frequency bit length Y = 1, it was originally 1 byte, but it was shortened to 3 bits, for example, the shortest 2 bits Huffman codeword and the net coefficient bit length 1 combined. Is done.
[0032]
The processes of S2 to S11 are repeated until the end of the effective data of the coefficient sequence (S12). Then, it is checked whether or not the last of the valid data is the 63rd (that is, the last of the coefficient sequence) (S13), and if it is before the 63rd, an EOB code word is written in the bit stream (S14). If it is 63rd, the Huffman coding for the coefficient sequence of the block is terminated, and the process proceeds to the Huffman coding of the next block.
[0033]
In the embodiment described above, the Huffman coding for the AC coefficient is performed based only on the coefficient value element without considering the run length element, so that it is only necessary to prepare a one-dimensional Huffman table. Compared to the JPEG compression apparatus, the burden of RAM or ROM is reduced, and in particular, it becomes easy to implement in an apparatus with severe memory restrictions or an ASIC with a limited number of gates. For a high-quality image such as an output image of a recent digital camera, an image quality comparable to that of the conventional image can be secured. Furthermore, since a one-dimensional Huffman table having a configuration adapted to the conventional two-dimensional Huffman table is used, and the output bitstream includes frequency information that can be generated by the conventional two-dimensional Huffman table, image development is performed using conventional JPEG. This can be done using a deployment device.
[0034]
The above embodiment is merely an example for explaining the present invention, and is not intended to limit the present invention only to this embodiment. Therefore, the present invention can be implemented in various forms other than the above-described embodiment. For example, if it is not necessary to perform image expansion using a conventional JPEG expansion device, the frequency information for generating the AC Huffman table included in the output bitstream can be generated only by the AC one-dimensional Huffman table used for compression. In addition, an event that the coefficient value is zero may be added to the AC one-dimensional Huffman table.
[0035]
By the way, in this specification, the image compression method according to the present invention is mainly described, and image development is not particularly described. However, the compression and decompression of the image are in an integrated relationship, and if the compression conversion operation is reversed, the image can be decompressed. Therefore, those skilled in the art will understand the decompression conversion operation if the compression conversion operation is known. Therefore, the technical scope of the present invention described in each claim includes not only a compression device and a method exactly as described, but also a deployment device and a method that are in a relationship of two sides of the same.
[Brief description of the drawings]
FIG. 1 is an explanatory diagram showing a two-dimensional Huffman table used for Huffman coding of AC coefficients in a conventional JPEG baseline method.
FIG. 2 is a block diagram showing the configuration of an embodiment of the present invention.
FIG. 3 is an explanatory diagram showing a one-dimensional Huffman table used for Huffman encoding of AC coefficients in the embodiment.
FIG. 4 is a flowchart showing an overall flow of Huffman encoding of AC coefficients in the embodiment.
FIG. 5 is a flowchart of processing for detecting the end of valid data of an AC coefficient sequence.
FIG. 6 is a flowchart of processing for obtaining the bit length of the absolute value of an AC coefficient.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Pre-processing part 3 DCT encoding part 5 Entropy encoding 7 Bit stream production | generation part 9 Quantization table 11 DC one-dimensional Huffman table 13 AC one-dimensional Huffman table

Claims (7)

In an apparatus for compressing digital images by the JPEG baseline method,
AC Huffman encoding means for Huffman encoding AC coefficients generated by DCT encoding of the digital image;
A one-dimensional Huffman table for AC that defines a first Huffman codeword group assigned to each one-dimensional event group related only to the value of the AC coefficient;
With
The AC Huffman encoding means uses the AC one-dimensional Huffman table to perform Huffman encoding based only on the value of the AC coefficient without considering the zero-run length of the AC coefficient generated by the DCT encoding. And
The first Huffman codeword group does not include the Huffman codeword assigned for the event that the value of the AC coefficient is zero;
When the event that the value of the AC coefficient is zero is Huffman encoded, the AC Huffman encoding means replaces the value of the AC coefficient that is zero with another predetermined value, and corresponds to the predetermined value. An image compression apparatus that converts a Huffman codeword assigned to an event.   The image compression apparatus according to claim 1, wherein the another predetermined value is one. Bitstream generating means for generating an output bitstream including the result of the Huffman encoding;
The bit stream generation means is assigned to each of a two-dimensional event group related to a combination of a non-zero AC coefficient value and a zero run length of the AC coefficient, and the zero stream length is zero when the zero run length is zero. The image compression apparatus according to claim 1, wherein an appearance frequency table for generating a second Huffman codeword group that matches one Huffman codeword group is included in the output bitstream. In a method of compressing a digital image by the JPEG baseline method,
When Huffman coding the AC coefficient generated by DCT coding of the digital image, the first Huffman codeword group assigned to each one-dimensional event group related only to the value of the AC coefficient is used. Performing Huffman coding based only on the value of the AC coefficient without considering the zero-run length of the AC coefficient generated by the DCT coding,
The first Huffman codeword group does not include the Huffman codeword assigned for the event that the value of the AC coefficient is zero;
When the event that the value of the AC coefficient is zero is Huffman coded, the value of the AC coefficient that is zero is replaced with another predetermined value, and the Huffman codeword assigned to the event corresponding to the predetermined value A method of compressing an image, wherein:   5. The image compression method according to claim 4, wherein the another predetermined value is 1.   When generating an output bitstream including the result of the Huffman coding, each is assigned to a two-dimensional event group related to a combination of a non-zero AC coefficient value and a zero run length of the AC coefficient, and 5. The appearance bit table for generating a second Huffman codeword group that matches the first Huffman codeword group when a zero run length is zero is included in the output bitstream. Image compression method. When compressing digital images using the JPEG baseline method,
When Huffman coding the AC coefficient generated by DCT coding of the digital image, using a first Huffman codeword group respectively assigned to a one-dimensional event group related only to the value of the AC coefficient, Huffman coding is performed based only on the value of the AC coefficient without considering the zero-run length of the AC coefficient generated by the DCT coding,
The first Huffman codeword group does not include the Huffman codeword assigned for the event that the value of the AC coefficient is zero;
When the event that the value of the AC coefficient is zero is Huffman coded, the value of the AC coefficient that is zero is replaced with another predetermined value, and the Huffman codeword assigned to the event corresponding to the predetermined value A computer-readable recording medium carrying a program for operating a computer to convert to a computer.

Images