Method of and apparatus for compressing image data.

BACKGROUND OF THE INVENTION

Field of the Invention

0001The present invention relates to a method of and an apparatus for compressing image data having graduation with respect to every pixel block for reducing the amount of data of image information

Description of the Prior Art

0002In an image processor such as a process scanner, image data is obtained for every pixel, which is a constant small area, by reading an original image. The image data employed in the field of printing are usually larger in amount than those employed in television, and the amount of the information thereof reaches several to tens of mega bites for each image. Such a large amount of image data would require enourous memory capacity if they were stored in a data base as they were; thus the cost for data transmission would increase greatly

0003In order to cope with this problem, data compression is executed on the image data for reducing the total amount of information. Several types of data compression methods are well known in the art, such as that using orthogonal transformation and the like.

0004However, the data compression methods in the art suffer from a trade-off between a compression rate (or compressibility) and quality of a reproduced image produced by the compressed image data. If the compression rate is increased, the quality of the image is lowered. On the other hand, the compression rate is decreased in order to raise the quality of the image. This is not a big issue for the image data employed in a television because the quality of the images of the television is relatively low compared with that employed in the field of printing.

0005The data compression methods in the art have another issue concerning repetition of data compression and data restoration. If the data compression is execused by high compression rate, the image quality of a restored image obtained by restoring the compressed data is lowered as compared with the original image. Thus, when the data compression and restoration are repeated several times, the image quality is gradually lowered.

SUMMARY OF THE INVENTION

0006The present invention is directed to a method of and an apparatus for compressing image data depicting an image with respect to each pixel block including a plurality of pixels, the method comprising the steps of: (a) executing an orthogonal transformation operation on the image data with respect to each pixel block, to thereby obtain a coefficient table including coefficients of a series of orthogonal functions of the image data for each pixel bolck, (b) preparing a plurality of code tables, each of the plurality of code tables validating a part of the coefficients of the coefficient table, (c) obtaining valid coefficients of the coefficient table with respect to each code table, the valid coefficients being validated by each code table, (d) selecting a most relevant code table most relevant to the coefficient table within the plurality of code tables on the basis of the valid coefficients, and (e) collecting the valid coefficients validated by the relevant code table, to thereby form compressed data comprising the valid coefficients for each pixel bolck.

0007Preferably, the step (d) comprises the steps of: (d-­1) computing representative values for respective code tables, each of the representative values being correlated with a sum of absolute values of the valid coefficients, and (d-2) comparing the representative values with each other, thereby selecting the most relevant code table.

0008According to an aspect of the present invention, the plurality of code tables are classified into a plurality of groups, and the step (c) comprises the steps of: (c-1) preparing a statistical value indicative of a standard deviation of image data for each pixel block, (c-2) selecting one of the plurality of groups for each pixel block in response to the statistical value, and (c-3) computing the valid coefficients only for code tables included in the selected group.

0009The present invention is also directed to a method of and an apparatus for compressing image data depicting an image with respect to each pixel block including a plurality of pixels, the method comprising the steps of: (a) preparing a statistic value indicative of a standard deviation of the image data within each pixel block, (b) selecting one of prescribed procedures of data compression in response to the statistic value, (c) executing data compression on the image data according to a selected procedure, to thereby obtain first compressed data representing the image, (d) storing the first compressed data with identification data specifying the selected procedure, (e) executing data restoration on the first compressed data, to thereby obtain restored image data depicting the image and including the identification data therein, (f) identifying the selected procedure by means of the identification data included in the restored image data, and (g) executing data compression on the restored image data according to the selected procedure identified in the step (f), to thereby obtain second compressed data representing the image.

0010Preferably, the restored data includes a plurality of components each of which is formed as digital data and the identification data is expressed by a least significant digit of a component of the restored data, the component representing a part of the image at a prescribed position in one pixel block.

0011According to another aspect of the present invention, each of the prescribed procedure comprises the steps of: (1) executing an orthogonal transformation operation on the image data with respect to each pixel block, to thereby obtain a coefficient table including coefficients of a series of orthogonal functions indicative of the image data for each pixel block, (2) preparing a plurality of code tables peculiar to each of the prescribed procedures, each of the plurality of code tables validating a part of the coefficients of the coefficient table, (3) obtaining valid coefficients of the coefficient table with respect to each code table, the valid coefficients being validated by each code table, (4) selecting a most relevant code table most relevant to the coefficient table within the plurality of code tables on the basis of the valid coefficients, and (5) collecting the valid coefficients validated by the relevant code table, to thereby form compressed data comprising the valid coefficients for each pixel block.

0012Further, each of the plurality of code tables for a same procedure may validate a same number of the coefficients.

0013Moreover, the coefficients of the coefficient table may be expressed by means of digital data, and repective code elements in each of the plurality of code tables may be designate respective numbers of effective digits for respective valid coefficients.

0014As a function of the relevant code table, only a part of coefficients in the coefficient table are validated. Accordingly, the number of coefficients, that is, the amount of information representing the image is fairly reduced.

0015The restored image data including the identification data enables to execute data compression on the restored image data by the same sequence which was emplyed to obtain the first compressed data. Accordingly, the deterioration of the image quality is suppressed even if data compression and restoration are repeared several times.

0016When the identification data are expressed by means of the least significant digit of the restored data, the total amount of the restored image data is further reduced because the identification data does not exist independently.

0017Acccordingly, an object of the present invention is to increase the compression rate (or compressibility) without lowering the image quality too much.

0018Another object of the present invention is to suppress the deterioration of the image qaulity when the data compression and restoration are repeated several times.

0019These and other objects, features, aspects and advantages for the present invention will become more apparent from the following detailed description of the present invention when taken in conjection with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

0020<ul id="ul0001" list-style="none"><li>Fig.1 is an explanatory diagram showing pixel blocks in an image;</li><li>Fig. 2 is an explanatory diagram showing pixel array in a pixel block;</li><li>Fig. 3 is an explanatory diagram showing examples of image data and transformation coefficients;</li><li>Fig. 4 is a conceptual diagram showing an exemplary code table;</li><li>Fig. 5 is an explanatory diagram showing exemplary relation between standard deviation of image data and numbers of appearance thereof;</li><li>Fig. 6 is a schematic block diagram showing an apparatus to which an embodiment of the present invention is applied;</li><li>Figs. 7A, 7B and 8 are flow charts for illustrating the embodiment of the present invention;</li><li>Fig. 9 is a conceptual diagram showing an exemplary subtable;</li><li>Figs. 10A and 10B are conceptual diagrams showing the structure of compressed image data;</li><li>Fig. 11 is a schematic block diagram showing a restoration apparatus;</li><li>Fig. 12 is a flow chart showing the procedure of a method of recording group information;</li><li>Figs. 13, 15 and 16 show how the sheets containing Figs. 13A, 13B, 15A, 15B, 16A and 16B are to be arranged in order to stand in the proper relation;</li><li>Figs. 13A, 13B, 15A, 15B, 16A and 16B when arranged according to Figs. 13, 15 and 16 are explanatory diagrams showing exemplary cycles of compression and restoration, and</li><li>Fig. 14 illustrates SN ratios of restored images.</li></ul>

DESCRIPTION OF THE PREFERRED EMBODIMENT

0021A preferred embodiment of the present invention will be described in sequence with reference to the following items: <ul id="ul0002" list-style="none"><li>A. a method of producing compressed image data;</li><li>B. a method of producing restored image data;</li><li>C. structure of an aparratus and operation thereof;</li><li>D. exemple of data compression and restoration ; and</li><li>E. modifications.</li></ul>

0022Fig. 1 is a conceptual diagram showing pixel blocks in an image 1. The image 1 is composed of XxY pixels P, and is divided into VxH pixel blocks B_vh each having a plurality of pixels. Fig. 2 is a conceptaul diagram showing the structure of one pixel block B_vh. The pixel block B_vh is composed of MxN pixels P_ij and image data f_ij is obtained for every pixel. Although the pixel block B_vh is composed of 8x8 pixels in the example shown in Fig. 2, the same may be composed of 4x4 pixels or 16x16 pixels.

0023When data compression is executed, the image data are first subjected to orthogonal transformation with respect to every pixel block B_vh in the image 1. Several types of the orthogonal transformation can be applied such as discrete Fourier transformation, Hadamard's transformation and the like. In this embodiment, a Discrete Cosine Transformation (hereinafter referred to as "DCT") is employed. The following equation expresses a coefficient F_mn associated with a specific function in a series of transformation functions for DCT.<maths id="math0001" num=""><img file="EP0349847A2_D0001.tif" /></maths><u style="single">m</u> and <u style="single">n</u> have values generally within ranges of zero to (M - 1) and zero to (N - 1), respectively.

0024Fig. 3 is an explanatory diagram showing the image data f_ij within one pixel block and the transformation coefficient F_mn obtained through the expression (1). Fig. 3 (a-1) and Fig. 3(b-2) show tables FT which have the transformation coefficients F_mn. A tranformation coefficient F₀₀ shows an average value (hereinafter referred to as "direct current component") of the image data f_ij, and remaining tranformation coefficients F_mn (hereinafter referred to as "alternate current components") show distribution of the image data f_ij. Each of the tranformation coefficients F_mn is expressed by means fo binary data of 8 bit, for example. However, the values of the coefficients are shown in decimal notation in Fig. 3, for convenience of illustration.

0025As can be understood from Fig. 3, the transformation coefficients F_mn show the following character. As well as positions (m,n) at which the transformation coefficients F_mn are nonzero, the values of the coefficients F_mn are considerably varied in response to the distribution of the image data f_ij. Those of the transformation coefficients F_mn having large absolute values sprecifically represent the distribution of the original image data f_ij, and the number of those tranformation coefficients F_mn having large absolute values is increased in response to spatial uneveness of the image data f_ij.

0026The inventors of the present invention have directed their attention to the character of the transformation coefficients F_mn, to employ a code table which validates only a part of the alternate current components in a coefficient table FT. Fig. 4 is an explanatory diagram showing a code table. Within this code table CT, only a part of coordinates (m,n) are supplied with an effective bit number I_mn. The effective bit number I_mn indicates that the transformation coefficient F_mn in the position (m,n) is validated by the table CT. The effective bit number I_mn further indicates the maximum bit number for expressing the transformation coefficient F_mn in binary system. That is, the value of the transformation coefficient F_mn no more than a maximum effective value F_max (= (2^Imn) -1) is validated as it is. Referring to Fig. 4, symbol "-" indicates that the effective bit number is zero. All of the effective bit numbers I_mn may be of the same value (8 bits, for example). However, the compression rate is raised by employing appropriate values (such as 8 bits or 3 bits, for example) which are smaller than the bit number (8 bits) of the original transformation coefficients F_mn, and which are varied with the coordinates (m,n). Even if a transformation coefficient F_mn has a value larger than the maximum effective value F_max, its value is limited to the maximum effective value F_max by the effective bit number I_mn. The effective bit numbers I_mn are empirically determined to avoid such situation as much as possible and even if the transformation coefficient F_mn is limited to the maximum effective value F_max, influence thereof can be neglected.

0027A plurality of the code tables CT are prepared in advance of the data compression. The respective code tables CT have different patterns and values of the effective bit numbers I_mn from each other. The patterns of the effective bit numbers I_mn are empirically determined in order to reproduce an image of higher image quality with a smaller number of the transformation coefficients. For example, the coefficient tables FT like those shown in Fig. 3 are first prepared for typical images such as a human figure, a landsape, a still life and the like. Then, by comparing the distribution of the coefficients in those coefficient tables FT with each other, the code tables are defined so as to represent major types of the distribution of the coefficients frequently seen in those coefficient tables.

0028When the transformation coefficient table FT is obtained for every pixel block B_vh in data compression operation, an optium code table (or a relevant code table) is selected for each coefficient table FT from the plurality of code tables CT. For the purpose of such selection, the sum S_F of the effective transformation coefficients F_mn (hereinafter simply referred to as "effective sum") is obtained in accordance with the following expression:<maths id="math0002" num=""><img file="EP0349847A2_D0002.tif" /></maths>

0029The effective sum S_F is obtained for each code table CT with respect to one coefficient table FT. The code table CT having the maximun effective sum is selected as the optium one to the coefficient table FT. The reliability of the selection of the optium code table as a funcion of the effective sum S_F is further improved if the same number N_c of the transformation coefficients F_mn (hereinafter refered to as "effective component numbers") are validated by any of the respective code tables CT. The optium code table is identified by a table number defined for each code table CT.

0030Compressed image data are produced on the basis of an average value of the original image data f_ij, the table number of the optium code table CT and the values of the transformation coefficients F_mn validated by the optium code talbe CT. The effective component number N_c is far smaller than the total number MxN of the transformation coefficients F_mn within a coefficient table and the value of each transformation coefficient F_mn is expressed by a smaller effective bit number I_mn, whereby the compression rate is quite increased. The compressed image data can be encoded by well-known Huffman codes or the like, in order to further increase the compression rate.

0031In this embodiment, the code tables CT are classified into a plurality of groups in response to characteristics of the image in order to still further improve the compression rate. To this end, the code tables CT are classified into a plurality of compression groups (code table groups) in respone to a standard deviation σ of the image data f_ij obtained for every pixel block B_vh as follows: <tables id="tabl0001" num="0001"><table frame="all"><tgroup cols="2" colsep="1" rowsep="0"><colspec colnum="1" colname="col1" colwidth="78.75mm" /><colspec colnum="2" colname="col2" colwidth="78.75mm" /><tbody valign="top"><row><entry namest="col1" nameend="col1" align="right">σ &lt; σ₁:</entry><entry namest="col2" nameend="col2" align="left">compression group g₁ including code tables CT₁₁ to CT_1L</entry></row><row><entry namest="col1" nameend="col1" align="right">σ₁ ≦ σ &lt; σ₂:</entry><entry namest="col2" nameend="col2" align="left">compression group g₂ including code tables CT₂₁ to CT_2L</entry></row><row><entry namest="col1" nameend="col1" align="right">σ₂ ≦ σ &lt; σ₃:</entry><entry namest="col2" nameend="col2" align="left">compression group g₃ including code tables CT₃₁ to CT_3L</entry></row><row><entry namest="col1" nameend="col1" align="right">σ₃ ≦ σ :</entry><entry namest="col2" nameend="col2" align="left">compression group g₄ including code tables CT₄₁ to CT_4L</entry></row><row rowsep="1"><entry namest="col1" nameend="col2" align="justify">(where σ₁, σ₂ and σ₃ represent prescribed threshold values)</entry></row></tbody></tgroup></table></tables> Number L of the code tables may be varied with the compression groups g₁ to g₄. Fig. 5 illustrates distribution G of frequency A of the standard deviation σ in an overall image 1. Referring to Fig. 5, the frequency A is shown in logarithmic scale along the vertical axis.

0032The first threshold value σ₁ is determinerd such that the image in the pixel block is fairly even when the standard deviation σ is less than the same. In such a pixel block, because the alternate current components of the transformation coefficients are small, the image quality is not so much lowered even if a code table having a small value of the effective component number N_c is employed for the data compression. Since the frequency A of a pixel block having smaller standard deviation σ is generally higher as understood from Fig. 5, the compression rate is much improved by treating all of such pixel blocks with the code tables having a small effective component number N_c.

0033When the standard deviation σ is between the first and second threshold values σ₁ and σ₂, this indicates that the image is loosely changed in the pixel block. In this case, it is better to employ a code table having an effective component number N_c larger than that of the compression group g₁, in order to maintain the image quality. Since this compression group g₂ also corresponds to a range of relatively high frequency A, the compression rate attained by employing the code tables having a relatively small value of the effective component number N_c in the compression group g₂ is also high.

0034While the frequency A becomes relatively low with respect to a range of the standard deviation σ more than the second threshold value σ₂, the compression rate would be decresed if all of such pixel blocks were processed with a code table having a large effective component number N_c. Thus, the third threshold value σ₃ is employed for further classification of the code tables.

0035Respective effective component numbers N_c1 to N_c4 of code tables CT_1k to CT_4k (k= 1 to L) of the respective compression groups g ₁ to g₄ are in the following relation: 0 ≦ N_c1 ≦ N_c2 ≦ N_c3 ≦ N_c4 ≦ MxN - 1      (3)

0036The effective component numbers of the code tables are constant within the respective compression groups. When N_c1 = 0, the compressed image data of a pixel block B_vh belonging to the compression group g₁ is expressed only by the average of the pixel data f_ij, that is, only by the direct current component F₀₀.

0037Prior to the data compression operation, the threshoid values σ₁ to σ₃ and the code tables CT_1k to CT_4k classified into the compression groups g₁ to g₄ are prepared. One of the compression groups g₁ to g₄ is selected in response to the standard deviation σ of the block B_vh, and the optium code table is selected from code tables included in the selected compression group.

0038Restored image data f_ij in a pixel block are obtained by executing Inverse Discreet Cosine Transformation (or IDCT) on the transformation coefficients F_mn of compressed image data.

0039IDCT is expressed as follows:<maths id="math0003" num=""><img file="EP0349847A2_D0003.tif" /></maths>

0040The restored image data is so formed as to indicate which one of the compression groups g₁ to g₄ is employed for producing the compressed image data. Such information indicating the compression group is hereinafter referred to as "group information".

0041In order to record the group informaiton in the restored image data f_ij, similar information is also restord in the compressed image data. In the compressed image data, the code table number has a function of representing the group information, as hereinafter described.

0042The group information is integrated into the restored image data f_ij by using the least significant digits of the restored image data at prescribed pixel positions within a pixel block, for example.

0043When the restored image data are expressed in binary system, two values ("0" and "1") of a least significant bit distinguish two compression groups. The value of "0" of a least significant bit is hereinafter simply referred to as "even" while the value "1" is referred to as "odd".

0044Since four compression groups are employed in this embodiment, the group information can be expressed by two bits. Within the pixel block B_vh shown in Fig. 2, two image data f₁₁ and f₃₃ are selected and their values are adjusted as follows to record the group information: <tables id="tabl0002" num="0002"><table frame="all"><tgroup cols="2" colsep="1" rowsep="0"><colspec colnum="1" colname="col1" colwidth="78.75mm" /><colspec colnum="2" colname="col2" colwidth="78.75mm" /><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left">compression group g₁ :</entry><entry namest="col2" nameend="col2" align="left">both of f₁₁ and f₃₃ are made even</entry></row><row><entry namest="col1" nameend="col1" align="left">compression group g₂ :</entry><entry namest="col2" nameend="col2" align="left">f₁₁ is made odd and f₃₃ is made even</entry></row><row><entry namest="col1" nameend="col1" align="left">compression group g₃ :</entry><entry namest="col2" nameend="col2" align="left">f₁₁ is made even and f₃₃ is made odd</entry></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left">compression group g₄ :</entry><entry namest="col2" nameend="col2" align="left">both of f₁₁ and f₃₃ are made odd</entry></row></tbody></tgroup></table></tables>

0045The image data f₁₁ is made even in accordance with the following expression, for example: f₁₁ = ff₁₁ &amp; 254      (5) where ff₁₁ : image data before being made even f₁₁ : image data after being made even &amp; : bit-AND operator

0046The bit-AND operator &amp; expresses operation of obtaining logical product for each bit of binary numbers. The expression (5) is expressed in 8-bit binary numbers as follows; e₇ e₆ e₅ e₄ e₃ e₂ e₁ e₀ = a₇ a₆ a₅ a₄ a₃ a₂ a₁ a₀ &amp; 1 1 1 1 1 1 1 0      (6) where e_i (i = 0 to 7) : each bit of image data f₁₁ a_i (i = 0 to 7) : each bit of image data ff₁₁ Through the function of the bit-AND operator &amp;, the expression (6) is transformed as follows: e₇ e₆ e₅ e₄ e₃ e₂ e₁ e₀ = a₇ a₆ a₅ a₄ a₃ a₂ a₁ 0      (7) That is, whatever value the image data ff₁₁ has, the least signification bit e₀ of the image data f₁₁ is always zero after the transformation by the expression (5). Further, the difference between the image data f₁₁ and ff₁₁ is not more than 1 and hence influence of the difference on the image quality can be neglected.

0047On the other hand, the image data is made odd through the following expression: f₁₁ = (ff₁₁ &amp; 254) + 1      (8) Since the least signification bit of the image data f₁₁ in the left hand side is always 1, the image data f₁₁ is an odd number. Further, since the maximum value of the first term (f₁₁ &amp; 254) in the right hand side is 254, the maximum value of the image data f₁₁ is 255. Thus, the maximum value 255 of the image data is not exceeded by the operation of the expression (8). Further, the difference between the image data f₁₁ and ff₁₁ is not more than 1 and hence influence of the difference on the image quality can be neglected.

0048Although the above description has been made on the operation of the image data f₁₁, the same also applies to the image data f₃₃.

0049As a funtion of the least signification bits e₁₁₀ and e₃₃₀ of the image data f₁₁ and f₃₃ after being made even or odd, the group information N_g is expressed in the following 2 bit binary number: N_g = e₃₃₀ e₁₁₀      (9) Relation between the group information N_g and the compression groups g₁ to g₄ is as follows: <tables id="tabl0003" num="0003"><table frame="all"><tgroup cols="2" colsep="1" rowsep="0"><colspec colnum="1" colname="col1" colwidth="78.75mm" /><colspec colnum="2" colname="col2" colwidth="78.75mm" /><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left">compression group g₁ : N_g = 00</entry><entry namest="col2" nameend="col2" align="left">(10a)</entry></row><row><entry namest="col1" nameend="col1" align="left">compression group g₂ : N_g = 01</entry><entry namest="col2" nameend="col2" align="left">(10b)</entry></row><row><entry namest="col1" nameend="col1" align="left">compression group g₃ : N_g = 10</entry><entry namest="col2" nameend="col2" align="left">(10c)</entry></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left">compression group g₄ : N_g = 11</entry><entry namest="col2" nameend="col2" align="left">(10d)</entry></row></tbody></tgroup></table></tables>

0050The respective bits forming the group information N_g are not restricted to those of the image data f₁₁ and f₃₃, but the least significant bits at two arbitrary pixel positions can be employed. Although the group information N_g is expressed in a 2-bit binary number in this embodiment because of four compression groups, the bit number may be increased if more compression groups are employed.

0051As hereinabove described, the restored image data are so formed as to include the group information, whereby the restored image data can be identified as being in the same compression group as the previous one in which the original image data before compression were included.

0052Fig. 6 is a schematic block diagram showing the structure of an image data compressor emboding the present invention. Referring to Fig. 6, an image data compressor AP comprises an image memory 5, a block read controller 6, a buffer memory 7, a two-dimentional DCT transformer 8, a linear quantizer 9, an encoder 10, a recording controller 11, a group-information read controller 12, a standard deviation calculator 14, a group information controller 15, a ROM 17 and a controller 13. The operation of the image data compressor AP will be described along flow charts shown in Figs. 7A and 7B.

0053At a step S1, the threshold values σ₁ to σ₃ of the standard deviation are inputted through a parameter input terminal 16 provided in the exterior of the data compressor AP, and set in the group informaiton controller 15. Steps S2 to S4 and S11 to S13 are adapted to sequentially select pixel blocks B_vh to be processed.

0054At a step S5, the block read controller 6 reads image data f₁₁ within one pixel block B_vh, and stores the same in the buffer memory 7.

0055A step S6 is adapted to determine the group information N_g. At a step S61, a decision is made as to whether or not the image data f_ij are those of a restored image. If the image data f_ij are those of a restored image, the image data f_ij has data indicating that they are restored image data, in a header portion of a data file of the image data f_ij, for example. The aforementioned decision is made in accordance with presence/absence of such data.

0056If the image data f_ij are not those of a restored image but those of an original, steps S62 and S63 are carried out. At the step S62, the image data f_ij are supplied from the buffer memory 7 to the standard deviation calculator 14, which in turn calculates the standard deviation σ. At the step S63, the standard deviation σ is supplied from the calculator 14 to the group information controller 15, which in turn determine the group information N_g, That is, the group information controller 15 compares the standard deviation σ with the threshold values σ₁ to σ₃, to determine which one of the compression groups g₁ to g₄ the pixel block belongs to. Then the controller 15 determines the group information N_g by the expressions (10a) to (10d).

0057When the image data f_ij are those of a restored image, on the other hand, a step S64 is carried out. At this step, the group information N_g is synthesized by respective least significant bits e₁₁₀ and e₃₃₀ of image data f₁₁ and f₃₃ at two pixel positions, as shown by the expression (9). Referring to Fig. 6, the image data f₁₁ and f₃₃ are read out from the buffer memory 7 in the group-­information read controller 12. The read controller 12 synthesizes the group information N_g through the image data f₁₁ and f₃₃, and supplies the same to the group information controller 15. Thus, when the group information N_g is supplied from the read controller 12, the controller 15 holds the group information N_g and performs no operation of determining the group information N_g on the basis of the standard deviation σ.

0058When the group information N_g is determined at the step S6, the image data f_ij are transmitted from the buffer memory 7 to the two-dimentional DCT transformer to be subjected to DCT at a step S7.

0059At a step S8, the transformation coefficients F_mn are linearly quantized. As is well known in the art, "linear quantization" means proccessing of dividing a transformation coefficient F_mn by a predescribed value α ("quantum width") and making the value F_mn/α to be integral when the values F_mn/α includes decimal part. The value of the quantum width α may be fixed, or may be varied with the standard deviation σ. Such linear quantization may not be performed, as hereinafter described.

0060At a step S9, an optimum code table is selected with respect to a coefficient table FT* having the lineary quantized transformation coefficients F_mn*.

0061Fig. 8 is a flow chart showing the procedure of selecting the optimum code table carried out by the encoder 10 at the step S9. Referring to Fig. 8, meaning of each symbol is as follows: <tables id="tabl0004" num="0004"><table frame="all"><tgroup cols="2" colsep="1" rowsep="0"><colspec colnum="1" colname="col1" colwidth="78.75mm" /><colspec colnum="2" colname="col2" colwidth="78.75mm" /><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left">max:</entry><entry namest="col2" nameend="col2" align="left">parameter for obtaining maximum value of effective sum S_F</entry></row><row><entry namest="col1" nameend="col1" align="left">Max:</entry><entry namest="col2" nameend="col2" align="left">parameter representing a code table number of a code table having the maximum effective sum S_F</entry></row><row><entry namest="col1" nameend="col1" align="left">ℓ:</entry><entry namest="col2" nameend="col2" align="left">parameter representing a code table number</entry></row><row><entry namest="col1" nameend="col1" align="left">L:</entry><entry namest="col2" nameend="col2" align="left">number of the code tables</entry></row><row><entry namest="col1" nameend="col1" align="left">c:</entry><entry namest="col2" nameend="col2" align="left">parameter representing a serial number of a effective current component</entry></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left">N_c:</entry><entry namest="col2" nameend="col2" align="left">number of the effective alternate current components</entry></row></tbody></tgroup></table></tables>

0062At a step S21, the compression group of the pixel block under process is identified as a function of the compressing group information N_g. Steps S22 to S30 are adapted to select that code table within L pieces of code tables of the compression group, which code table has the maximum effective sum S_F (see expression (2)).

0063At the steps S23 to S26 for obtaining the effective sums S_F, a subtable T shown in Fig. 9, which is stored in the ROM 17, is employed. The subtable T specifies coordinate values (m_ℓc, n_ℓc) of the coordinate positions in which the effective bit numbers I_mn are nonzero. The subtables T are prepared for respective code tables CT_g1 to CT_gL within one compression group g_i (i = 1 to 4). Where ℓ = 0 to (L - 1) and c = 0 to (N_c -1). The subscript g of the code tables CT_g1 to CT_gL represent the compression group g_i. As to one code table CT_g, the coordinates (m_ℓc, n_ℓc) are sequentially called from the subtable T, and absolute values of the transformation coefficients F_mn* of the transformation coefficient table FT* at the coordinates (m_ℓc, n_ℓc) are added up at the step S24.

0064The code table having the maximum effecitve sum S_F is selected as the optimum code table. When the optimum code table is thus selected, the effective transformation coefficients F_mn* of the coefficient table FT* are specified by the optimum code table and are expressed in the effective bit numbers I_mn (six bits, for example), while the other transformation coeffcients F_mn* are neglected. At a step S31, the table number Max of the optimum code table and the transformation coefficients F_mn* specified as effective by the optimum code table are transformed into a Huffman code train, to form compressed image data (steps S31 and S10 ). The average value of the original image data f_ij may be simultaneously Huffman-coded at this time, while only the average value may be encoded by a method such as the so-­called predictive encoding, independently of the step S31.

0065Compressed image data (encoded data) D_f thus obtained are transmitted to the data recording controller 11 from the encoder 10, and further outputted to the image memory 5 or a transmission path for an external circuit (not shown). A selection command for this output is inputted in the parameter input terminal 16 by an operater in advance of the compression of the image data, and initialized in the data recording controller 11.

0066Figs. 10A and 10B are conceptual diagrams showing the file structure of the encoded compressed image data D_f. A header 18 of Fig. 10A includes the threshold values σ₁, σ₂ and σ₃ etc., and the code trains of respective pixel blocks are arranged following the header 18. Fig. 10B is the internal structure in a code train 19 of each pixel block, which is formed of the table number Max of the optimum code table and the code train of the transformation coefficients F_mn* specified as effective by the optimum code table.

0067Fig. 11 is a schematic block diagram showing a restoration apparatus for restoring the encoded compressed image data D_f. This restoraiton aparatus IAP comprises an image memory 20, a read controller 22, a buffer memory 23, a decoder 24, a ROM 25, an inverse quantizer 26, a two-­dimentional IDCT Transformer 27, a group information recoder 28 and a recording controller 29.

0068The compressed image data D_f are read out from the image memory 23 or a transmission path 21 by the read controller 22, and supplied to the buffer memoy 23 and further transmitted to the decoder 24. The ROM 25 stores a code table identical to that employed for the compression of the image data, a Huffman code table for decording the Huffman code train, and the like. The compressed image data D_f are decoded into the transformation coefficients F_mn* by the decoder 24. The transformation coefficients F_mn* are inverse-quantized by the inverse quantizer 26. That is, the transformation coefficients F_mn* (=F_mn/α) linearly quantized with the quantum width α is multiplied by the quantum width α. The transformation coefficients F_mn thus obtained are inverse-converted by the two-dimentional IDCT Transformer 27, whereby the image data ff_ij of each pixel within the pixel block are obtained.

0069On the basis of the image data ff_ij thus obtained, the group information recoder 28 records the group information in the image data ff_ij. Fig. 12 is a flow chart showing the recording procedure. At a step S31, the group information recoder 28 reads out the code table number Max within the compressed image data D_f from the buffer memory 23. At a step S32, the group information N_g is determined from the code table number Max. Through steps S33a to S33c and S34a to S34d, the image data ff₁₁ (f₁₁) and ff₃₃ (f₃₃) are made odd/even. The steps S34a to S34d are carried out through the expressions (5) and (8).

0070The image data f_ij in which the group information N_g is recorded are stored in the image memory 20 by the recording controller 28. Image data f_ij for the overall image can be obtained by performing the restoration processing operation on all of the pixel blocks.

0071In order to re-compress image data f_ij once restored, the group information N_g is read out from the least significant bits of two image data f₁₁ and f₃₃. Data compression is performed according to the compression group identified by the group information N_g. Firstly, an example 1 will be described in which the linear quantization of F_mn is not performed.

0072Figs. 13A and 13B are explanatory diagrams showing image data f_ij and transformation coefficients F_mn in a first example of repeating compression and restoration. Image data f_ij (0) of Fig. 13A(a) represents image data of an original, and f_ij(k) (k = 1 to 3) of Fig. 13A(d), Fig. 13B(f) and (h) represents image data obtained by the k-th restoration after alternate data compression and restoration. F_mn(k) (k = 1 to 3) of Fig. 13A(c), Fig. 13B(e) and (g) represent transformation coefficients obtained by the k-th compression after alternate data compression and restoration.

0073In this example, the transformation coefficients F_mn are not subjected to linear quantization. Therefore, the transformation coefficients F_mn of decimal values obtained by DCT are not made to be integers but directly preserved as the compressed image data D_t and restored therefrom. In Figs. 13A and 13B, only integral parts of the transformtion coefficients F_mn are shown for convenience of illustration. On the other hand, the image data f_ij(k) are regularly preserved and restored in integral values.

0074First, the image data f_ij(0) of the original are DC transformed to the transformation coefficients F_mn ⁰(1) of Fig. 13A(b). These transformation coefficients F_mn ⁰(1) are not yet subjected to the selection of effective alternate current components by a code table, but have 19 alternate current components.

0075Then, the transformation coefficients F_mn(1) of Fig. 13A(c) are obtained as the result of the selection of the effective alternate current components by the code table. These transformation coefficients F_mn(1) include six effecitve alternate current components F₀₁, F₀₂, F₁₀, F₁₁, F₁₃ and F₁₄.

0076The restored image data F_ij(1) of Fig. 13A(d) are obtained by IDCT of the transformation coefficients F_mn(1).

0077In the example shown in Fig. 14, the threshold value σ₁ is set at 4.0 and the threshold values σ₂ and σ₃ are set to be at least 5.0. Since the standard deviation σ of the image data f_ij(0) of the original is 4.862 as shown in Fig. 13A(a), the image data f_ij(0) belong to the compression group g₂.

0078On the other hand, since the standard deviation σ of the image data f_ij(1) restored after the first data compression is 3.621, the image data f_ij(1) would belong to the compression group g₁ if the same were classified by the value of the standard deviation σ. However, as hereinabove described, the information indicating that the image data f_ij(0) of the original belonged to the compression group g₂, i,e., the group information N_g = 01 (binary notation) is recorded by even/odd of the data f₁₁(1) and f₃₃(1) (underlined in Fig. 13A(d)) of the image data f_ij(1). In more concrete terms, the data f₁₁(1) is adjusted to be odd and g₃₃(1) is adjusted to be even, as hereinabove described. Acccordingly, from the second data compression, the image data f_ij(1) is specified to belong to the compression group g₂.

0079Since the transformation coefficients F_mn(k) are accurately preserved as decimal values in the example, DCT is carried out reversibly. Therefore, the restored image data f_ij(2) and f_ij(3) obtained after the second cycle of the data compression and restoration are equal to the restored image data f_ij(1) after the first cycle of the data compression and restoration.

0080Fig. 14 illustrates transition of imsge quality of the restored images as SN ratios of the image data f_ij(k). SN ratios R_SN corresponding to the case shown in Figs. 13A and 13B (Example 1) are shown with white circles. The SN ratio R_SN is defined as follows: <maths id="math0004" num=""><img file="EP0349847A2_D0004.tif" /></maths> where f_ij represents a difference between image data of an original and that of a restored image. Symbols X and Y in the expression (11b) represent numbers of colums and rows of pixels existing in the whole image (see Fig.1).

0081As understood from Fig. 12, the SN ratio R_SN is not changed after the restored image data f_ij(1) in the example 1, that is, the deterioration of image quality is prevented.

0082Fig. 14 also shows a reference example, in which data compression and restoration are repeated without applying the present invention, for example. According to the reference example in which the compression groups is identified by the value of standard deviation σ, the image data is decided to belong to the compression group g₁ in the second data compression operation. Therefore, the image quality of a restored image after the second data compression is extremely deteriorated, and the value of its S_N ratio R_SN is also extremely reduced as shown in Fig. 14.

0083Examples 2 and 3 will be described next, in which the quantization of F_mn is performed. In the Example 1, the transformation coefficients F_mn are not subjected to the linear quantization. However, if the transformation coefficients F_mn are linearly quantized, storage capacity required for preserving the transformation coefficients F_mn can be extremely reduced and the compression rate can be further improved.

0084Figs. 15A and 15B are explanatory diagrams showing image data f_ij(k) and transformation coeffcients F_mn*(k) with lineary quantization on the transformation coefficients F_mn*. The subscripts * are attached to the transformation coefficients indicating that the same are linearly quantized.

0085Since the value of the quantum width α is "1", the term "linear quantization" herein used is the same as to round to the nearest whole number. The round-off operation is performed by counting fractions over 1/2 as one and disregarding the rest.

0086The SN ratio R_SN of the image data f_ij(k) in the Example 2 shown in Figs. 15A and 15B is shown with black squares in Fig. 14. Although the SN ratio R_SN of the Example 2 where the coefficients F_mn are linearly quantized decreases more quickly than that of the Example 1 where the linear quantization is not executed, the SN ratio R_SN of the Example 2 decreases more slowly than that of the reference example, and the image quality is also lowered more slowly. It is because the errors caused by the round-­off operation in the linear quantization of the transformation coefficients F_mn are accumulated, dissimilarly to the case of the Example 1.

0087In order to further suppress the progress of the deterioration of the image quality, the following method can be employed: Before performing IDCT on the transformation coefficients F_mn*(k), values of the transformation coefficients F_mn*(k), which are effective alternate current components, are modified as follows: <tables id="tabl0005" num="0005"><table frame="all"><tgroup cols="2" colsep="1" rowsep="0"><colspec colnum="1" colname="col1" colwidth="78.75mm" /><colspec colnum="2" colname="col2" colwidth="78.75mm" /><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left">when F_mn* > 0, FF_mn = F_mn* + 0.5</entry><entry namest="col2" nameend="col2" align="left">(12a)</entry></row><row><entry namest="col1" nameend="col1" /></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left">when F_mn* < 0, FF_mn = F_mn* - 0.5</entry><entry namest="col2" nameend="col2" align="left">(12b)</entry></row></tbody></tgroup></table></tables> The modified trensformation coefficients FF_mn are subjected to the IDCT. Application of the expression (12a) or (12b) is hereinafter referred to as "modification of the transformation coefficients F_mn*".

0088Figs. 16A and 16B show examples of the image data f_ij(k) and the transformation coefficients F_mn* in the Example 3 where the modification of the transformation coefficients is performed. The SN ratio R_SN of the Example 3 is shown with white squares in Fig. 14. The SN ratio R_SN of the Example 3 holds relatevely large, and the SN ratio R_SN is maintained constant after the second cycle of the data compression and restoration.

0089The expressions (12a) and (12b) are introduced on the basis of the following idea: When the image data f_ij(1) of Fig. 15A, which are restored with no modification of the transformation coefficients F_mn*, are compared with the image data f_ij(0), the restored image data f_ij(1) have smaller differences of data values between the pixels. That is, the image data after one cycle of the data compression and restoration are averaged in the pixel block to some extent. On the other hand, the absolute values |F_mn*| of the transformation coefficients F_mn* denote amplitudes of respective frequency components of the image data f_ij restored as a function of the transformation coefficients F_mn*, as shown in the expression (4). Therefore, when the absolute values |F_mn*| are increased, the amplitudes of the respective frequency components are increased so that the restored image data f_ij represent more undulation in the pixel bock. The expressions (12a) and (12b) are equivalent to the increase of the absolute values |F_mn*|. Therefore, when the transformation coefficients F_mn* are modified by the expressions (12a) and (12b) to restore the image data, an effect of averaging of the image data by the data compression is cancelled by the modification by these expressions.

0090The expressions (12a) and (12b) are generally expressed as follows: <tables id="tabl0006" num="0006"><table frame="all"><tgroup cols="2" colsep="1" rowsep="0"><colspec colnum="1" colname="col1" colwidth="78.75mm" /><colspec colnum="2" colname="col2" colwidth="78.75mm" /><tbody valign="top"><row><entry namest="col1" nameend="col1" align="left">when F_mn* > 0, FF_mn = F_mn* + α</entry><entry namest="col2" nameend="col2" align="left">(13a)</entry></row><row><entry namest="col1" nameend="col1" /></row><row rowsep="1"><entry namest="col1" nameend="col1" align="left">when F_mn* < 0, FF_mn = F_mn* - β</entry><entry namest="col2" nameend="col2" align="left">(13b)</entry></row></tbody></tgroup></table></tables> where both of α and β are positive values, and preferably within a range of 0.3 to 0.5.

0091When the transformation coefficients F_mn* are thus modified and restored, as the Example 3 shown in Fig. 14, the SN ratio R_SN is substantially equal to that of the Example 1 where the linear quantization is not performed and a large effect on suppressing progress of the deterioration of the image quality is attained. Further, since the transformation coefficients F_mn* are linearly quantized and preserved, the compressed image data D_t can be reduced to achieve high data compressibility.

0092Although the code tables are classified into a plurality of code table groups depending on the standard deviation of the image data in the above embodiment, such classification may be performed by means of variance or various statistical vlaue similar to the same, other than the standard deviation. That is, classification may be performed by means of some statistical value statistically representing distribution state of the image data.

0093Although the above embodiment has been described with reference to a method of compressing image data as to one image, the present invention may be applied to respective color separation images as to a color image. For example, respective color separation images of YMCK printed images, RGB signal images or Y_S/I_S/Q_S signal (chroma signal) images are compressed respectively. In this case, data compression suitable for each color separation image can be achieved by varying the number of threshold values of standard deviation and values thereof, the number of code tables in each code tables group, the numbers of effective alternate current components of the code tables, and the patterns of effective bit numbers of the code tables for each color separation image.

0094In this case, it is preferable to determine pixel positions (i, j) ((1, 1) and (3, 3) in the above embodiment) for recording the group information in the image data f_ij to be varied with the respective color separation images. Since values of the least significant digits of the image data f_ij in these pixel positions are adjusted in recording of the group information, such adjustment causes a small change of the image quality. Hence the change of the image quality is reduced as a whole if the pixel position whose least significant digit values are adjusted are not overlapped in overprinting respective color separation images. similarly, these pixel positions are preferably not adjacent to other pixel positions in the respective color separation images.

0095Although the group information is expressed by means of the least significant digits of the image data f_ij of the restored image in the above embodiment, the group information may be recorded as other data separately from the image data f_ij. In this case, the image data are defined as that including not only the data f_ij (hereinafter referred to as "image density data") representing density values of respective pixel positions but other data representing the group information.

0096The method of the data compression and restoration is not restricted to that of the above embodiment, but various methods such as that disclosed in Japanese Open Laying Gazette No.109086/1980 can be applied. The present invention is applicable to a method of obtaining a statistical value substantially representing standard deviation of image data within pixel blocks and selecting one of prescribed procedures of data compression with response to the statistical value. In the enbodiment, four procedures of the data compression operation using one of the code table groups different for the compression groups g₁ to g₄ correspond to the "plurality of prescribed procedures".

0097If the purality of prescribed procedures are not distinguished by means of the compression groups dissimilarly to the above enbodiment, some identification data for identifying the selected procedure is included in the restored image data. In this case, compressed image data are also preserved with such identification data. In the above enbodiment, the image data f₁₁ and f₃₃ in the restored image data f_ij and the code table number Max in the compressed image data D_t play the role of the identifying data, respectively.

0098According to the present invention, only a part of the transformation coefficients are employed to form the compressed image data, whereby the compression rate is increased without lowering the image quality.

0099Further, the image data are produced to include the identification data so that the same procedure as that employed in previous data compression can be applied to re­compression of the restored image data, whereby progress of the deterioration of the image quality is suppressed even if the data compession and restoration are repeated many times.

0100Because the identification data are expressed by means of the least significant digits of image data in prescribed pixel positions within a pixel block, there is no need to independently preserve the identification data. Consequently, the amount of restored image data is reduced while the deterioration of the image quality of the restored image is minimized.

0101Although the present invention has been described and illustrated in detail, it is clearly understood that the same is by way of illustration and example only and is not to be taken by way of limitation, the spirit and scope of the present invention being limited only by the terms of the appended claims.