JPH02295278A - Method and device for converting, encoding, and decoding gradation image signal - Google Patents

Abstract

PURPOSE:To encode a gradation image with high compressibility while holding its subjective picture quality sufficiently high by selecting a conversion coefficient in consideration of human visual characteristics. CONSTITUTION:The encoding and decoding device has a converter 1, a coefficient selector 2, a coefficient weighting parameter generator 5, a quantizer 3, an unequal-length encoder 4, a transmission line 6, an unequal-length decoder 7, a reverse quantizer 8, and a reverse converter 9. Then parameters which have larger values according to the importance to human sight are generated for respective conversion coefficients and when a coefficient is selected, a coefficient which has higher than certain importance to the human sight is selected. Thus, the conversion coefficient is selected in consideration of the human visual characteristics, so while the subjective picture quality is held sufficiently high, the gradation image can be decoded with the high compressibility.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a technology for encoding and decoding image signals.

(Prior Art) One of the typical conventional image coding methods is transform coding.

Regarding transform encoding, see William Caperat (W
illiamK. Pratt) rDigital Image Processing
sing, Wiley-Interscience
It is explained in detail in Jpp 667-699 et seq. "Reference 1").Transform encoding will be explained below based on Reference 1.

FIG. 14 shows a basic block diagram of the transform encoding method.

Note that in the following, the position coordinates of each pixel constituting the image signal are indicated by θ, k). Let j be a coordinate indicating the position in the vertical direction, and k be a coordinate indicating the position in the horizontal direction. The i-th original image signal to be transmitted is Fi(i, k). Fi
, k) θ = 1 to M, k = 1 to L) each corresponds to one image (referred to as a [frame]). M is the number of pixels in the vertical direction, and L is the number of pixels in the horizontal direction. Hereinafter, i will be referred to as an image number.

The original image signal Fi (j, k) is divided into blocks of NxN pixels, and a transformer performs linear transformation such that j = lk-1 to convert the NxN transformation coefficient fH (u, v).
is obtained for each block. In (1), Ac(j, u
) is a column-direction transformation matrix that transforms the original image signal along the vertical direction, and AR(k,v) is a row-direction transformation matrix that transforms the original image signal along the horizontal direction. In the coefficient selector 61, only a part of the conversion coefficient "Mf; (u, v) is selected based on the coefficient allowance ai (u, v), and the selected conversion coefficient fTi (u, v) is obtained. The value of the coefficient is set to zero.For example, select the coefficient where a4(u,v)>T, and discard the coefficient where cyi(u,v)<T.Here, T
is a certain constant.The coefficient loss measuring device 63 calculates the loss OH(u,v) of the coefficient fi(u,v) per image.The quantizer 62 calculates TH(u,v). .Quantizer 62
quantizes fT1(u,v), obtains a transform coefficient quantization level number fq·NH(u,v), and sends it to the transmission line 64.

On the receiving side, f obtained from the transmission path is processed by inverse quantization 65.
TNi (u, ■) is deconcentrated and dequantized transform coefficient rT
Obtain i(u,v). In the inverse transformer 66, f;(u
, v) is subjected to linear transformation of 1=l k=1 for each block to obtain a decoded signal for one block consisting of N×N pixels. In (2), Be(j,
u) is the matrix [AC(j, u month's inverse matrix is Nori, u) elements, BR(k, V) is the matrix [AR(j, u month's inverse matrix is (Nori, u) elements,
k, v) elements, all coefficients fi
Transmission by selecting some coefficients fTi (u, v) from (u, v), and converting the selected coefficients by quantization into a signal rTNi (u+■) with a smaller number of levels than the original image signal. They are reducing the amount of information.

(Problem to be solved by the invention) In order to increase the information compression rate in transform encoding, it is possible to design an appropriate coefficient selector and cut off transform coefficients fi (u, v) with statistically small amplitudes. good. For this reason, conventionally,
A method has been proposed in which the variance of the amplitude of fi (u, v) is statistically determined over a large number of images, and the transformation coefficient fi (u, v) with a small variance is discarded (Reference 1, p. 673). Furthermore, a method has been proposed in which the margin is determined for each image as shown in FIG. 14 and used as a criterion for coefficient selection. Furthermore, in order to reduce the amount of generated information, the following methods have also been proposed. In other words, in order to improve coding efficiency, each block is divided into four classes depending on the magnitude of AC energy, and then the variance of the amplitude of the transform coefficient Fi (u, v) is calculated for each class and the one with the smallest variance within each class is calculated. Conversion coefficient fi(
A method of truncating u, v) has also been proposed (Wein-Hsiung Chen et al., “Adaptive Coding of Monochrome and Color Images,” Adaptive Coding of M
onochrome amd Color Image
s, IEEE Transacitoons on Co.
mmunication magazine J November 1977 issue, p.
p1285-1292, hereinafter [Reference 2]). The coefficient truncation method in Reference 1 is exactly the same for all Bronok {,
Go alone. However, when focusing on a certain block, the amplitude of the transform coefficient fi(u,v) is not necessarily small just because the variance determined over the entire image is small; on the contrary, it may be large. The variance is just a statistic obtained over the entire image, and the magnitude of the transformation coefficient Fi(u,v) of a certain block is often much larger than its variance.

As described above, in the conventional transform encoding method according to Document 1, coefficients are truncated in exactly the same way for all blocks, and therefore, there is a drawback that coefficients cannot be truncated in a way that matches the local characteristics of an image. To improve this point,
It may be determined which coefficients to truncate for each block. However, if this was done, it would be necessary to encode which coefficients were truncated for each block and send it from the transmitting side to the receiving side, which would increase the amount of transmitted code. Literature 2
In this case, each block is divided into four classes according to the AC energy, and the coefficient to be rounded down is the same in each class. Therefore, it is not necessary to encode which coefficients are truncated for each block, and it is sufficient to encode each class, so that the amount of transmitted codes does not become so large. However, the basic idea of block classification according to AC energy and truncation of coefficients in Reference 2 is based on the rate distortion theory (rate disto theory).
tion theory). To apply late distortion theory to image encoding, encoding distortion must be quantitatively defined. In order to facilitate theoretical handling, the mean square error is often defined as encoding distortion, but in literature 2, it is also defined as the rate distortion theoretical function (rate distortion function).
tion), it is assumed that the mean squared error is taken as the encoding distortion.

In another method, A1iHabibi
rSurvey of Adaptive Image Coding Techniques, Survey of Ada
ptive Image Cladding Technique
iques, IEEE Trans. on Comm
．． Magazine J November 1977 issue, pp1275-1284
As introduced on page 1279 of Reference F31), in each block, the area of conversion coefficients is divided into triangular areas by straight lines parallel to the diagonals, and a triangle of an appropriate size is selected. Send only the inner coefficients. Then, this triangle is divided into 1a bands in the area parallel to the diagonal, and the variance of the coefficients within each band is assumed to be equal.From the viewpoint of minimizing error energy, the bits for the coefficients for each band are Assignments were being made.

As explained above, the coefficient selection method in Reference 1, Reference 2, and Reference 3 evaluates the error that occurs when encoding an image using the mean square error, and minimizes it under a certain amount of generated information. This is what I am trying to do. However, the mean squared error is not necessarily appropriate as encoding distortion. The most important thing in image encoding is to encode the image with as little amount of information as possible, and to make sure that when humans view the encoded image, distortion is as invisible as possible.

In other words, it is to improve subjective image quality. In this way, when thinking about humans and taking human visual characteristics into account to improve subjective image quality, the mean squared error is often not appropriate as a measure for defining encoding distortion. Therefore, all of the methods in Reference 1, Reference 2, and Reference 3 use the assumption that the mean squared error is small, and it is difficult to say that they sufficiently take human visual characteristics into account, and the subjective image quality is insufficient. There was a drawback.

(Means for Solving the Problems) In the present invention, on the encoding side, an input image signal is divided into a plurality of blocks, and the input image signal is linearly transformed for each block to obtain transformation coefficients, and human visual characteristics The selected transform coefficients are selected based on coefficient weighting parameters that consider The coefficients are unequal length decoded, the unequal decoded quantized coefficients are dequantized to obtain decoded transform coefficients, and the composite transform coefficients are linearly inverse transformed to obtain a decoded image signal.

In addition, in the present invention, on the encoding side, the input image signal is divided into a plurality of blocks, and the input image signal is linearly transformed for each block to obtain transformation coefficients, and coefficient weighting parameters that take into consideration human visual characteristics are provided. means for selecting transform coefficients based on; means for quantizing the selected transform coefficients; and means for unequal length encoding the quantized transform coefficients;
Decoding {j'ill includes means for decoding quantized coefficients encoded with unequal lengths, and means for dequantizing quantized coefficients encoded with unequal lengths to obtain decoded transform coefficients. , the grayscale image is encoded and decoded by means of performing linear inverse transform on the decoded transform coefficients to obtain a decoded image signal.

(Function) In the present invention, conversion coefficients are selected taking into consideration human visual characteristics, and while maintaining sufficiently high subjective image quality,
It is possible to compose grayscale images with a high compression rate. The principle is that for each conversion coefficient, the more important it is to human vision, the larger the parameter is generated, and when selecting coefficients, the more important it is to human vision, the more important it is to human vision. It is in.

In this way, excellent encoding characteristics can be obtained from the viewpoint of improving subjective image quality with a constant amount of information.

(Example) Hereinafter, an example of the present invention will be described in detail with reference to FIGS. 1 to 13.

FIG. 1 shows an example of a block diagram of an encoding device and a decoding device according to the present invention. The converter 1 inputs an original image signal Fiθ,k) from a terminal 100, linearly transforms it using NXN pixels as one block, and obtains transformation coefficients Fi(u,v).

In the coefficient weighting parameter generator 5, when the linear transformation is a two-dimensional discrete cosine transformation, the generated signal W(u, v
) is Staffen Eri
cson) ``Frequency Weighted Inch-Frame Hybrid Coding, (Frequency
Weighted Interframe Hybrid
Coding) J Report No. TRI
TA-TTT-8401, Telecommunica
tion Theory) Electrical Eng.
ineering, Royal Institute
of Technology, Stockholm
J (January 1984, hereinafter referred to as [Reference 4]), it can be decided to select coefficients that take human visual characteristics into account, and W (U, V) has a larger difference as the frequency component increases. Figure 2 shows an example of W(u,v) when the block size is 32 x 32 pixels (N=32). In FIG. 2, only values for low frequency components that are actually important are shown. The coefficient selector 3 sets a certain threshold value WT, and W(u,v)>W
Select only the coefficient T. In the example in Figure 2, WT=0
．． When set to 79, the coefficients selected are those inside the dotted line. This is the area shown in Figure 3 (■). When coefficients are selected using the method according to document 3, the area approximately the same size as ■ is inside the solid line in FIG. This is the area indicated by ■ in FIG. ■ and ■
A comparison is shown in Figure 5. In other words, ■
is an area that is commonly selected in ■ and ■, ■ is an area that is selected in ■ but not in ■, and ■ and ■ is an area that is selected in ■ but not in ■. Comparing ■ and ■, the difference is that ■ has a value of W (u, v) of 0.79 ~
A coefficient with a relatively large value of 0.82 is selected, whereas a coefficient with a relatively small value of W (u, v) of 0.61 to 0.78 is selected in ■. In other words, ■ selects visually more important coefficients, whereas
(2) visually selects less important coefficients and discards more important coefficients. Therefore, if you compare the same amount of information, it will be understood that the image quality of ■ is better than ■ when human visual characteristics are taken into account.

Although Reference 1 and Reference 2 attempt to select coefficients with larger variance values, it is necessary to note that coefficients with large variance values are not necessarily visually important coefficients. In References 1 and 2, as a result of measuring the coefficient variance, the value is small, so visually important coefficients are often discarded. In such a case, W(u, v) is calculated based on the value of W(u, v) as in the present invention.
v) Better image quality can be obtained by selecting a coefficient such that <WT.

Here, the differences between the present invention and Document 4 will be explained. The basic difference is that in the present invention, coefficients are selected and then quantized, whereas in Document 4, coefficient selection processing is not performed at all, and only quantization is performed. Therefore, in the method of Document 4, even visually unimportant coefficients are often encoded with unequal length and sent, and the degree of improvement in image quality is often small despite the generation of a larger amount of information. For example, assume that the X marks in FIG. 6 are significant coefficients when encoding using the method of Document 4. Here, a significant coefficient is a coefficient that does not become zero even after quantization. In the method of Reference 4, a minimum rectangle surrounding all significant coefficients is found and the coefficients inside the rectangle are encoded with unequal length. At that time, coefficients located inside the rectangle are subjected to unequal length encoding, regardless of whether they are significant coefficients or not. In the case of Figure 6, the thickness of this rectangle is k in the U direction.
c''=15, lc'=12 in the v direction. Therefore, in the case of FIG. 6, kc'lc' = 1 in the method of Document 4.
5×12=180 coefficients are to be encoded. Among them, 170 quantized coefficients are zero, and 10 are significant coefficients. Suppose that one quantized coefficient of zero can be encoded with one bit, and one significant coefficient can be encoded with an average of three bits. In this case, the quantization coefficient in FIG. 6 is approximately 170 X 1 + 10 X 3 using the method of Document 4.
= Can be encoded with 200 bits. In contrast, in the present invention, visually important coefficients are selected in advance. For example, only the coefficients located inside the area indicated by ■ in FIG. 3, that is, the area indicated by the solid line in FIG. 6, are selected and made to be encoding targets. Therefore, in the case of Figure 6, (u, v) = (1
The significant coefficients present in 5.12) are truncated as they are visually insignificant. Find the smallest rectangle that encloses the other nine significant coefficients, the size of which is k in the U direction. =12,1
c=9. Therefore, in the case of FIG. 6, ko1 in the present invention. =12×92108 coefficients are to be encoded. Of these, 99 quantized coefficients are zero, and 9 are significant coefficients. Assume that a quantized coefficient of zero can be encoded with one bit, and one significant coefficient can be encoded with an average of three bits, as described above. In this case, according to the present invention, the quantized coefficients shown in FIG. 6 can be encoded with approximately 108 + 9 x 3 = 135 bits. Since the method of Reference 4 required about 200 bits, the present invention was able to reduce the amount of code by about 33%. As described above, in the present invention, compared to the method of Document 4, the amount of information can be significantly reduced in many cases with almost no deterioration in image quality.

Although the basic configuration of the present invention is shown in FIG. 1, it is possible to make various improvements to it to further improve encoding efficiency. An example is shown in FIG. In comparison with FIG. 1, in FIG. 7, on the encoding side, the coefficient selector 11
A coefficient weighter 12 is inserted between the inverse transformer 20 and the inverse quantizer 18, and on the decoding side, a coefficient inverse weighter 19 is inserted between the inverse transformer 20 and the inverse quantizer 8. There is. The idea shown in FIG. 7 is to quantize visually important coefficients more finely and to quantize visually unimportant coefficients more roughly, and this method is introduced in Reference 4. That is, after weighting the transform coefficient fgi (u, v) selected by the signal W (u, v) generated by the coefficient weighting parameter generator by W (u, v), the quantizer 13 uniformly quantizes the selected transform coefficient fgi (u, v) by W (u, v). If so, then W(u
, v), the large fTi(u, ■) is roughly quantized,
The pair fTi(u,v) of W(u,v) is equivalent to being finely quantized.

Large coefficients of W(u,v) are visually important coefficients, and small coefficients are less important coefficients. On the decoding side, the coefficient weighter 19 divides the output of the inverse quantizer 18 by W(u,v) to obtain the decoded transform coefficient f
Obtain TH(u,v).

Furthermore, the encoding/decoding apparatus shown in FIG. 8 can be obtained by further developing from the block diagram of FIG. 7. A block diagram of the image encoder 22 and image composite unit 30 in FIG. 8 is shown in FIG. A block diagram of the complexity detector 23 in FIG. 8 is shown in FIG. 9. FIG. 8 shows a block diagram of the contrast detector 24 in FIG. In FIG. 8, on the transmitting side, the i-th original image signal F is sent from the terminal 100.
4(j,k) is input and added to the contrast detector 24, complexity detector 23 and image encoder 22. The contrast detector 24 receives the i-th input image signal Fi(j
, k), the contrast Ci (j, k) is determined and added to the image feature parameter encoder 26 and the image decoding encoding parameter setter 25. The value of Ciθ,k) is, for example, 0 when the image is high contrast, 1 when the image is medium contrast, and 2 when the image is low contrast.

In the complexity detector 23, the i-th input image signal Fi
A complexity Gi(j, k) is determined for θ, k) and added to the image feature parameter encoder 26. The value of Gi (i, k) is, for example, 0 for a block in which the image signal "i (J, k) has a high degree of complexity, and 1 for a block in which the image signal Fi (j, k) has a low degree of complexity." The feature parameter encoder 26 is C
i(j,k), Giθ,k) and area signal BiQ,k
) is unequal-length encoded using Huffman encoding or run-length encoding, and is added to the multiplexer 27.
i(i.k) will be explained in detail later. Ci(j
, k), Giθ,k), and Biθ,k) take constant values within each block. The image encoding parameter setter 25 is
Parameters indicating the quantization step size and coefficient truncation criteria are determined according to the values of Ci(], k), Giθ, k), and Biθ, k) and are added to the image encoding 22. The image encoder 22 encodes the i-th image signal Fiθ,k) according to predetermined encoding parameters and sends it to the multiplexer 1.
Add to 8. The multiplexer 18 first selects the output of the image feature parameter encoder 26 and then selects the output of the image feature parameter encoder 2.
2 is selected and output to the transmission line 28.

On the receiving side, the encoded image contrast signal, image complexity signal, area signal, and encoded image signal are received from the transmission line 28, and the encoded image signal is sent to the day multiplexer 2.
9, the coded image contrast signal, the coded additive complexity signal, and the coded image signal are separated from the coded image contrast signal, the coded additive complexity signal, and the coded image signal. The region signal is applied to an image feature parameter decoder 32, and the encoded image signal is applied to an image decoder 30. The image feature parameter decoder 32 corresponds to the image feature parameter encoder 26, and decodes the image contrast signal C: (j, k) and image complexity signal Gi for each block of the i-th image.
(j, k) and area signal Bi(j, k) are obtained and added to the image decoding parameter setter 21. The image decoding parameter setter 31 sets Ciθ,k for the image decoder 20.
), Gi (j, k), and Bi (i, k), the image decoder 30
Add to. The image decoder 30 performs an operation corresponding to the image encoder 16 according to the determined encoding parameters,
The encoded image signal is decoded to obtain the i-th decoded image I'i (j, k), which is output to the output terminal 101.

Next, an example of the complexity detector 23 is shown in the ninth example. The i-th original image signal Fi (i, k) is input from the terminal 100, and the average value detector 36 makes one block of N×N pixels and detects the average value AVE of Fi (j, k) within the block. 2
The digitization circuit 33 binarizes Fi(j,k) using AVE as a threshold value to obtain a binary signal LiO,k), and the change point coefficient unit 34
Add to. The change point coefficient unit 34 receives a binary signal Li(j,k)
The number of horizontal change points NHiθ,k) in each block of
and the number of vertical change points NyH(i,k).

NHi(j,k) takes a constant value for each block.

For example, the complexity measuring device 35 calculates Nni (j, k) and N
When the larger value of vi (i, k) is greater than a certain threshold, it is assumed that the degree of complexity is high, and Gi (3, k) = o is output to the terminal 107, and conversely, when it is smaller than a certain threshold, the degree of complexity is high. If the temperature is low, Gt(j,k)=1 is output to the terminal 107. Since complexity detection is performed for each block of N×N pixels, Gi(i,k) takes a constant value for each block.

Next, an example of the contrast detector 24 is shown in FIG. The i-th original image signal Fiθ,k) is inputted from the terminal 100, and the highest level detector 38 makes one block of N×N pixels and detects the highest level MAX within the block. On the other hand, the lowest level detector 39 detects the lowest level M in the block.
Detect IN. A subtracter 40 calculates MAX - MIN for each block and a difference histogram measuring device adds it to 43. In the differential histogram measuring device 43, MAX-MIN
The frequency distribution of is calculated for each image, and a result as shown in FIG. 12, for example, is obtained. As can be seen from an image such as a seal stamp or a fingerprint imprint, there is little level variation in the background portion of the image, and therefore MAX-MIN is considered to be small. In the non-background part, the opposite is true, the level fluctuation is large, and therefore MAX-MIN is considered to be large. Therefore, when a frequency distribution of MAX-MIN as shown in Fig. 12 is obtained, the area where MAX-MIN is smaller than the point dB that gives the valley of the frequency distribution is the non-background area, and the area where MAX-MIN is smaller than that is the non-background area. A region with a small MIN is defined as a background region. In FIG. 10, the valley detector 44 receives the difference classification distribution from the difference histogram measuring device 43, calculates the value dB of MAX-MIN that gives the valley of the difference classification distribution, and adds it to the background detector 45. The background detector 45 determines whether the level difference MAX-MIN of each block is smaller than aB, and when it is small, Bi (j, k) = 1, that is, background, and when it is large, Bi θ, k) = 0, that is, non-background. shall be. Hereinafter, Biθ,k) will be referred to as a region signal. 1st
In the background separation shown in FIG. 0, the background and non-background are separated for each block of N×N pixels, so Bi(j,k) takes a constant value for each block. However, the background area detection block size does not necessarily have to be NXN. A control circuit 46, which may be of other sizes, such as N/2 x N/2, provides synchronization and clock signals to and controls each section.

The contrast measuring device 42 receives the brightness frequency distribution of the original image signal for each NXN pixel from the histogram measuring device 4l, for example, and sets the frequency value of the brightness frequency distribution to a predetermined value n.
When the lowest luminance dB1 is greater than 1, and the highest luminance at which the frequency value is smaller than a predetermined value n1 is dB2, W" dB
When 2-dB1 is greater than a predetermined value W1, Ci
(j, k) = 0 high contrast; when W is smaller than a predetermined value W2, Ci θ, k): 2 contrast is low; otherwise, CiO, k) = x medium contrast. However, W1 > W2. FIG. 13 illustrates an example of the luminance frequency distribution and nl, W, dB1, and dB2. W=d
Let B2-dB1. However, there are many other ways to define W. For example, W=(dB2-dB1)/
It may be dB2+dB1.

Ciθ,k) takes a constant value for one block (NXN pixels).

FIG. 11 shows an example of the image encoder 22, image encoding parameter setter 25, image decoder 30, and image decoding parameter setter 31 in FIG. 8. In FIG. 11, on the transmitting side, the converter 47 receives the i-th image signal F from the terminal 100.
j (j, k) is input, divided into blocks of N×N pixels, linear transformation is performed according to equation (1), transformation coefficients fi (u, v) are obtained, and the coefficients are added to the selector 48.

The coefficient selector 48 includes a coefficient weighting parameter generator 52
A parameter W(u,v) indicating a criterion for coefficient selection is received, a coefficient is selected based on W(u,v), and the selected coefficient FTi(u,v) is added to the coefficient weighter 49.

The coefficient weighter 49 performs coefficient quantization in consideration of human visual characteristics according to the method proposed in Reference 4, so that the transform coefficients are weighted lower in advance as they become higher frequency components, and W(u,v)FTH(u , v) and add the result to the quantizer 50. The quantizer 50 sets the quantization step size Δ:(
j, k) and quantization step size Δi(j,
The coefficient W(u+
v) Quantize FTH (u, v) and set the quantization level number (
W(u,v)FTi(u.V))N is obtained and added to the unequal length encoder 51. The quantizer 50 also obtains a parameter kc, 1o indicating the existence of an effective coefficient, and passes this to the unequal length encoder 3.
Add to 2. The unequal length encoder 32 encodes coefficients W(u,v)FTH(u,v) quantized using Huffman encoding or the like.
v) compression encode N and kc, lc, compressing code,
is obtained and applied to terminal 104. The terminal 104 is connected to the multiplexer 27 in FIG. 8. On the receiving side in FIG. 11, the unequal length decoder 58 performs an operation corresponding to the unequal length encoder 51 and decodes the compressed code. Quantized coefficients (W(u,v)FTi(u,v))N and k.

, 1c is obtained and added to the inverse quantizer 57. Inverse quantizer 57
receives the quantization step size Δi(j,k) from the image decoding parameter measuring device 59, dequantizes the quantized coefficients to obtain W(u,v)fi(u,v), and performs coefficient inverse weighting. Add to container 56. The coefficient inverse weighter 56 operates inversely to the coefficient weighter 49, and W(u,v)f'i
(u,v) is divided by W(u,v) to obtain ri(u,v), which is added to the inverse transformer 55. W(u,v) is applied to the coefficient weighting parameter generator 55.

W(u,v) is obtained from the coefficient weighting parameter generator 54.

Next, a method for determining Δiθ,k) and Bi(i,k) will be explained. In FIG. 11, a method for determining Δi(i,k) will first be described.

In FIG. 11, Δiθ,k) is Bi(i,k)=0
That is, in the non-background area, it can be determined as shown in Table 1, for example.

Table 1 In the table, for example, Δ3 × Δ2 × Δ1 (6) Δ3I<Δ2I<ΔII (7) Δ1 × Δr (8
) Δ2 × Δ2I (9) Δ3 × Δ3' (10). The idea is to make the quantization step size Δi(j,k) smaller when the contrast is low and the complexity is high, and to increase the quantization step size Δiθ,k) when the contrast is high and the complexity is low. be. This idea focuses on the fact that the lower the contrast or the higher the complexity, the more noticeable coding distortions such as block distortion are to the human eye. However, Ci(
], k), Bi (j, k), the contrast is extremely low in the background region, that is, when Bi (j, k) = 1, so Δi (j, k) is calculated by all Δ in Table 1.
iθ,k) is set to be smaller than Δ0. In this way, encoding distortion such as block distortion becomes less noticeable, and images can be encoded with good image quality. Furthermore, at this time, the luminance may be averaged over all the background regions where Bi(i.k)=1 to obtain a background average level, and this level may be encoded.

There are 3 classifications regarding contrast and 2 classifications regarding complexity.
Although we have explained the case of classification, there are many ways to think about the number of classifications. For example, the number of classifications may be increased, such as 4 classifications for contrast and 4 classifications for complexity. Alternatively, the number of classifications may be reduced, such as two classifications for contrast and two classifications for complexity.

(Effects of the Invention) According to the present invention, conversion coefficients are selected taking human visual characteristics into consideration, and a grayscale image can be encoded at a high compression rate while maintaining sufficiently high subjective image quality.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an example of a block diagram of an encoding/decoding apparatus according to the present invention, FIG. 2 is a diagram showing an example of coefficient weighting parameters, and FIG. 3 is a diagram showing an example of a coefficient selection area according to the present invention. , FIG. 4 is a diagram showing an example of the coefficient selection area according to the conventional method, FIG. 5 is a diagram showing a comparison between the coefficient selection area according to the present invention and the coefficient selection area according to the conventional method, and FIG. FIG. 7 is a diagram showing an example of a block diagram of the encoding device of the present invention, and FIG. 8 is a diagram showing a comparison of coefficients to be encoded.
The figure shows an example of the encoding/decoding device of the present invention, FIG. 9 shows an example of a complexity detector, FIG. 10 shows an example of a contrast detector, and FIG. 11 shows an image encoder. , an image encoding parameter setting device, an image decoder, and a diagram showing an example using the image decoding parameter setting device, FIG. 12 is a diagram for explaining the background separation method, and FIG. 13 is a diagram for explaining the contrast detection method. and FIG. 14 are block diagrams of a conventional encoding/decoding device using transformation. In the figure, 1, 10, 47, 60... converter, 2, 11, 48,
61... Coefficient selector, 5, 15, 21, 52.54.
... Coefficient weighting parameter generator, 3, 13, 50,
62... Quantizer, 4, 14. 51... Unequal length encoder, 6, 16, 28. 64... Transmission line, 7, 17.
58... Unequal length decoder, 8, 18, 57, 65...
・Inverse quantizer, 9, 20, 55.60--inverse transformer,
12.49... Coefficient weighter, 19.56... Coefficient inverse weighter, 22... Image encoder, 23... Complexity detector, 24... Contrast detector, 25...
- Image encoding parameter setter, 26... Image feature parameter decoder, 27... Multiplexer, 29...
! Day multiplexer, 30... Image decoder, 31...
...Image decoding parameter setting device, 33. ... Binarization circuit, 34... Change point counter, 36... Average value detector, 38... Maximum level detector, 39... Minimum level detector, 40... Subtractor, 42... Contrast measuring device, 43... Differential contrast measuring device, 44...
Valley detector, 45... Background detector, 466... Control circuit, 63... Coefficient breakdown measuring device.

Claims (2)

[Claims] (1) On the encoding side, the input image signal is divided into multiple blocks, the input image signal is linearly transformed for each block to obtain transformation coefficients, and the transformation coefficients are calculated based on coefficient weighting parameters that take into account human visual characteristics. Select, quantize the selected coefficients, perform unequal length encoding on the quantized coefficients, and on the decoding side perform unequal length decoding and unequal length decoding on the unequal length encoded quantized coefficients. A method for encoding and decoding grayscale image signals, in which dequantized quantized coefficients are dequantized to obtain decoded transform coefficients, and the decoded transform coefficients are linearly inversely transformed to obtain a decoded image signal. (2) On the encoding side, the input image signal is divided into multiple blocks, and the input image signal is linearly transformed for each block to obtain transformation coefficients, and the transformation is performed based on coefficient weighting parameters that take into account human visual characteristics. means for selecting coefficients;
means for quantizing the selected transform coefficients; means for unequal-length encoding the quantized coefficients; and means for unequal-length decoding the quantized coefficients encoded on the decoding side. , a gradation system characterized by having means for dequantizing quantized coefficients that have been decoded with unequal length to obtain decoded transform coefficients, and means for linearly inversely transforming the decoded transform coefficients to obtain a decoded image signal. Image signal encoding/decoding device.