JPH0352712B2 - - Google Patents

Description

[Detailed description of the invention]

This invention uses quantization to stabilize fluctuations in compression ratio that occur when variable length codes are used for highly efficient encoding in encoding for the purpose of narrowband transmission or storage of images. Regarding the quantization method. To save transmission capacity by removing redundant signals due to the statistical properties of image signals;
In other words, it is well known that band compression is possible. If image signals can be biased by performing some kind of transformation by taking advantage of the strong spatial correlation between adjacent pixels, highly efficient encoding becomes possible. Here,
As transformation methods for imparting bias to images, linear predictive transformation methods and orthogonal transformation methods are widely known. The linear predictive conversion method is called DPCM
(Differential Pulse Code Modulation) uses the fact that the prediction error has statistical bias. This takes advantage of the fact that the spectrum of each frequency component that is the output is biased. On the other hand, in order to perform highly efficient encoding, distortion is generally applied to these transformed components within a visually permissible range, but the probability distribution of these transformed components is also biased. In this case, it is known that it is effective to provide a variable length code by taking advantage of the fact that entropy can be reduced. However, this statistical bias varies considerably depending on the image, with smooth images having a large bias and complex images having a large variation. If a variable length code is given to , there is a problem in that fluctuations occur on the time axis. Conventionally, a large buffer memory was required to correct this variation in the time axis. SUMMARY OF THE INVENTION In view of the above-mentioned points, it is an object of the present invention to provide a quantization method that stabilizes fluctuations in compression ratio by controlling quantization characteristics and reduces the capacity of a buffer memory. This invention will be explained below. This invention relates to a quantization method for image signals, in which an input image is divided into pixel groups m (1, 2) x n (≧2).
The pixel group in each block is converted into a frequency domain sequence by Hadamard transform, and then the currently selected quantization characteristic among the plurality of quantization characteristics is used. quantize, linearly predict at least the DC component of this sequence quantized value and convert it into a quantized value of a prediction error component, and form a plurality of combinations by comparing the sequence quantized values with each other. Corresponding to each combination, each sequence quantization value is variable-length encoded and stored in the buffer memory, read out from the memory at a constant bit rate and output, and the quantization characteristic for the next block line is selected according to the amount of code in the memory. It was designed to do so. First, an image signal encoding method to which the present invention method can be applied will be explained. Figure 1 shows an image divided into 2×2 blocks Bij (i=1,
2, ..., j = 1, 2, ...), pixel 1 is divided into scanning lines h, h+
1... are scanned sequentially. In addition,
Let i, (i+1), . . . be block lines. If the pixel signals (density signals) for pixel 1 of block Bij are x ₁ , x ₂ , x ₃ , x ₄ as shown in the figure, then
The normalized Hadamard transform can be expressed as the following equation (1) using a quadratic or quartic Hadamard matrix. 1/21 1 1-1x ₁ x ₂ x ₃ x ₄ 1 1 1-1 = 1/21 1 1 1 1-1 1-1 1 1-1-1 1-1-1 1x ₁ x ₂ x ₃ x ₄ = H ₀ H ₁ H ₂ H ₃ ...(1) Here, the sequence H ₀ , H ₁ , H ₂ , H ₃
correspond to the fourth-order frequency components transformed into the frequency domain by Hadamard transform, H ₀ is the DC component, H ₁ is the vertical component, H ₂ is the horizontal component,
_H3 indicates the slope component, respectively. By the way, in general images, the correlation between adjacent pixels is strong, so the energy of low-order components is high,
The energy of higher order components is observed to be low. Each component is then distributed with the number of bits of the quantization level according to the magnitude of its energy, but since the energy of the DC component is particularly large compared to other components (vertical, horizontal, and inclined), It is necessary to realize highly efficient encoding by adopting a conversion method that lowers the energy of the DC component. As such a conversion method, a method of linearly predicting each sequence obtained by Hadamard transform is known. Next, this linear prediction method will be explained. In the linear prediction method, the block size is 1×n or 2
This is possible when the block size is 2 x n as shown in Figure 2. First, Hadamard transform is applied to each block in the horizontal direction, and then the Hadamard transform is applied to each block in the vertical direction. Consider the case where linear prediction is performed at each time. Then, each sequence obtained by Hadamard transformation is designated as Hl ₀ and Hl ₁ and arranged as shown in FIG. 2 for convenience. However, l=0,
1,..., (n-1). And the sequence in block Bi _-1 ,j
From Hl ₀ ′, the sequence Hl ₀ in block Bij is
Prediction is made according to the following equation (2). Hl^ ₀ = Hl ₀ ′−Hl ₁ ′−Hl ₁ ...(2) The prediction error component ΔHl ₀ at this time is ΔHl ₀ = Hl ₀ −Hl^ ₀ = Hl ₀ −Hl ₀ ′+Hl ₁ ′+Hl ₁ ... ...(3), then this prediction error component ΔHl ₀ is the one that minimizes the distance between each pixel between blocks. Note that the sequence Hl ₁ is transmitted as is without any prediction. Here, in order to show the effect of linearly predicting each sequence obtained by Hadamard transform, a comparison will be made with the total energy of the image to be transmitted. If the variance of the image x according to the memoryless Gaussian distribution is σ ² x, the variance of each sequence y is σ ² y, the block size is N, and the variance when linear prediction is used for some sequences is σ ² Δy, then Rate Distortion
Based on Theory, the absolute value of the transmitted energy is given as the average value Pa as follows. Pa=10/N・_N-1 〓 ^y=0 log ₁₀ σ ² y ……(4) or Pa=10/N・_N-1 〓 〓 ^y=0 log ₁₀ σ ² Δy ……(5) And, The relative value Pr of the transmission energy is Pr=10log ₁₀ σ ² −Pa (6). From calculations using equations (4) to (6) above, results as shown in FIG. 3 were obtained for a standard image. In Fig. 3, the characteristic shows the change in the absolute power to be transmitted corresponding to the block size of the one-dimensional Hadamard transform, and the characteristic is 2
shows the change in the absolute power to be transmitted corresponding to the block size of the dimensional Hadamard transform. There is. Note that the total energy of the original image is 35.2 dB (rms). As is clear from this, the transmission energy of both the characteristics and the transmission energy decreases as the block size increases, whereas the transmission energy of the characteristics and the transmission energy remains constant (approximately 15 dB) almost regardless of the block size. . This means that
From the viewpoint of transmission energy, it is shown that it is efficient to select a block size of 4×1 or 2×2 and perform linear prediction. It has been found that it is efficient to perform Hadamard transform in the horizontal direction and perform linear prediction in the vertical direction. In this case, a buffer memory is required to store the entire sequence in the previous block line, but in order to save the capacity of this buffer memory, Hadamard transform is performed in the horizontal direction using a block size of 2 × 2, and similarly It may be considered to perform linear prediction in the horizontal direction.
In this case, the buffer memory only needs to have a capacity to store the fourth-order sequence, and the buffer memory shown in Figure 1 is
Each sequence H ₀ ′ of the previous block Bi, j _-1 in the same block line, such as Bi, j _-1 and Bij,
Sequence of current block Bij from H ₁ ′, H ₂ ′, H ₃ ′
H ₀ and H ₁ are predicted, and the sequence to be transmitted is DC prediction error component ΔH ₀ , vertical prediction error component ΔH ₁ , ΔH ₀ =H ₀ −H ₀ ^=H ₀ −H ₀ ′+H ₂ ′+H ₂ ...(7) ΔH ₁ =H ₁ −H ₁ ^=H ₁ −H ₁ ′+H ₃ ′+H ₃ ...(8), horizontal component H ₂ , and tilt component H ₃ . Here, the vertical component H ₁ is already composed of the sum and difference within the block Bij, and the transmitted energy is also the DC component H ₀
Since it is very small compared to , linear prediction has little effect. In other words, there is almost no difference in the magnitude of transmission energy between the vertical component H ₁ and its prediction error component ΔH ₁ , and linear prediction is performed only for the DC component H ₀ , and the sequence to be transmitted is determined by ΔH ₀ , H ₁ ,
H ₂ and H ₃ may be used. Table 1 below shows the sequence H ₀ a, H 1 a, H ₂ obtained by sampling a standard image at 7 bits/pixel and performing a Hadamard transform horizontally using a block size of ₂ ×2. A suitable quantization level is given to each of a and H ₃ a, and these are used as sequence quantization values. This quantization allows distortion D to each frequency component y within a range that does not cause visual deterioration in the spatial power spectrum of the image, and the amount of information R required for transmission at this time is Rate
From Distortion Theory, R [bit/pixel] = 1/2N _N-1 〓 ^y=0 Max {0, log ₂ σ ² y/D} ...(9) is given.

[Table] Pixel signals x ₁ to _{x 4} are input at 7 bits/pixel,
Output sequence of normalized Hadamard transform using block size 2×2 H _{0 a} to H ₃ a
requires a maximum of 8 bits. However, as a visually acceptable distortion level, the sequence
For example, white noise equivalent to 1 bit, 2 bits, ₂ bits, and 3 bits may be applied to H ₀ a, H ₁ a, H 2 a, and _{H 3} a, respectively. 1 bit, 2 bits, to each least significant bit (LSB) side of H ₃ a
By shifting the data by two or three bits, bit compression is performed and quantization is performed. This amount of bit compression is defined as the quantization characteristic value. After this,
Linear prediction is performed by expanding the quantized compressed values for the DC component H ₀ and the horizontal component H ₂ to obtain the prediction error component ΔH ₀ , and then the sequence quantized value ΔH ₀ ,
Based on the combination of H ₁ , H ₂ and H ₃ , combination number C ₀
~ Form _C3 . In Table 1, for example, in combination number C ₁ , only the quantized value ΔH ₀ of the DC prediction error component takes a value of "1" or "-1", and the other sequence quantized values
It shows that H ₁ to _{H 3} all take the value of "0", which means that the probability of occurrence of this combination is 23.9%. In addition, ±h indicates that the sequence quantization value can take any value, and the block code is a code that indicates a break within a block, and since there is little difference in the probability of occurrence of each combination, there are two Bits are allocated. Furthermore, the data code does not transmit any data for the combination number C ₀ , only the sign S of the sequence quantization value ΔH ₀ for the combination number C ₁ , and the sequence quantization value ΔH for the combination number C ₂ . ₀
This shows that 6 _bits of data are sent after the plus/minus sign S, and for _the combination number _{C3 ,} 3 _bits , 5 bit, 5 bit,
Allocating 7 bits, sequence quantization value ΔH ₀
This indicates that variable length code data corresponding to the combination of ~ _H3 is transmitted. Therefore, the code length to be transmitted is 2 bits for combination number _C0 , 3 bits for combination number _C1 , 9 bits for combination number _C2 , and ₂₂ bits for combination number C3. be. Note that the combination occurrence probability is determined as an average by testing various images with low spatial frequency such as portraits, images with high spatial frequency such as landscape photographs, etc. As described above, data below the distortion level given during quantization in the spatial spectrum of the image is discarded without being transmitted, and in the example of Table 1, the sequence quantization value H ₃ ,
H ₂ , H ₁ , and ΔH ₀ are compressed to 3 bits, 5 bits, 5 bits, and 7 bits, respectively, and then transmitted, resulting in an average amount of transmitted information of 2.1 bits/pixel, making it an extremely efficient code. It can be seen that it is a method of conversion. As mentioned above, using a variable length code means that the compression rate changes for each input pixel, and a buffer memory is used to absorb this difference between the encoding bit rate and the transmission bit rate. This buffer memory must have a capacity sufficient to absorb changes in the input/output bit rate, and the present invention saves the capacity of this buffer memory so that only a small capacity is required. FIG. 4 is a block configuration diagram showing an embodiment of the present invention, in which pixel signals x ₁ to _{x 4} for each block are input to a Hadamard transformer 10, converted to a frequency domain sequence, and then sent to a quantizer 20. and is quantized into sequence quantization values H ₀ to _{H 3} according to the quantization characteristic value QC. Then, the sequence quantization value H ₀ of the DC component is linearly predicted by the linear predictor 30 using the sequence quantization value H ₁ of the vertical component, and the prediction error component ΔH ₀ and the sequence quantization value H ₁ of the vertical component are predicted. , the sequence quantization value H ₂ of the horizontal component, and the sequence quantization value H 2 of the slope component
Each of H ₃ is input to an encoder 40 as described later, and a variable length code VC corresponding to the combination of each sequence quantized value ΔH ₃ to _{H 3} is formed. The variable length code VC formed by the encoder 40 is temporarily stored in the buffer memory 50 and then output as an image signal PD at a constant bit rate, and the memory controller 60 sequentially counts the amount of codes in the buffer memory 50 and stores the counting result. Control signal CN based on
is input to the quantization characteristic selector 70. Then, the quantization characteristic number ZF is determined according to the processing flow described later and inputted to the encoder 40, and also inputted to the ROM 80 for outputting the quantization characteristic value.
Quantization characteristic value corresponding to quantization characteristic number ZF from
The QC is read out and output, and quantization is performed in the quantizer 20 using the quantization characteristic value QC. In addition,
The contents of the ROM 80 and the ROM in the encoder 40 are as shown in Table 2 on the next page.

[Table] In other words, the ROM 80 stores a predetermined quantization characteristic value corresponding to the obtained quantization characteristic number ZF (0 to 3).
QC (H ₀ to _{H 3} ) is output and the ROM in the encoder 40 is
is the combination number obtained from the sequence quantization values H ₃ to ΔH ₀ in the quantization characteristic number ZF (0 to 3)
A predetermined variable length code VC is output according to _C0 to _C3 . For example, if the quantization characteristic number ZF determined by the quantization characteristic selector 70 is “2”, the ROM 80
The quantized characteristic value QC output from is the sequence
“2” for each of H ₀ , H ₁ , H ₂ , and H ₃ ,
“3”, “3”, “4”, and the code in the encoder 40 is
From the ROM, the combination number C ₀ ~ as mentioned above in Table 1
Data VC with a code length of 2 bits, 3 bits, 8 bits, or 18 bits is output depending on _C3 . In this way, in the present invention, the compression ratio is changed for each block line, and this can be done by changing the quantization characteristic value QC of the quantizer 20. A visually permissible compression ratio (output bit rate) is defined, and the quantization characteristic selector 70 first selects the code length of the next block line from the code length of the current block line according to FIG. 6, which will be described later. This is easy to predict because the images have strong spatial correlation. Next, it is detected whether or not there is an empty area in the buffer memory 50 that can accommodate this predicted value, and if there is no empty area, a quantization characteristic that coarsens the quantization is used for the next block line. Do it like this. Roughening the quantization characteristics has the effect of worsening the SN ratio, but since the encoding bit rate is higher than the output bit rate, this means that the image configuration of that block line is complex. At this time, even if the quantization is coarsened, the deterioration is not visually noticeable. On the other hand, if the encoding bit rate is lower than the output bit rate, it means that the block line had a smooth image structure, and in this case, the deterioration is easily noticeable visually.
In addition, there is a risk that the buffer memory will underflow, so it is necessary to perform fine quantization. The buffer memory 50 can be configured with a RAM (Random Access Memory) of about 256 bytes, for example, and the variable length code VC output from the encoder 40 is once input into a register in bit units, and then is converted into, for example, 8 bits. The signals are input to the buffer memory 50 in parallel in units. The buffer memory 50 is read out at a constant speed, for example, DMA
(Direct Memory Access) is stored as data in a magnetic recording device or the like. Further, if it is necessary to serially output data to a transmission line, the buffer memory 50 can be replaced with a serial-input/serial-output register. After all, while the input of the original pixel and the output from the buffer memory 50 have a constant bit rate, the input to the buffer memory 50 has a variable bit rate, and the buffer memory 50 is provided to absorb this excess or deficiency. memory controller 60
manages the buffer memory 50 for each block line, counts the amount of codes in the buffer memory 50 while the pixels of one block of the current image are input, and specifies the quantization characteristics for the next block line. There is. As described above, by changing the quantization characteristics while utilizing the visual characteristics in accordance with the output bit rate, that is, the average output bit length per input pixel, it is possible to maintain a constant output even with a small capacity buffer memory. It becomes possible to transmit at bit rate. The quantization characteristic number ZF to be used in the next block line may be encoded and transmitted next to the block line synchronization signal BL, and an example of its format is shown in FIG. Here, the state of the memory controller 60 and the quantization characteristic selector 70 that determine the quantization characteristic number ZF will be explained with reference to the flowchart of FIG. 6. Buffer memory 50 in 1 block line
The input bit length to is IL, the output bit length is OL, the capacity of buffer memory 50 is V, and buffer memory 5 is
Let the empty area of 0 be EP, and let the predicted value of the input bit length for the next block line be IL. If the bit length remaining in the buffer memory 50 after the end of the previous block line is PKi _-1 , then
The remaining bit length PKi of the buffer memory 50 after the end of the current block line is PKi = PKi _-1 + IL - OL ... (10), and the empty area EP of the buffer memory 50 at this point is EP = V - PKi ... (11) It is. The memory controller 60 performs the above calculation every time a block line ends, and calculates the bit length input to the buffer memory 50 in the next block line.
IL is predicted, and the quantization characteristic number ZF to be used for the next block line is determined from the relationship between IL, PK, EP, OL, and IL. In other words, when the processing routine for determining the quantization characteristic number ZF starts (step
S1), it is determined whether the block line BL is "1" (step S2), and the block line BL is
If BL is "1", the remaining bit length PK is set as the input bit length IL (step S4), and the block line BL is set as the input bit length IL (step S4).
is not "1", the input bit length IL is added to the remaining bit length PK in the previous block, and the output bit length OL is subtracted, and the value is set as the new remaining bit length PK.
(step S3), and then the remaining bit length obtained as described above from the capacity V of the buffer memory 50.
Find empty area EP by subtracting PK (step
S5). Then, the remaining bit length PK is the output bit length OL
Determine whether the following is true (step
S6), if the remaining bit length PK is larger than the output bit length OL, the predicted value IL of the input bit length is set as the sum of the input bit length IL and the bit length variation α between blocks (step S7); It is determined whether the empty area EP is less than or equal to the predicted value IL of this input bit length (step S8). If the empty area EP is less than the predicted value IL of the input bit length, "+
1” is set as the quantization characteristic number ZF (step
S9), set 1/2 of the input bit length IL as the new predicted value IL of the input bit length (step S10), and set the empty area EP
It is determined whether or not is smaller than the predicted value IL of the input bit length (step S11). If the empty area EP is smaller than the predicted value IL, a further calculation of "+1" is performed (step S12), and then,
Even when the empty area EP is greater than or equal to the predicted value IL, 1/4 of the input bit length IN is set as the new predicted value IL (step S13). Then, it is further determined whether or not the empty area EP is smaller than this predicted value IL (step S14), and if it is smaller, it is calculated by "+1" (step S15), and return processing is performed. If is greater than or equal to the predicted value IL, return processing is performed as is. In addition, in the above processing step S8, if the empty area EP is larger than the predicted value IL, the process also returns. On the other hand, in the above processing step S6, if the remaining bit length PK is less than the output bit length OL,
The value obtained by adding the bit length variation α between blocks to the remaining bit length PK is set as the predicted value IL of the input bit length (step S20), and it is determined whether this predicted value IL is larger than the output bit length OL ( step
S21). If the predicted value IL is larger than the output bit length OL, return processing is performed and the predicted value
If IL is less than or equal to the output bit length OL, the value obtained by subtracting "-1" from the current quantization characteristic number ZF is set as the new quantization characteristic number ZF (step S22), and then the input bit length IL and the remaining bit length PK are set as the new quantization characteristic number ZF. The added value is set as the predicted value IL of the input bit length (step S23), and it is determined whether this predicted value IL is smaller than the output bit length OL (step S24). If the predicted value IL is greater than or equal to the output bit length OL, the input bit length IL
The value obtained by doubling the value and adding the remaining bit length PK is set as a new predicted value IL (step S26). In addition, in processing step S24, the predicted value IL is equal to the output bit length.
If it is smaller than OL, calculate "-1" from the current quantization characteristic number ZF (step
S25) and proceeds to step S26 above. Then, it is further determined whether or not this predicted value IL is smaller than the output bit length OL (step S27), and the predicted value
If IL is greater than or equal to the output bit length OL, return processing is performed, and if the predicted value IL is smaller than the output bit length OL, the current quantization characteristic number "-"
The value obtained by "1" is set as a new quantization characteristic number ZF (step S28) and the process moves to return processing. Next, Hadamard transformer 10, quantizer 20,
Specific circuit examples of the predictor 30 and the encoder 40 are shown in FIG. ₇ and _FIG . The sequence H ₀ a to H ₃ a converted into the frequency domain is input to the quantization unit 20, and the sequence quantized value H ₂ of the horizontal component is quantized by the quantizer 22 constituting the quantization unit 20. , the sequence quantized value H ₁ of the vertical component quantized by the quantizer 23, and the sequence quantized value H ₃ of the slope component quantized by the quantizer 23 are input to the subsequent variable length code conversion circuit 40, respectively. be done. Further, the sequence quantized value H ₀ of the DC component, which is the output of the quantizer 21, is inputted together with the sequence quantized value H ₂ to a predictor 30 as shown in FIG.
The prediction error component ΔH 0 predicted by ΔH ₀ is input to the variable length encoder 40 . Note that the quantizers 21 to 24
are each composed of a parallel shifter or a shift register, and for the sequence H ₀ a to H ₃ a converted by the Hadamard converter 10, the ROM 80
The bits are simultaneously shifted to the lower bit side by the number of bits according to the quantization characteristic value QC transmitted from the quantization characteristic value QC. For example, the quantization characteristic selector 70
When the quantization characteristic number ZF is set to “1” by
QC is “1”, “2” for sequence H ₀ to _{H 3} ,
“2” and “3” are output, 1 bit for DC component H ₀ a, 2 bits for vertical component H ₁ a,
2 bits for horizontal component H ₂ a, slope component H ₃ a
For each, three bits are shifted to the lower bit side. Also, the predictor 30
The expander 31 is composed of, for example, a parallel shifter, and since the input sequence quantization value H ₀ of the DC component and the sequence quantization value H ₂ of the horizontal component are different in level, the level of the horizontal component H ₂ is different. The delay circuits 32 and 33 each serve to delay the time by one block. On the other hand, the encoder 40 is a PLA (Programmable
Logic Array), and the combination number determined by this comparator 41
C ₀ to _{C 3} are input to a ROM (Read Only Memory) 42 as address signals. Then, the quantization characteristic number transmitted from the quantization characteristic selector 70
Code length read from ROM42 according to ZF
CL is set in counter 43, and block code BC is set in shift register 45. When the code length CL is set in the counter 43, the signal P
goes into the set state, starts counting down by the clock pulse CP, and when the counter 43 reaches 0, the signal P goes into the reset state. When signal P is in the set state, clock pulse CP and signal P are connected to the shift signal via AND gate 46.
Send SS to shift register 44. Here, the shift registers 44 and 45 are parallel-in/serial-out type shift registers. Further, the horizontal component H ₂ , vertical component H ₁ , slope component H ₃ from the quantization unit 20 and the prediction error component ΔH ₀ from the predictor 30 are input to the shift register 44 . The shift data shifted out from the block code BC is input to a shift register 45 storing a block code BC, and the output of the shift register 45 is input to a serial-in/parallel-out type register 48. When a fixed number of bits are input to the register 48, the entire data is read out in parallel, and after being temporarily stored in the buffer memory 50, it is transmitted to another device or stored as an image signal PD. In such a configuration, when pixel signals x ₁ to _{x 4} read out by simultaneous reading of two scanning lines or reading of one scanning line and a delay of one scanning line are input to the Hadamard transformer 10, these pixel signal
x ₁ to _{x 4} are converted into frequency domain sequences H ₀ a to H ₃ a and input to the quantization unit 20 . When the quantization characteristic number ZF is, for example, "3", from Table 2, the quantizer 21 of the quantizer 20 shifts the DC component H ₀ a by 3 bits to the lower side,
The quantizer 22 shifts the horizontal component H ₂ a to the lower side by 4 bits, and similarly the quantizer 23 shifts the horizontal component H 2 a to the lower side.
H ₁ a is shifted to the lower side by 4 bits, and the quantizer 23 shifts the slope component H ₃ a to the lower side by 5 bits. Here, positive and negative signs are shifted from the upper bit side. Among the sequence quantized values H ₀ to _{H 3} obtained in this way,
H ₁ to _{H 3} are input to the comparator 41 of the variable length encoder 40 and are also input to the shift register 44 .
Then, the DC component H ₀ is input to the predictor 30,
In this predictor 30, the DC component H ₀ along with the horizontal component H ₂
A prediction error component ΔH ₀ is formed, and this prediction error component ΔH ₀ is input to the comparator 41 and also to the shift register 44 . Here, the comparator 41 is
Programming is configured in PLA, and combination numbers C 0 to C ₀ according to Table 1 described above are based on the input sequence quantization values ΔH ₀ , H ₁ , H ₂ , and H ₃ .
_C3 is determined, and one of the combination numbers _C0 to _C3 is set as a binary signal, for example, "1". For example, if the prediction error component ΔH ₀ is "1", other sequences H ₁ ~
If H ₃ is all “0”, combination number C ₁
only becomes “1”, and other combination numbers C ₀ and C ₂ ,
_C3 is all "0". In this way, the combination number determined by the comparator 41
When C ₀ to _{C 3} are input to the ROM42, the ROM42
uses “1” of the combination numbers C ₀ to _{C 3} as an address signal, reads the code length data CL and block code BC stored at the corresponding address corresponding to the quantization characteristic number ZF, and reads out the code length data CL and block code BC.
CL is set in the counter 43 and block code BC is set in the shift register 45.
Therefore, for example, if the quantization characteristic number ZF is "2"
When the combination number _C2 is determined and input to the ROM 42, the code length "9" is set from the ROM 42 in the counter 43, and "10" is set in the shift register 45, corresponding to the countdown of the counter 43. The contents of shift registers 44 and 45 are shifted out. Therefore, 9 bits below "10" are input to the shift register 48. Further, when the comparator 41 determines that the combination number C _{is 0} , the counter 43 contains the code length "2".
is set, and “00” is stored in the shift register 45.
is set, and the shift register 48
only "00" is entered. The data serially input to the shift register 48 in this way is
The signals are converted in parallel and stored in the buffer memory 50 in units of, for example, 8 bits, and after time axis correction is output as an image signal PD. As mentioned above, for example, low bit rate digital
Assuming a VTR, in order to perform real-time processing at an input speed of the original image of 5.6M pixels/second and a transfer speed of 14M bits/second to the transmission path or recording device, the buffer memory 50 outputs a band-compressed signal. will be output at a bit rate of 2.5 bits/pixel, but due to variations in the compression ratio depending on the pixel, the buffer memory 5
If the input to 0 becomes less than or more than 2.5 bits/pixel, underflow or overflow of buffer memory 50 may occur.
In order to avoid this, a large capacity buffer memory was conventionally required, but by using the method of the present invention as described above, it is possible to output at a constant bit rate without excess or deficiency even with a small capacity buffer memory. In the embodiment shown in FIG. 7, dividers 11 to 14 are inserted into the output section of the Hadamard transformer 10 for ``1/2'', but the quantization section 2
0, it is common to perform the calculation of "1/2".

[Brief explanation of drawings]

FIG. 1 is a diagram for explaining the principle of this invention.
Fig. 2 is a diagram for explaining the linear prediction method used in this invention, Fig. 3 is a diagram for explaining the relationship between block size and Hadamard transform, and Fig. 4 is a block configuration showing an embodiment of this invention. 5 is a data transmission format diagram, FIG. 6 is a flowchart showing an example of the operation of selecting a quantization characteristic value in this invention, and FIGS. 7 and 8 are block configurations showing details of FIG. 4. It is a diagram. 1... Pixel, 10... Hadamard transformer, 21
~24...quantizer, 30...predictor, 31...
Expander, 32, 33... Delay circuit, 40... Variable length encoding circuit, 41... Comparator, 42, 80...
ROM, 43... Counter, 44, 45, 48...
...Shift register, 50...Buffer memory, 6
0...Memory controller, 70...Quantization characteristic selector.

Claims (1)

[Scope of Claims] 1. An input image is divided into a group of adjacent blocks consisting of a pixel group m(1, 2)×n (≧2), and the pixel groups in each block are transformed into a frequency domain by Hadamard transform. After converting into a sequence of and converting the sequence quantization values into values, forming a plurality of combinations thereof by comparing the sequence quantization values with each other, variable-length encoding the sequence quantization values corresponding to each of these combinations, and storing the sequence quantization values in a buffer memory; An adaptive quantization method for an image signal, characterized in that the image signal is read out from the memory at a constant bit rate and outputted, and the quantization characteristic for the next block line is selected depending on the amount of code in the memory. 2 Divide the input image into adjacent blocks of 2×2 pixels, and transform the pixels in each block into DC components, vertical components, horizontal components, and tilt components in the frequency domain by Hadamard transform. After that, quantization is performed using the currently selected quantization characteristic among the plurality of quantization characteristics, and at least the DC component of the sequence quantized value is linearly predicted and converted into a quantized value of the DC prediction error component. With,
A plurality of combinations are formed by comparing the sequence quantization values in one block, and each combination is given a predetermined variable length code and stored in a buffer memory. At the same time, the input bit length to the buffer memory in the first block line is IL, the output bit length from the buffer memory is OL, the capacity and empty area of the buffer memory are V and EP, and the input to the next block line is Let IL be the predicted value of the bit length IL, select a quantization characteristic that finely quantizes when the first condition of V-EP≦OL, IL^≦OL, and select the second condition of EP≦IL^. When the conditions are met, a quantization characteristic that coarsens the quantization is selected, and when conditions other than the first and second conditions are met, the quantization characteristic used for quantization of the current block line is used as the quantization characteristic of the next block line. An adaptive quantization method for image signals, which is characterized in that it can be used as is for quantization.