JPH026471B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an interframe encoding device that highly efficiently encodes an image signal (reduces the bit rate) by utilizing the correlation between successive screens. Conventionally, this type of interframe coding apparatus has been constructed as shown in FIG. In the figure, a is an encoder, and b is a decoder. In FIG. 1, 1 is an A/D converter, 2 is a subtracter, 3 is a scalar quantizer, 4 is an adder, 5 is a frame memory, and 12 is a D/A converter. Now, the A/D converter 1 digitizes the analog image signal 6 of the f-th frame (f is an integer) which is sequentially raster-scanned from left to right from the top of the screen to the bottom. The sample sequence of this digital image signal 7 is S ^f _t (here, the sample sequence number in raster scanning), and the predicted signal 8 consisting of sample values at the same pixel position one frame before the S ^f _t is P^ ^f _t , the image signal 7 and the predicted signal 8
The inter-frame difference signal 9 of ε ^f _t is a color quantized signal 10 for each sample of the inter-frame difference signal 9.
ε^ ^f _t , the scalar quantized signal 10 and the prediction signal 8
The reproduced image line signal 11 obtained by adding the above is defined as S^ ^f _t . At this time, the encoder shown in FIG. 1a performs the following processing. ε ^f _t =S ^f _t −P^ ^f _t ε^=ε ^f _t +qk ^f S^ ^f _t =P^ ^f _t +ε^ ^f _t =S ^f _t + ^qk kHere, P^ ^f _t =S^ ^f _t ·Z ^−f represents a delay of one frame), and q ^f _t is the scalar quantization noise. That is, there is a correlation between the image signals of the current frame and the previous frame, and their values are close. Therefore, the inter-frame difference signal ε ^f _t is close to 0 and has lower power than the original signal S ^f _k . On the other hand, from the above equation, the quantization noise of the reproduced image signal is the same as the quantization noise of the inter-frame difference signal. If scalar quantization is performed using the interframe difference signal ε ^f _t that has lower power than the original signal S ^f _k , quantization noise will not increase even if the number of quantization levels is reduced. Therefore, the inter-frame difference signal ε ^f _t is scalar quantized into ε^ ^f _t using a small number of quantization levels so that the quantization arithmetic is minimized based on its amplitude probability distribution P(εk). It is sufficient to allocate each code to each quantization level of this ε^ ^f _t and perform high-efficiency encoding for transmission. The decoder shown in FIG. 1b performs the calculation S^ ^f _t = P^ ^f _t + ε^ ^f _t = S ^f _t + qk ^k from the transmitted color quantized signal 10 to eliminate quantization noise. A reproduced image signal 11 containing q ^f _t is obtained. As described above, the conventional interframe encoding device scalar quantizes the difference signal between the image signal of the current frame and the image signal of the previous frame corresponding to the same position by reducing the number of levels for each sample. According to this method, since the image signal includes Gaussian noise, it is difficult to encode only the significant interframe difference signal generated by the movement of the subject. Furthermore, since the interframe difference signal is scalar quantized sample by sample, it is impossible to reduce the bit rate below a predetermined bit rate. This invention was made to eliminate the above-mentioned drawbacks, and is resistant to disturbances such as Gaussian noise.The present invention is a frame-by-frame method that vector quantizes only the inter-frame difference signals determined to be significant motion in block units. The present invention provides an encoding device. The present invention will be explained in detail with reference to the drawings. Now, as shown in Figure 2, samples on four temporally consecutive scanning lines of the f-1 numbered frame (in the case of television, one frame consists of two fields that are scanned in a skip). The blocks are grouped into blocks of 4x4 in a grid pattern, and this is
S ^f-1 _l = [S ₁ , S ₂ , ..., S ₁₆ ] as two vectors
Expressed as ^f-1 _t . On the other hand, in the fth frame after one frame (after two fields in television scanning), a block of samples corresponding to S ^{f -1} _l and the same position on the screen S ^f _l = [S ₁ , S ₂ ,...,S ₁₆ ] ^f _l . Here, l is an integer and represents the block sequence number when the block is shifted from left to right on the screen and further from top to bottom. The present invention provides an interframe encoding device that performs motion detection vector quantization based on the interframe difference signal in block units to realize highly efficient encoding. FIG. 3 shows an example of the configuration of the encoding section of the vector quantization type interframe encoding device according to the present invention, and FIG. 4 shows an embodiment of the configuration of the decoding section of the device. In FIG. 3, 14 is a raster/block scan converter, 15 is a mid-tread type limiter, and 1
6 is a block scanning frame memory, 17 is a vector quantization encoder, 18 is a vector quantization decoder, 19 is a transmission data buffer, and 20 is a threshold control circuit. Furthermore, in FIG. 4, 31 is a receiving data buffer, and 33 is a block/raster scan converter. In the drawings, the same reference numerals and the same or equivalent parts are shown in FIGS. 1, 3, and 4. Next, the operation of the encoding section of the interframe encoding device shown in FIG. 3 will be explained. First, the digitized image signal 7 of the f-th frame passes through a raster/block scan converter 14 and is converted from raster scanning to block scanning as shown in FIG. 2 on the screen. In block scanning, each sample of the image signal is divided into [S ₁ , S ₂ ,
S ₃ , S ₄ , S ₅ , S ₆ , ...S ₉ , S ₁₀ , ..., S ₁₃ , S ₁₄ ,
S ₁₅ , S ₁₆ ] _l is scanned in this order, and the block scan moves from left to right on the screen and then from top to bottom to output a block scan image signal 21. (Here, the main scanning direction and sub-scanning direction within the block may be interchanged.) The block scanning image signal 21, that is, S ^f _l
are the block scan prediction signals 22 read out from the block scan frame memory 16 P^ ^f _l = P^ ₁ , P^ ₂ ,
..., P^ ₁₆ ] ^f _l for each element (for each corresponding sample)
is subtracted through subtractor 2. Next, the prediction error signal 23 which is the output of the subtracter 2 ε _l =[ε ₁ , ε ₂ ,
..., ε ₁₆ ] _l is the suppression that suppresses minute level fluctuations and overloads in the time axis direction through a mid-trend type limiter 15 having input/output conversion characteristics as shown in Fig. 5 (a in the figure is a limiter level constant). The prediction error signal 24 is converted into ε′ _l =[ε′ ₁ , ε′ ₂ , ..., ε′ ₁₆ ] _l . This suppressed prediction error signal 24 is vector quantized in block units by a vector quantization encoder 17 and a vector quantization decoder 18, resulting in a vector quantization prediction error signal 26 ε^ _l = [ε^ ₁ ,
ε^ ₂ , ..., ε^ ₁₆ ] converted to _l . The vector quantized prediction error signal 26 is added to the block scan prediction signal 22 in an adder 5 for each element to obtain a reproduced block scan image signal 27 S^ ^f _t = [S^ ₁ , S^ ₂ , . . . S^ ₁₆ 〕 ^f _l
Play. This reproduced block scan image signal 27 is delayed by one frame in the block scan frame memory 16 and is used as the block scan prediction signal 22 for the next frame. The above inter-frame difference calculation is expressed as an inter-vector calculation shown in the following equation. ε _l = S ^f _l − P^ ^f _t ε^ _l = ε _l + Q _l S^ ^f _t = P^ ^f _t + ε^ _l = S ^f _l + Q _l P^ ^f _t = S^ ^f _t・Z ^-f Here, Z ^-f is a delay of one frame, and Q _l is an operational error caused by the mid-trend type limiter 15 and vector quantization of ε _l . In the above process, the vector quantization encoder 17 performs the following processing on the suppressed prediction error signal 24 in units of blocks to perform high-efficiency encoding. In other words, the case where the component corresponding to the average value and standard deviation from the average value in the block of ε′ _l = [ε ₁ , ε ₂ , ..., ε ₁₆ ]′ _l exceeds the threshold T Vector quantization is performed as a certain significant block, and if the threshold value T〓 is not exceeded, it is treated as an insignificant block so that the vector quantization prediction error signal 26 ε^ _l = 0.
Details of vector quantization will be described later. Vector quantization encoded data 25, which is the output of the vector quantization encoder 17, is sent to a transmission data buffer 19. The transmission data buffer 19 sends an information generation amount control signal 28 to the threshold control section 20 to control the amount of information generation of the vector quantized encoded data 25 so that the data buffer does not exceed a predetermined capacity. The quantized encoded data 30 is maintained at a constant transmission rate. The threshold control section 20 feeds back a threshold signal 29 to the vector quantization encoder 17 based on the information generation amount control signal 28. Threshold T〓
It is best to move it up and down in frame units. In this case 1
If the frame cycle period T〓>2·a, frame frames will be dropped. Next, the operation of the decoding section of the interframe encoding device will be explained with reference to FIG. The vector quantized encoded data 25 subjected to speed conversion through the reception data buffer 31 is converted into a vector quantized prediction error signal 26 ε^ _l by the vector quantized decoder 18. The adder 5 and the block scan frame memory 16 perform the reverse procedure of interframe coding on the signal ε^ _l to restore the reconstructed block scan image signal 27 S^ ^f _t . That is, as a vector operation, S^ ^f _t = P^ ^f _t + ε^ _l = S ^f _l + Q _l . The block scan image signal 27 is converted into a raster scan reproduced image signal 34 by a block/raster scan converter 33, and is D/A converted to obtain an analog reproduced image signal. Next, a vector quantizer that compresses the block scan prediction error signal, that is, the interframe difference signal, and encodes it with high efficiency will be described in detail. FIG. 6d is a block diagram showing an embodiment of a vector quantization encoder and FIG. 6b is a vector quantization decoder according to the present invention. In the figure, 36 is a DC separation circuit, 37 is a normalization circuit, 38 is an address counter, 39 is an output vector code table, 40 is an input/output vector distortion calculation circuit, 41 is a minimum distortion detection circuit, 42 is an index latch, 50 54 is a motion detection circuit, 54 is a gain regeneration circuit, and 55 is a DC synthesis circuit. Note that the same reference numerals in the figures indicate the same or corresponding parts. The operation of the vector quantizer will be explained below. In the vector quantization encoder of FIG. 6a,
For the suppression prediction error signal 26 ε′ _l , the DC separation circuit 3
6 and the normalization circuit 37 perform the following processing to convert it into _an input vector 44Xl . That is, ε′ _l =
[ε′ ₁ , ε ₂ , ..., ε′ ₁₆ ] Intra-block average value mg of _l
and the within-block standard deviation (or equivalent component) σ _l , m _l =1/16 ₁₆ 〓 ^j=1 ε′ _j (j=1, 2, ..., 16) σ _l = [1/16 ₁₆ 〓 ^j=1 (ε′ _j −m _l ) ² 〓 ^1/2 or σ _l = ^max _j ε′ _j − ^min _j ε′ _j or σ _l =1/16 ₁₆ 〓 ^j=1 | ε′ _j −m _l | X _j = ( _ε ′ _j −m _l ) _/ σ _l X _l [ X ₁ , _{X 2 , ...}
get. The average value signal _50ml and the deviation signal _51σl are sent to the motion detection circuit 52, and compared with the threshold signal 29, it is determined whether ε^ _ll is an interframe difference signal for a significant motion. used for The motion detection circuit 52 determines that when m _l and σ _l are smaller than the threshold T〓, m _l and σ _l are sufficiently small and can be quantified to zero, so m _l < T〓 and σ _l ＜T〓
Then, quantize as m _l =σ _l =0. After performing this processing, the processed average value m ₁ and deviation σ _l are sent to the index latch 42 as a motion block average/deviation signal 53. Index latch 42 is m _l
=σ _l =0, it is determined that ε' _l is an insignificant block with no movement, and only an insignificant block code (1 bit) is transmitted as the vector quantized encoded data 25 of ε' _l . When m _l ≠0 or σ _l ≠0, a significant block code (1 bit), m _l , σ _{l ,} and an output vector index signal 49 to be described later are sent as vector quantized encoded data 25. That is, the significant/insignificant block identification code can also be used as a prefix when vector quantized encoded data 49 is variable length encoded (for example, Huffman coating). At this time, a block of 16 samples with no motion is efficiently coded to 1 bit (ie, 1/16 bit for each sample) for the random block code. Next, the principle of vector quantization, which realizes highly efficient coding even for significant blocks, will be explained. Now, we have K input signal sample series (K is a positive number).
Collectively input vector X = [X ₁ , X ₂ , ..., X _k ]
shall be. In this case, the K-dimensional Euclidean signal space
Let the predetermined division of R ^k ( X ∈R ^k ) be R ₁ , R ₂ , ..., R _N
and the representative point of each division (e.g. center of gravity) y i=
Let Y = N set of [yi ₁ , yi ₂ , ..., yik]
Let [ y ₁ , y ₂ , ..., y _N ]. Mapping the representative point y i as the output vector of the input vector X included in the division R _i is called vector quantization. At this time, vector quantization Q _V is defined by the following equation. Q _V :R ^k →Y Here, R _i =Q ^-1 _V ( y i)=[ X ∈R ^k :Q( X = y i _N ∪ ⁱ⁼¹ R _i =R ^k , R _i ∩R _j =0 (i≠j) The above vector quantization Q _V is the decoding D _V and the decoding D _V
is represented as a cascade of . The encoding C _V is the set of output vectors in R ^k Y = [ y ₁ , y ₂ , ...,
y _N ] index set I = [1, 2, ...,
N], and the decoding D _V is a mapping from I to Y. That is, C _V : R ^k → I, D _V : I → Y Q _V = D _V · C _V. In the vector quantization, the encoded output I is transmitted or recorded, so extremely efficient encoding can be realized. The set Y of output vectors in vector quantization consists of a large number of input vectors X l _generated from a model of ^the input signal source based on the amplitude probability distribution P( It can be determined by clustering training in which the sum of distortions of vectors (temporarily set at initialization) 〓 ^l min ⁱ d( X _l , y i) is minimized. FIG. 7 shows an example of the arrangement of output vectors in the signal space R16 when K= ¹⁶ . Vector quantization is a cascade of encoding, which is a mapping of the output vector to the index at the closest distance (minimum distortion) to the input vector X , and decoding, which is the mapping from the index to the output vector.
^At this _time ^, CV : X → i = [ i | min ⁱ d ( , y i). Here , _if we ^define _the distortion ^d ₍ ^_ Using Hausdorff's distance to _measure d ( X , y i)= max ^j |X _j −y _ij | Using _the sum ^of absolute values, d( y _ij | In order to carry out the above vector quantization, the vector quantization encoder shown in FIG. 6A writes a set of output vectors generated by clustering in the output vector code table 39 in advance. Now, when the input vector 44 X _l is input, the address counter 38 is
..., sequentially counts up to N, and reads out the output vector y i corresponding to the index i from the output vector code table 39 in the order of y ₁ , y ₂ , . . . , y _N . Next, the distortion d between the input vector X _l and each output vector y i read out sequentially
( X _l , y i) is calculated by the distortion calculation circuit 40. This distortion signal 44 between the input and output vectors is sent to the minimum diameter detection circuit 41, and min d( X _l , y i) (i=1,
2,...,N) is found. this is,
i=1, 2, ..., N and sequential address counter 3
8 counts up, it is successively compared with the past minimum distortion, and if the distortion is small, it is detected as the current minimum distortion. If the strobe signal 48 is sent every time the minimum distortion is detected and the address signal 49 at this time is taken into the index latch 42, then i=
When N, the index of the minimum distortion output vector is obtained in the index latch 42. The index latch 42 uses the index of this minimum distortion output vector and the motion block average/deviation signal 2.
5 and the significant/insignificant block identification code are output as vector quantized encoded data 25. Next, the operation of the vector quantization decoder shown in FIG. 6b will be explained. The significant/insignificant block identification code, index, average value, and deviation signal sent as vector quantized encoded data 25 are taken into an index latch 42. The index is
It is sent as an address signal to the same output vector code table 39 as the vector quantization encoder section and reads out the corresponding output vector 56 y i = [yi | min ⁱ d( X _l , y i)]. For this output vector y i, the gain regeneration circuit 54 receives the deviation σ _l and performs gain regeneration.
The DC synthesis circuit 55 receives the average value m _l and performs DC regeneration to obtain a reproduced prediction error signal 26 ε^ _l .
That is, ε^ _i = σ _l・y _ij + m _l (i=1, 2, ..., 16) ε^ _l = [ε^ ₁ , ε^ ₂ , ..., ε^ ₁₆ ] _l However, m _l = σ If _l = 0, ε^ = [0, 0, ..., 0] _l . Through the above processing, the vector quantization type interframe encoding device according to the present invention reduces the quantization level by suppressing the time axis type particulate noise and overload of the interframe difference signal by the mid-to-length limiter, and further, By detecting motion on a block-by-block basis and using the average value and deviation of the block to distinguish between presence/absence and non-existent blocks, the accuracy of motion detection against noise is improved and highly efficient coding is achieved. Furthermore, for significant blocks with movement, highly efficient coding is achieved by separating the mean values in block units, normalizing them by gain, and vector quantizing them in a multidimensional signal space.
The image signal is subjected to normal value separation and normalization processing so that the amplitude rate distribution of the input vector becomes similar for different scenes, thereby improving the efficiency of vector quantization. Since vector quantization is encoded in block units, the countermeasure against transmission errors between frames in interframe encoding is that the block scanning image signal 21
This can be achieved by directly transmitting vector quantization. Although the above description has been made assuming that the block size of vector quantization is 4×4, it goes without saying that this block size can be freely selected. Furthermore, in the present invention, vector quantization has been explained using a full search that requires distortion calculation with all output vectors, but vector quantization can also be performed using high-speed vector quantization that generates output vectors in a tree structure and performs a tree search. good. Further, different values may be used for the average value and the deviation for the motion detection threshold value. As described above, the vector quantization type interframe coding device according to this invention is configured to perform high efficiency coding by DC separated normalized vector quantization including noise suppression and motion detection of the interframe difference signal. This has the effect of realizing high-quality video transmission at an extremely low bit rate.

[Brief explanation of drawings]

FIGS. 1a and 1b are block diagrams of a conventional interframe encoding device. FIG. 2 is an explanatory diagram of the pixel arrangement on the screen of the interframe block scanning image signal. FIG. 3 is a configuration diagram showing an embodiment of the encoding section of the vector quantization interframe encoding device according to the present invention. Fourth
FIG. 1 is a configuration diagram showing an embodiment of a decoding section of a vector quantization interframe encoding device according to the present invention.
FIG. 5 is an explanatory diagram of the input/output conversion characteristics of the mid-tread type limiter for interframe difference signals according to the present invention. FIGS. 6a and 6b are block diagrams showing an embodiment of a vector quantizer according to the present invention, in which a represents an encoder, and b
is a diagram showing a decoder. FIG. 7 is an explanatory diagram of the relationship between input and output vectors in a 16-dimensional space in vector quantization. In the figure, 1 is an A/D converter, 2 is a subtracter, 3 is a frame memory, 4 is a scalar quantizer, 5 is an adder, 12 is a D/A converter, 14 is a raster/block scan converter, 15 is a mid-length limiter, 16 is a block scanning frame memory, 17 is a vector quantization encoder, 18 is a vector quantization decoder, 19 is a transmission data buffer, 20 is a threshold control circuit, 31 is a reception data buffer, 3
3 is a block/raster scan converter, 36 is a DC separation circuit, 37 is a normalization circuit, 38 is an address counter, 39 is an output vector code table, 4
0 is a distortion calculation circuit, 41 is a minimum distortion detection circuit, 42 is an index latch, 54 is a gain regeneration circuit, 55
is a DC synthesis circuit. In addition, in the figures, the same reference numerals indicate the same or corresponding parts.

Claims (1)

[Claims] 1. A frame memory that always stores at least one frame of image signals, and a frame memory that stores K input signal sequences (K
When the latest image signal, which has been block-blocked for each frame, is input, the block-blocked predicted signal sequence corresponding to the same position on the screen at least one frame before is read out from the storage unit, and the block-shaped predicted signal sequence is read out from the storage unit. a subtracter that calculates a difference signal sequence; a mid-tread limiter that suppresses minute fluctuations and overloads in the inter-frame difference signal; and a subtracter that separates the average value within a block of the suppressed inter-frame difference signal sequence and calculates the deviation from the average value. A DC separation normalization circuit that normalizes the signal by its component and converts it into an input vector, and a block in which the intra-block average value and the deviation component are less than a predetermined threshold value is regarded as an insignificant block, and all blocks in the inter-frame difference signal sequence are set to zero. A motion detection circuit that executes the processing and input vectors of significant blocks whose deviations from the average value exceed a threshold value are selected in advance to form a predetermined number of representative points in the K-dimensional signal space R ^k , that is, an optimal arrangement. the output vector; A corresponding output vector is selected from the vector identification code, multiplied by the deviation component, and the average value is added to reproduce the inter-frame difference signal sequence, and a vector quantization complex is performed to make the inter-frame difference signal sequence zero for the meaningless block. a coding unit; an adder that adds the inter-frame difference signal sequence reproduced after vector quantization and the predicted signal sequence to reproduce an image signal and writes it into the frame memory; and the significant/insignificant block identification code and an output vector. A vector quantization type interframe coding device characterized by comprising a transmission data buffer that variable-length encodes the identification code and the significant block average value and deviation, and controls a threshold value to keep the amount of information generated constant. . 2. Normalization of the inter-frame difference signal sequence or deviation component from the intra-block average value for motion detection is the difference between the maximum and minimum values of the intra-block signal sequence or the absolute value of the difference from the intra-block signal sequence average value. 2. A vector quantization type interframe encoding device according to claim 1, characterized in that a summation is used. 3 In a vector quantization encoder, the distortion measure (distance in K-dimensional signal space) between input and output vectors
A distortion calculation unit that calculates as distortion the maximum value of the absolute values of the differences between each element of the input vector and the output vector, or the sum of the absolute values of the differences between each element of the input vector and the output vector. 2. The vector quantization interframe encoding device according to claim 1, further comprising a distortion calculation unit that calculates the distortion.