JPS63217878A - Prediction tree search vector quantizing system - Google Patents

Abstract

PURPOSE:To efficiently encoding a vector index by setting a range of a tree search by using a precedence index. CONSTITUTION:A belt signal (x)=(x0, x1,...xN-1) in which N pieces of signals have formed one block is inputted from an input terminal 1. It is supposed that a pair gamma<i>,<0>, gamma<i>,<1> of the i-th reference vectors are outputted from a code book 7. A distance calculating circuit 2 calculates a distance of the input vector (x) and the reference vectors gamma<i>,<0>, gamma<i>,<1> and outputs its result by '0', '1'. A vector instructing circuit 6 updates (i) to 2i in case '0' is inputted from a minimum distance deciding circuit 3, and updates (i) to 2i+1 in case '1' is inputted. As a result, from the code book 7, subsequently, gamma<2i>,<0>, gamma<2i>,<1> or gamma<2i+1>,<0>, gamma<2i+1>,<0> are outputted and repeated by a necessary number of times. An index delaying circuit 5 delays a vector index, and outputs a vector index of the vicinity of an encoding block.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a predictive tree search vector quantization method for image signals, etc., and applies correlation between vector indexes to tree search vector quantization.

(Prior Art) Vector quantization methods and predictive vector quantization methods are conventionally known. In the vector quantization method, an input signal such as an image signal is represented by a vector having multiple components, and each vector is approximated by the vector closest to the input vector among the representative vectors prepared in advance in a table called a codebook. It is something that is transmitted. A full search vector quantization method and a tree search vector quantization method are known as methods for searching for the closest vector.

The exhaustive search vector quantization method calculates the distance between all representative vectors in the codebook and the input vector to find the closest vector, and while it provides the best performance, it requires a huge amount of distance calculation operations. On the other hand, the tree search vector quantization method uses a codebook with a tree structure to sequentially reduce the candidates for the optimal representative vector and arrive at the optimal representative vector.While the amount of calculation can be reduced, the quantization performance is Deteriorates somewhat. In any case, in the vector quantization method, one representative vector is associated with an input vector, and the correlation between representative vectors is not utilized.

On the other hand, the predictive vector quantization method is a method that makes it possible to utilize the correlation between representative vectors, and there is a method shown in FIG. 6 or FIG. 7.

Figure 6 uses the correlation between vector indices indicating representative vectors, and assigns a variable length code to the code of the vector index according to the probability of appearance of the index of the encoded vector that depends on the preceding vector index. .

This method enables encoding close to the control key of the vector index. In other words, the correlation between vector indices can be utilized. In FIG. 7, the codebook for the encoded vector is switched depending on the preceding vector index, and the correlation between vector indexes can also be used.

(Problem to be solved by the invention) The method shown in Fig. 6 has the advantage of simplifying the processing because it is possible to perform processing in the area of entropy encoding, but the method shown in Fig. 6 has the advantage of simplifying the processing. It is necessary to prepare a variable length code table. Further, although the method shown in FIG. 7 provides ideal vector quantization characteristics, it is necessary to prepare codebooks equal to the number of states classified by the preceding index.

Therefore, it has been difficult to easily implement the methods shown in FIGS. 6 and 7.

An object of the present invention is to provide a simple prediction tree search vector quantization method that solves this problem.

(Means and effects for solving the problem) In order to achieve the above object, the present invention sets the range of tree search for encoded vectors by the preceding vector index, and has a simple configuration. The feature is that the correlation between vector indices can be used by.

(Embodiment) FIG. 1 is a diagram explaining an embodiment of the present invention, in which 1 is a vector input terminal, 2 is a distance calculation circuit, 3 is a minimum distance judgment circuit, 4 is a vector index output terminal, and 5 is a diagram for explaining an embodiment of the present invention. 6 is an index delay circuit, 6 is a vector instruction circuit, and 7 is a codebook.

Vector signal x = (xO+ X 1 + ”' X N-1
) is input. Codebook 7 outputs the first reference vector pair γi”Oi・1,
γ. The distance calculation circuit 2 calculates the distance between the input vector
If , output 1. This result is output as vector index information, and the vector instruction circuit 6
is input to. The vector instruction circuit 6 updates i to 21 when 0 is input from the minimum distance determination circuit 3, and updates l to 21+1 when 1 is input. The result is codebook 7 or 21.0 21.1 Again, 21+
1.0, etc., then γ, γγ21+1°1 are output and repeated as many times as necessary.

For example, if this number of times is 5, the tree search shown in FIG. 2 is executed and one of the representative vectors vol v, l . . . V31 is determined. As a result, the collection of 0 and 1 codes output from the minimum distance determination circuit 3 is vOr
. . . This becomes a double code indicating V31, which forms a vector index. The vector index is output from the index output terminal 4 and is also input to the index delay circuit 5. The index delay circuit 5 delays the vector index and outputs the vector index near the encoded block. For example, as shown in FIG. 3, according to the encoding order, the index ld2 of the immediately adjacent block to the left and the index ld2 of the block immediately above
, outputs. This ' dt + i d2 is input to the vector instruction circuit 6. The vector instruction circuit 6 not only updates the reference vector number l using the output from the minimum distance determination circuit 3 as described above, but also uses 'dt and 1d2 to update the initial vector number Pid1. It has a function to determine id2.

For example, if the signal of the left block is known from idl and id2, the signal of the encoded block may be predicted to some extent. For example, in FIG. 2, if it is assumed that the signals can be limited to numbers 10 and below, it is sufficient to set the initial vector to 1-10 and execute the following search. If you start from number 1, vOr...5 to get V31
In contrast, starting from number 10, two searches and an index can be expressed with two bits, whereas 5 bits are required to represent the index. The larger the starting number (the lower the tree search), the better the efficiency, but the more often predictions will be incorrect. Therefore, depending on the hit rate of the vector predicted from idl and id2, for combinations that are difficult to predict, the initial vector is set in the upper stage of the tree search, and for combinations that can be accurately predicted, the initial vector is set in the lower stage of the tree search. Just set it.

As a means of determining the initial vector, (id□.

The initial vector number Pido + l d 1 may be stored in a table whose address is 1d2).

In the example shown in Figure 2, a 10-bit address provides 102
Four pieces of data will be prepared. When the number of tree search stages increases, approximate characteristics can be easily achieved by configuring addresses using only the upper bits of each index. For example, if Ldl and id2 are each 12 bits, then id.

This method uses the upper 6 bits of id2 and the upper 4 bits of id2.

Since the method described above can determine the initial vector without requiring any additional bits, the receiving side can similarly determine the initial vector, and there is no problem of mismatch between the transmitter and the receiver.

Furthermore, as a method for remediating the case where the prediction is incorrect, it is also possible to transmit information regarding the initial vector as additional information. For example, if the number of tree search stages is 8, the number of stages in which the initial vector exists can be specified using 3-bit information.

In addition, we have shown a method of limiting the initial vector Pidl, id2 to the search in the lower stage;
2, it is also possible to search for vectors in its vicinity. This is shown in FIG. That is, when predicted vector 10 is given, higher-order vector 5 is examined. If 10 is selected as a result, it is determined that the prediction was correct, and the search begins in the lower row, and in the example shown in the figure, vector 2
1, vector 44 is reached. Vector 5 test result 1
If 1 is selected, it is determined that the prediction is wrong, and the upper vector 2 is further examined. If 5 is selected as a result, 5
It is determined that there is an encoded vector under , and vector 1
Enter the search below 1.

In other words, by sequentially searching the upper vectors of the predicted vector, determining the highest point of the tree that contains the encoded vector, and then entering the search in the lower stages, the target vector is always reached even if the prediction is wrong. It can be so. In this method, the amount of code increases because it is passed to the upper stage, but it becomes possible to search from a wide range of vectors even when the prediction is incorrect. A second embodiment will be shown using FIGS. 1 and 5. This method is based on the following idea.

In FIG. 5, it is assumed that the index predicted by the preceding indexes id1 and ld2 is 010. Here, the index number is displayed in binary. At this time, vectors close to vector 010 have a high probability of occurrence,
It is considered that a far vector has a low probability of occurrence. According to the principle of vector quantization, it is known that distortion per code amount can be minimized by quantizing vectors with higher probability of occurrence more precisely and quantizing vectors with lower probability of occurrence more coarsely. Precise quantization corresponds to advancing the search to a deeper step, that is, to the lower stage, in tree search vector quantization. Coarse quantization corresponds to stopping the search at a shallow step, that is, at an upper stage. Further, the closeness between vectors is reflected in the distance between nodes in the tree structure, or the distance to reach a common node when tracing this structure upward. Thus, distant indices diverge at an early stage of the main structure, and near indices diverge at a later stage of the main structure. The accuracy of quantization can be adjusted by advancing the predicted index by a certain constant d steps from the point at which the index of the encoded vector is branched.

FIG. 5 shows the case d-2. For example, if the encoded vector is originally quantized to ooooo, it branches at the prediction index and node 0. Quantization is stopped at the stage where the process advances from node 0 to d-2 steps, ie, oo and ooo.

In the figure, the stopping point is indicated by a black circle. If d is set to a sufficiently large value, this corresponds to normal tree search vector quantization. Therefore, in this method, the characteristics can be changed depending on the number of stages in which the predicted vector is set and the number d of stages after prediction failure.

The above method can easily be realized as a function of the vector instruction circuit 6 in FIG.

In both of the above-mentioned embodiments, the tree search method is changed depending on the preceding index. This is a modification of the first method.

The first method aims to reduce the code amount of the index by selecting only vectors in the vicinity of the predicted index as encoding vector candidates, and the second method The main purpose of far vectors is to reduce the code amount of the index by terminating quantization midway through. Therefore,
The two methods can be used together.

For example, this can be achieved in the following way. Assume that the predicted index in FIG. 5 is 010. By comparing the distance between the input vector and vectors 010 and 01, it is determined whether the tree search should proceed upward or downward. That is, if vector 01 is closer to the input vector, it advances upward and sends out the code "1". If vector 010 is closer, it will proceed downward,
At the same time as sending the code “0”, the following search (0100
, 0101). If upward is selected, it is determined whether to search upward or downward again based on vectors 01 and 0. If vector 01 is closer to the input vector, a downward search is performed, a code "0" is sent out, the process moves to vector 011, and the subsequent search continues. As the search proceeds upward, the downward search is eventually selected and all states can be reached. Therefore, in this process, the end point determined by the predicted index of the tree search can be reached. The above operation can be easily realized by the vector instruction circuit 6 shown in FIG.

Note that although the above explanation has been given using a two-branch tree search as an example, it goes without saying that the same concept can be applied to an n-branch tree search.

(Effects of the Invention) As described above, according to the present invention, the range of tree search can be set using the preceding index, so that vector indexes can be efficiently encoded. Therefore, the correlation between indexes can be efficiently utilized with a simple configuration, and it can be applied to encoding systems for image signals and the like.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is an explanatory diagram showing an example of a tree search method showing the principle of the present invention,
FIG. 3 is a diagram explaining the present invention in detail, FIGS. 4 and 5 are diagrams explaining a modified method of the present invention, and FIGS. 6 and 7 are diagrams explaining the conventional technology of the predictive vector encoding method. It is. 1... Vector input terminal, 2... Distance calculation circuit, 3
... Minimum distance determination circuit, 4... Vector index output terminal, 5... Index delay circuit, 6...
Vector instruction circuit, 7...Codebook.

Claims (3)

[Claims] (1) In a vector quantization method that uses correlation between vector indices, there are a means for predicting the index of a coded vector based on a preceding vector index, a means for using the predicted vector as an initial vector for tree search, and an initial vector. 1. A predictive tree search vector quantization method, characterized in that the prediction index has means for continuing the tree search, and the hierarchy of the tree search is determined according to the prediction accuracy rate of the prediction index. (2) In a vector quantization method that uses correlation between vector indices, means for predicting the index of a coded vector using a preceding vector index, means for determining an end point of tree search using the predicted index, and an input A predictive tree search vector quantization method comprising means for sequentially obtaining indices for vectors by tree search, and employing the indices obtained up to the end point as the index for the input vector. (3) In a vector quantization method that utilizes the correlation between vector indices, there is a method for predicting the index of a coded vector based on the preceding vector index, and a method for performing upward tree search using the vector indicated by the predicted index as an initial vector. Predictive tree search vector quantization characterized in that it has a means for determining whether to perform a downward search, uses a code indicating whether it is an upward or downward search as index information, and continues normal tree search afterward if a downward search is selected. method.