JPH02186835A - Signal encoder - Google Patents

Abstract

PURPOSE:To reduce a quantization error and the quantity of information by using a random access variable code book also with an on-going fixed code book, and cascade-connecting two vector quantizers with different block size. CONSTITUTION:The encoder is provided with a vector quantization encoder 1 which performs the vector quantization of an inputted digital signal by using the fixed code book 2 on one side, and a vector quantization encoder 4 which performs the vector quantization of a digital signal transmitted as being unprocessed by using a variable code book 3 when distortion is larger than an allowable value on the other side. And those two vector quantization encoders 1 and 4 are cascade-connected. Thereby, it is possible to obtain a signal encoder by the vector quantization by which the quantization error for a specific input vector can sufficiently be reduced, and simultaneously, the quantity of information can be controlled to a minimum.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a signal encoding device, and particularly to a signal encoding device using a vector quantizer that encodes digital signals with high efficiency.

[Prior art] Figure 6 shows, for example, the Technical Report of the Institute of Electronics, Information and Communication Engineers, IT85.
61 is a block diagram showing the configuration of the conventional Bertor quantized signal encoding device shown in FIG. A fixed codebook (2) that stores a plurality of the quantized representative vectors (9) is connected to a vector quantization encoder (20) that outputs the index as encoded data (11), and outputs the index as encoded data (11). (11) A fixed codebook (
2) is connected.

Next, the operation will be explained. k-dimensional human vector (8
)X-[X+, X2,...X,], and the quantization representative vector Yi-[Y,, Y2... ..., yh] set to [Y]
-[YI, Y2, ..., YN] (9)
shall be. The vector quantization encoder (20) determines the quantization representative vector Yi that minimizes the distortion d+ with respect to the input vector X(8) defined by the following equation. In addition,
Here, if di is defined as square strain, di-IIX-Yi12-Σ (Xi Y++)
It becomes 2j-1. Here, dt<df, l≠i.

As encoded data (11), the quantized representative vector Y
The index i that separates i by 1 is output and transmitted.

In vector quantization decoding W (21), the quantization representative vector Yi corresponding to the index l given by the input encoded data (11) is read from the codebook (2), and the decoded vector X (19) is read out from the codebook (2). is played.

That is, −Yi becomes.

As described above, the conventional vector quantization signal encoding device registers the quantization representative vector (9) between prefectures in the fixed code book (2).

[Problems to be Solved by the Invention] Since the conventional vector quantization signal encoding device is configured as described above, when the number of dimensions of the vector is high, the H-eye side stored in advance in the codebook is It is difficult to completely optimize the quantization representative vector for all information source input vectors, and there is a problem in that it is necessary to reduce excessive quantization errors that occur in unique input vectors.

This invention was made to solve the above problems, and is a signal code using vector quantization that can sufficiently reduce the quantization error even for unique human vectors, and at the same time minimize the amount of information generated. The purpose is to obtain a conversion device.

[Means for Solving the Problems] A signal encoding device according to the present invention includes a vector quantization encoder that vector quantizes an input digital signal using a fixed codebook, and a vector quantization encoder that vector quantizes an input digital signal using a fixed codebook. , the other vector quantization encoder vector quantizes the digital signal transmitted unprocessed using a variable codebook, and the two vector quantization encoders This is a continuous connection of devices.

[What is the effect of the signal generated by the vector quantizer in this invention? ] In the encoding device, when the minimum distortion is larger than the set distortion threshold, the input vector is not encoded at all by the preceding stage encoder and is transmitted to the next stage vector quantization encoder. Here, further distortion measurement is performed, and if it is larger than the threshold value, the human vector will be output to the decoder side as the original signal, or if it is smaller than the distortion,
The index number of the representative vector that gives the minimum distortion is output to perform efficient encoding while reducing the amount of information generated.

In the variable codebook, the codebook area for each feature m of the representative vector is H, and representative vectors having the same feature are arranged in order of frequency of appearance, so that a search can be performed with high frequency and with high accuracy.

By using the other vector quantization encoder connected to the variable codebook as the first stage and one vector quantization encoder as the second stage, it is possible to create a learning encoder that reduces the amount of information generated. .

[Example] An embodiment of the present invention will be described below.

In FIG. 1, one vector quantization encoder (1)
A fixed codebook (2) configured similarly to the conventional one is connected to the i'iJ variable codebook, which can be written and read at any time. )It is connected to the.

Next, the operation will be explained. The input vector
(12), and the index (12) is input to the selector (5).

In addition, in FIG. 1, one vector quantization encoder (1)
The quantization block size was 4×4.

Here, in one vector quantization encoder (1), the minimum distortion di and an arbitrarily set distortion threshold 1iiT
A comparison is made with i, and the following two processes are performed according to the comparison result. Further, a select signal (17) for identifying the process is supplied to the selector (5) and transmitted.

The encoding algorithm will be explained below according to the flowchart of FIG.

Process I: When di≦Ti, set the select signal (17) to “O” and at the same time,
Index 1 (12) is output to selector (5).

Process ■: When di>Ti, set the select signal (17) to “1” and at the same time,
Input vector X to the other vector quantization encoder (4)
(8) is transmitted.

Next, in the case of the above process (■), two types of processing are performed successively by the other vector quantization encoder (4).
It is shown below. However, the minimum distortion with respect to the input vector was set to dj, and the quantization block size of the other vector quantization encoder (4) was set to 2×2.

Process m: When dj≦Tj, the select signal (16) is set to O'', and at the same time, the index j (13) is transmitted to the selector (5).

Processing TV: When dj>Tj, at the same time as setting the select signal (16) to "1",
The input vector X'' (18) is transmitted to the selector (5), which inputs the input vector
Write in 3).

Here, operations for reading and writing quantization vectors to the variable codebook (3) are performed as shown in FIG.

First, the feature quantity is divided into X pieces, and the quantization vectors that take the value of each feature quantity in X and their index numbers are registered in order of appearance frequency. For the human vector, X feature quantities are immediately extracted, and a quantized vector that matches the values of all the X special quantities is selected. Variable code book (3)
If there is no match for the quantized vector in
Follow the instructions below.

■) When all X features are different.

The input vector sequence with the i-th feature value α(i) is
α(i)), and the quantized vector sequence in the variable codebook (3) is lj (α(i)), then dj−Σ Iy
j(α(i)) i Utai 2 -XCα(i))I...Choose the quantization vector yj that minimizes the value of (A), delete it, and write the human vector x instead. .

11) When two or more feature quantities match.

From the set of quantized vectors 7 with matching feature values, (
A) Calculate the formula, select y/j and rewrite in the same way.

In the manner described above, the quantization vectors in the variable codebook are read, written, and rewritten.

On the decoder side, after inputting the encoded data (11) output from the encoder side to the selector (5), a select signal (1
6) and (17), it is selected which one of the vector quantization decoder (6) and the other vector/child decoder (7) will be used to perform decoding. A flowchart in this case is shown in FIG. If one vector quantization decoder (6) is selected, index 1 (12
) according to the fixed codebook (2) decoded vector y
, (15) are extracted and output as an output vector. Furthermore, when the other vector quantization next encoder (7) is selected, the decoded vector y+ (14) is extracted from the variable codebook (3) according to the index j (13), or the input vector (18) will be output as is as the decoded vector (14). However, in the latter case, the input vector X' (18) will be registered in the variable codebook (3) at the same time.

The block size of one vector quantization encoder (1) and one vector quantization encoder (6) is 4.
X4, whereas the other vector m-child encoding B(
4) and the other vector quantization decoder (7) are composed of four 2×2 blocks. Therefore, when distortion calculation is performed in one vector quantization encoder (1) and the minimum distortion di becomes a value equal to or greater than the threshold value T1, the original image of the input vector ) is divided into four 2×2 blocks, and the minimum distortion calculation is performed for each in the other vector quantization encoder (4). Therefore, among the four blocks, if the matching with the variable codebook (3) is successful, its index j (13) will be the encoded output, but if the distortion is large and the matching is not successful, the input vector (2×2 size)X' (18) is output as is. At this time, it is necessary to write which blocks are encoded and which blocks are not encoded as header information of the select signal (16).

In the above embodiment, two vector quantization encoders with different block sizes are connected in cascade, one hector quantization encoder (1) and one vector quantization decoder (6).
A fixed codebook (2) is connected to the other vector quantization encoder (4) and a variable codebook (3) to the other vector quantization decoder (7). , and the variable codebook for the former,
And if you connect a fixed codebook to the latter (
The same effect as in FIG. 1 can also be obtained using the method shown in FIG.

[Effects of the Invention] As explained above, the present invention transmits an input vector extracted based on the minimum distortion during vector quantization, and sequentially stores the input vector as a new quantization representative vector, which can be written and read at any time. By using a variable codebook and a conventional fixed codebook together, and cascading two vector quantizers with different block sizes, quantization errors are reduced and the amount of information generated is significantly reduced. It turns out.

Also, by setting the other vector quantization encoder connected to the variable codebook, the distortion threshold T of one vector quantization encoder connected to the fixed codebook
The value of i can be set small, which increases the SN ratio and can be expected to improve image quality.

Furthermore, since each quantization vector is arranged so that the structure of the variable coatbook can be classified by feature quantity, the number of matching operations can be reduced compared to the conventional method by detecting the feature quantity of the -degree human force vector. , which can speed up the search.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of a signal encoding device according to the present invention, FIG. 2 is a flowchart of the device according to the present invention, and FIG. 3 is a vector read/write of a variable codebook according to the present invention. FIG. 4 is a flowchart 1 of the decoder operation, FIG. 5 is a block diagram showing an embodiment of the present invention, and FIG. 6 is a block diagram of a conventional device. In the figure, (1) is one vector quantization encoder,
(2) is a fixed codebook, (3) is a variable codebook, (4) is the other vector quantization 7-block encoder, (5) is the selector, (6) is one vector quantization decoder, ( 7
) is the other vector quantization decoder, (8) is the human vector X, (9) is the quantization representative vector Yi, (10) is the quantization representative vector Y, (11) is the encoded data, (1
2) is index i, (13) is index j, (
14) is the decoded vector (yj or x), (15
) is the decoded vector y i, (16) is the select signal 2, (17) is the select signal 1, (18) is the human vector X - (19) is the decoded vector, (20) is the vector quantization encoder, ( 21) is a hector quantization complexer. In addition, in the figures, the same reference numerals indicate the same or equivalent parts. Agent: Patent attorney Masuo Oiwa (2 others) Flowchart of decoder Figure 4 Procedural amendment (voluntary) 8 Human power vector 9, Quantization representative vector 11: Encoded data 19, Decoding vector Block diagram of conventional device Figure 6

Claims (1)

[Claims] One vector quantization encoder connects a fixed codebook consisting of a plurality of quantized representative vectors, and one vector quantization encoder has a region for each characteristic value of the representative vectors, and the representative vectors having the same characteristics are arranged in order of frequency of appearance for writing and reading at any time. A signal encoding device configured by continuously connecting the other vector quantization encoder to which a possible variable codebook is connected.