JPH077460B2 - Vector quantizer - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantizer for high efficiency coding of digital images.

2. Description of the Related Art Conventionally, orthogonal transformation such as Hadamard transformation is often used as high-efficiency image coding. In a general image, since there is a strong correlation between adjacent pixels, it is possible to reduce the transmission amount by performing orthogonal transformation. FIG. 4 shows the probability density distribution of sample values of pixels and one of the orthogonal components (Hi (i = 1,2, ...,
n)) represents the probability density distribution. Since the distribution of the orthogonal components is concentrated at the origin as shown in this figure, high-efficiency coding with less distortion can be achieved by performing fine quantization with the quantization representative value around the origin as shown in FIG. Done.

Problems to be Solved by the Invention However, the method of independently quantizing each orthogonal component as described above does not use the correlation between orthogonal components,
It is not always efficient quantization. FIG. 6 is a two-dimensional view of the distribution range of the two orthogonal components Hi and Hj. In this way, there is often a negative correlation between the orthogonal components, in which one value is large and the other value is small. The ◯ mark in FIG. 6 is a two-dimensional representation of the quantized representative values of Hi and Hj as shown in FIG. Since Hi and Hj independently have 8 (3 bits) quantized representative values, both have 8 × 8 = 64 (6 bits) quantized representative values.
However, in FIG. 6, only 35 of the 64 quantized representative values are included in the existence range of Hi and Hj, which is a very useless quantization. Further, in the conventional method, the number of quantized representative values is always limited to a power of 2 (bits), so that it is not possible to give quantized representative values symmetrical about the origin as shown in FIG. Therefore, even in one-dimensional quantization, optimum quantization becomes difficult.

The present invention has been made in view of the above points, and an object thereof is to provide a vector quantization device that enables efficient quantization.

Means for Solving the Problem The present invention divides image information into blocks of n pixels each, and performs an n-th order orthogonal transformation for each block,
First quantizing means for quantizing the n orthogonal components obtained by the orthogonal transforming means by a quantizer having a specific quantizing characteristic for each component; and n quantums obtained by the first quantizing means. The converted value is mi And a second quantizing means for performing mi-dimensional vector quantization for each of the l sets, each of which is divided into l sets, and 1 block per block obtained by the second quantizing means. The vector quantizer is characterized by outputting a quantized value.

Operation According to the present invention, the number of bits of each orthogonal component is reduced by the first quantization and the vector quantization is performed between the orthogonal components by the second quantization, so that the vector quantization can be performed relatively easily. Become. Further, by using vector quantization, efficient quantization can be performed as compared with conventional quantization, and image quality and compression rate can be improved.

Embodiment FIG. 1 shows a block diagram of an embodiment of the present invention. 1 in FIG. 1 is an input part of the apparatus, 2 is an nth-order orthogonal transformer, 3
-8 is the first quantizer, 9-11 is the vector quantizer, and 12 is the output part of this device.

In the apparatus of FIG. 1, first, n sample values input from the input section 1 are orthogonally transformed by the nth-order orthogonal transformer 2. The n orthogonal components obtained by the nth orthogonal transformer 2 are input to the quantizers 3 to 8 and are first quantized. Next, the first quantized orthogonal components are collected by mi (1 = 1,2, ..., L) units and input to the mi-dimensional vector quantizers 9 to 11 and second quantized. Then, l quantized values obtained from the second quantizers 9 to 11 are output to the output section 12.

As described above, l vector quantized values are obtained from n sample values. Moreover, each quantizer can be easily implemented by using a memory.

Here, a more specific block diagram of the present invention is shown in FIG. FIG. 2 is an example applied to high-efficiency coding using the 8th-order Hadamard transform.

In FIG. 2, 13 is an input part of the apparatus, 14 is an 8th order Hadamard transformer, 15 to 22 are ROMs for executing the first quantization, and 23 to 2
RO is for performing the second quantization (vector quantization)
M and 26 are output parts of this device.

The 8 sampled values of 8 bits input from the input part 13 are 8
Next, it is input to the Hadamard converter 14. Next, the eight 11-bit orthogonal components obtained by the eighth-order Hadamard transformer 14 are input to the ROMs 15 to 22 and are first quantized. Of the orthogonal components compressed to 7 bits to 3 bits by the first quantization, the components other than the output of the ROM 15 are input to the ROMs 23 to 25 and secondly quantized (vector quantization). By the above two-stage quantization, a total of four quantized values of 32 bits are obtained and output to the output section 26.

As described above, the apparatus shown in FIG. 2 can compress information of 64 bits per block into 32 bits. In the second quantization, orthogonal components having a relatively high correlation are collectively vector-quantized, and orthogonal components having a small correlation with other orthogonal components such as the output of the ROM 15 are output as they are.

Next, a concrete example of the second quantization is shown in FIG. FIG. 3 shows the distribution range of two orthogonal components having the same two-dimensional distribution as FIG. 6 of the conventional example. The broken line in FIG. 3 indicates the position of the quantization representative value of the first quantization according to the present invention. In this figure, both Hi and Hj have 13 (4 bits) quantized representative values, and 13 × 13 = 169 (8 bits) quantized representative values in total. Of these representative values, 64 (6 bits) quantized representative values represented by a circle are selected as the representative values of the second quantization (vector quantization). As a result, it is clear that the quantization of FIG. 3 enables the quantization to be very thin and have a small distortion, even though the same 6 bits (64) output as the quantization of FIG. 4 is output. . As described above, the quantization of the present invention makes it possible to greatly improve the image quality and the compression rate. The second quantization in FIG. 3 is 2 ⁸ × 6 = 1536bits, where the input (address) is 4 + 4 = 8bits and the output is 6bits.
It can be implemented as a device with a ROM capacity of, and can be realized very easily. Similarly, the total ROM capacity required for the second quantization (vector quantization) of the device shown in FIG. 2 is 2 ^{(4 + 4)} × 6.
+2 ^{(5 + 5)} × 9 + 2 ^{(4 + 4 + 3)} × 10 = 31232bit
s, which is easily realizable.

Although the Hadamard transform is used in the present embodiment, the present invention can be applied to all other orthogonal transforms such as the cosine transform. Further, although the vector quantization of the present invention is configured by using a memory, it can be realized by a logic circuit. Further, in order to further improve the image quality, it is possible to omit the first quantization and directly perform vector quantization.

EFFECTS OF THE INVENTION As described above, according to the present invention, since vector quantization is performed in the second quantization, it is possible to use even a quantized representative value that has not been used in the conventional example, so that the image quality and compression rate are greatly improved. Is possible. Further, in the conventional quantization, the number of quantized representative values for each orthogonal component was limited to a power of 2, but in the present invention, an arbitrary number can be assigned, which also improves the image quality.

Further, since the vector quantizer of the present invention has a small number of input bits due to the first quantization, it can be easily realized by using a memory, and its practical effect is large.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a block diagram of a more specific embodiment of the present invention, FIG. 3 is an explanatory diagram of vector quantization used in the present invention, and FIG. 4 is orthogonal. FIG. 5 is an explanatory diagram of conversion, FIG. 5 is an explanatory diagram of quantization in the conventional example, and FIG. 6 is a two-dimensional distribution diagram of quantized representative values in the conventional example. 2 ... Orthogonal transformer, 3 ... First quantizer, 9 ... Vector quantizer, 14 ... Hadamard transformer, 23 ... ROM.

Claims (3)

[Claims] 1. An orthogonal transformation means for dividing image information into n (n is an integer) blocks of pixels and performing an nth-order orthogonal transformation for each block, and n obtained by the orthogonal transformation means.
A first quantizing means for quantizing each of the orthogonal components by a quantizer having a specific quantizing characteristic, and n quantizing values obtained by the first quantizing means are mi. And a second quantizing means for performing mi-dimensional vector quantization for each of the l sets, each of which is divided into l sets, and 1 block per block obtained by the second quantizing means. A vector quantizer which outputs a quantized value. 2. A vector quantization apparatus according to claim 1, further comprising Hadamard transform means as the orthogonal transform means. 3. A value obtained by concatenating mi quantized values obtained by the first quantizing means is used as an address, and the mi quantized values are mi-dimensional vector quantized at the location indicated by this address. The vector quantization apparatus according to claim 1, wherein mi-dimensional vector quantization is performed using a memory that stores the quantization value of the result.