JP2996979B2 - High efficiency coding method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency encoding method for a moving image such as a TV image, and converts one frame of A / D-converted image information into a discrete numerical value. This is an encoding method that performs complete processing in units of / 1 field, and can be realized with simple hardware, so not only image transmission but also high-efficiency encoding that is useful for miniaturization and high density of image recording devices such as digital VTRs It is about the method.

[Summary of the Invention] The present invention relates to a high-efficiency encoding method of a TV image signal, which processes image information in units of one frame per field, and one unit (1) of small pixel blocks including a plurality of pixels. ) Image, and for each small pixel block, a linear transformation suitable for human visual characteristics is adaptively selected, and the result of the linear transformation is taken as a vector, and a partial linear space with the vector is obtained. The component of the upward projection vector is obtained as an encoded output.

This method is completed in frames / fields, and
Since it can be mainly composed of linear adders and multipliers and does not require a large-capacity frame memory, a compact and lightweight real-time high-efficiency coding device for moving images with a large amount of information such as HDTV is realized. can do.

Further, since a linear transformation adapted to human visual characteristics is adaptively performed for each local portion of the image, a reproduced image with high image quality can be obtained.

[Prior Art] Conventionally, high efficiency coding methods for general images can be roughly classified into two types.

One is the TUSE, as seen in the MUSE system, ED-TV system, etc.
A moving image such as a V-image is to be processed, and encoding of real-time operation is a direct target. These adaptively perform filtering in the spatio-temporal frequency domain by subsampling or the like in accordance with the amount of motion between frames / fields, and use a large-scale hardware using large-capacity memory elements and arithmetic elements. The encoding is performed in real time by the ware (this is referred to as type-I).

On the other hand, as a method for processing a still image, a finite number of pixels that regard an image as a set of statistical data, pay attention to its statistical properties, and minimize an approximation error for many images. A block pattern is prepared, and for an input original image to be encoded, the best pattern is selected from the prepared block patterns, and the original image is reconstructed by combining the patterns. The encoded output is a sequence of numbers of the selected pattern (hereinafter, referred to as
Type II). This is a method generally referred to as vector quantization, and although coding with considerable efficiency is possible, since a long time is required for calculation, still images are targeted.

[Problems to be Solved by the Invention] The configuration of a small and lightweight moving image recording apparatus to be applied to a high-definition TV or the like is as follows: (1) Real-time operation (2) Encoding / decoding The device is compact and lightweight. (3) Editability in frame / field units is a requirement. In the above-mentioned type-I, aside from the decoding device, the coding device becomes large in principle and performs coding on the premise of field correlation and frame correlation, so that it is not suitable for editing at the code level. On the other hand, type-
In II, real-time operation is almost impossible at the current technology level.

SUMMARY OF THE INVENTION An object of the present invention is to provide a new encoding method which fulfills the requirements (1), (2) and (3), which are not satisfied by the conventional type-I and type-II.

[Means for Solving the Problems] According to the present invention, an original image that has been discretely sampled is represented by a vertical m horizontal n
Is divided into small blocks (m, n: integers), and the pixel values of the small blocks are arranged in a row to obtain an mn-dimensional vector x. Here, L k (k: integers) rows mn columns prepared in advance of the matrix Te ⁱ and a set of matrix Td ⁱ of mn row row k column ^{(Te i, Td i) (} i =
1, ..., L) for, Td ^i, of the L which has been subjected to Te ⁱ vector Td ⁱ
From Te ⁱ x (i = 1,..., L), (Te ⁱ , Td ⁱ ) giving the value closest to x is selected, and a set of k-dimensional vector Te ⁱ x and number i (Te ⁱ , i ), Wherein the small blocks are m = n and are inclined at 45 degrees.

Further, according to the present invention, in the high-efficiency encoding method, the L (Te ⁱ , Td ⁱ ) configuration is prepared by using a previously prepared L
For a mn-order square matrix A ⁱ (i = 1,..., L), a matrix Q of mn rows and k columns and a matrix P of k rows and mn columns have a norm of QPx with A
choose the one which is closest to the ⁱ x, the Te ⁱ P, characterized in that the Td ⁱ and Q.

Further, according to the present invention, in the high-efficiency encoding method, a square window matching the small block x of the original image is applied to the configuration of the L mn-order square matrices A ⁱ (i = 1,..., L). The result of applying L two-dimensional FIR filters H ⁱ (i = 1,..., L) prepared in advance to each pixel of x is given by A ⁱ x
It is characterized by being given as.

[Operation] According to the present invention, for example, image information in units of one frame / field is processed, and one unit (one image) is divided by a small pixel block including a plurality of pixels. A linear transformation suitable for human visual characteristics is adaptively selected for each block, and a component of a projection vector on a linear space as a result of performing the linear transformation is obtained as an encoded output.

[Example] Hereinafter, an example of the present invention will be described in detail with reference to the drawings.

An image F of one frame is a set of real values sampled in a square lattice of M × N. If F is divided by small blocks composed of pixels of length m × n, and each small block is x, x can be regarded as a vector belonging to the mn-dimensional real linear space Vmn.

At this time, the linear transformation from Vmn to the k-dimensional space Vk is Te, the encoded output of Te · x, and the linear transformation from Vk to Vmn is Td
And Td.Te.x as the decoded output.

Since Te is an operation for the small pixel block x, the Te is switched for each block. That is, a set of _L (Te, Td) sets is S _L , and When || · || is a certain norm, for each x, | Td ⁱ Te ⁱ x−x || is minimized (Td ⁱ , Te ⁱ ) to the pixel block x. An adaptive encoder / decoder is used.

 Next, the construction method of Te and Td will be described.

First, prepare L 3 × 3 two-dimensional FIR filters, And the impulse response of h ⁱ _(3,3) is h ⁱ _{(−1, −1)} ,..., H ⁱ _(1,1) as shown in FIG. At this time, m which is just applied to the position of a certain block x of the original image F
The image of × result of fitting the rectangular window of n, h ⁱ
Applying _(3,3) is a square matrix A of order mn defined below
It is equivalent to applying ⁱ to x.

Where a _{and k} are the elements in row k of column A ⁱ , The set of _L matrices obtained in this way is defined as ML.

A ⁱ is typically is a linear transformation to Vmn from Vmn, which was eigenvalue expansion, and ^{A i = U i Λ i mn} (U i) t (5). Here, U ⁱ is a unitary matrix having an eigenvector as a column vector, and Λ ⁱ mn is a diagonal matrix having an eigenvalue on a diagonal.

At this time, k eigenvalues having larger eigenvalues are selected from Λ ⁱ mn, and a diagonal matrix Λ ⁱ k is set to 0 for the others, and Te ⁱ and Td ⁱ are determined as follows.

At this time, in an mn-order square matrix W of any rank k, Wx
Comes closest to the W = T but the A ⁱ x on Euclidean norm
This is the case for d ⁱ Te ⁱ .

In this example, a square matrix A ^{i of} degree mn is expanded into eigenvalues and k rows
A matrix Te ⁱ of mn columns and a matrix Td ⁱ of mn rows and k columns were obtained. In a norm other than the Euclidean norm, a matrix P of k rows and mn columns and a mn row k such that QPx is closest to A ⁱ x Find the matrix Q of columns, It can also be.

In other words, for each small block of the original image, a spatial frequency component unique to the small block is extracted by A ⁱ , and the component vector is approximated by a projection vector (ie, Te ⁱ x) on a lower-order k-dimensional space Vk. and, the projection vector components, and transmission and recording as a code, to reproduce with Td ⁱ (Figure 2).

As a specific example, in the case of a (m, n) = 4, k = 4, L = 8, 8-bit quantized image, the information amount of the original image small block x is 8 × 4 ×
4 = 128 bits, code amount is 4 × 8 + 3 = 35 bits (however, Te
(the four coefficients of ^ix are represented by 8 bits), so that the compression ratio is 35/128 = 0.27.

If the pixel average value (DC component) is separated and extracted for each small block x and transmitted, the code amount is 5 × 8 + 3.
= 43 bits, the compression ratio is 0.34.

Incidentally, an example of a set H _L of the spatial filter on which to base in order to constitute the M _L in Figure 3. h ¹ and h ⁵ are pixel blocks with strong vertical correlation (images including high-frequency components in the horizontal direction), h ² and h ⁶ are pixel blocks with strong correlation in horizontal directions (images including high-frequency components in the vertical direction), h ^3, h (image containing high frequency components of the right oblique direction) ⁷ strong pixel block correlation of the left oblique direction, h ^4, h ⁸ is high-frequency component of the strong pixel block (left oblique direction correlation obliquely rightward ) (These filters simulate the response of simple cells present in the cerebral visual cortex).

FIG. 4 shows a configuration diagram of an encoding device that realizes this method.

The apparatus shows the case of L = 8 described above, where the input x is coded in eight ways by matrix and vector multipliers 1A to 1H, each of which is immediately associated with a corresponding matrix and vector multiplier 2A to 2A. Decoded by 2H. Multiplier encoding and decoding of eight from 2A~2H output Td ^{i-Te i} x is compared with the input x in the norm calculator 3A~3H two vectors (performs appropriate norm calculation), eight To obtain the norm output (scalar value) of. The outputs of these norm calculators 3A to 3H are input to the comparator 4 together with the outputs of the multipliers 1A to 1H, where the index i giving the minimum norm is detected, and (Te ⁱ , i) is output as an encoded output. I do.

Next, how to divide small blocks of mn, as shown in Fig. 5, a method of arranging them in a horizontal and vertical grid order, and in the case of m = n shown in Fig. 6, tilting by 45 ° to a rhombus There is a method of arranging them like offset sampling. The latter method is effective in reducing block distortion occurring at an edge of an image in an oblique direction.

In the present invention, as apparent from FIG. 4, the configuration of the encoder requires only a multiplier and an adder for performing a small-scale matrix operation and a norm calculation, and a size comparator for norm comparison. Basically, it is sufficient to have a memory for a field in the in-frame completion processing, and a memory for at least one line in the in-field completion. As described above, since the configuration of the encoding apparatus has a structure in which elements having simple functions are arranged in an orderly manner, it is easy to realize small and lightweight hardware suitable for LSI.

[Effects of the Invention] According to the present invention, a high-efficiency encoding method that operates in real time and can be edited in units of frames / fields can be realized, and the hardware for that purpose can be small and lightweight.

[Brief description of the drawings]

FIG. 1 is a diagram showing a general example of a 3 × 3 spatial filter, FIG. 2 is a diagram illustrating the principle of encoding / decoding of the present invention as a model on a linear space, and FIG. FIG. 4 is a diagram showing a specific example of a spatial filter, FIG. 4 is a configuration diagram of an encoding device according to the present invention, and FIGS. 5 and 6 are diagrams showing a specific example of the arrangement of m × n small blocks.

Claims (3)

(57) [Claims] 1. An original image discretely sampled is defined by a vertical m horizontal n
Is divided into small blocks (m, n: integers), and when the pixel values of the small blocks are arranged in a line as a mn-dimensional vector x, L k (k: integers) rows mn columns prepared in advance Of a matrix Te ⁱ and a matrix Td ⁱ of mn rows and k columns (Te ⁱ ,
^{Td i) (i = 1,} ..., for L), Td ^i, vector Td of L subjected to ^{Te i i Te i x (i} = 1, ..., L)
(Te ⁱ , Td ⁱ ) that gives the value closest to x is selected from the following, and a high-efficiency coding method using a set (Te ⁱ , i) of a k-dimensional vector Te ^ix and a number i as a coded output 5. The high-efficiency coding method according to claim 1, wherein the small blocks are m = n and are inclined at 45 degrees. 2. An original image which is discretely sampled is defined by a vertical m horizontal n
Is divided into small blocks (m, n: integers), and when the pixel values of the small blocks are arranged in a line as a mn-dimensional vector x, L k (k: integers) rows mn columns prepared in advance Of a matrix Te ⁱ and a matrix Td ⁱ of mn rows and k columns (Te ⁱ ,
^{Td i) (i = 1,} ..., for L), Td ^i, vector Td of L subjected to ^{Te i i Te i x (i} = 1, ..., L)
(Te ⁱ , Td ⁱ ) that gives the value closest to x is selected from the following, and a high-efficiency coding method using a set (Te ⁱ , i) of a k-dimensional vector Te ^ix and a number i as a coded output In the configuration of the L pieces of (Te ⁱ , Td ⁱ ), the mn rows k with respect to the L pieces mn-order square matrix A ⁱ (i = 1,..., L) prepared in advance.
Choose the one that is closest to the A ⁱ x in norm is QPx in matrix P matrices Q and k rows mn columns column, high efficiency coding for the Te ⁱ P, characterized in that the Td ⁱ and Q Method. 3. An original image which is discretely sampled is defined by a vertical m horizontal n
Is divided into small blocks (m, n: integers), and when the pixel values of the small blocks are arranged in a line as a mn-dimensional vector x, L k (k: integers) rows mn columns prepared in advance Of a matrix Te ⁱ and a matrix Td ⁱ of mn rows and k columns (Te ⁱ ,
^{Td i) (i = 1,} ..., for L), Td ^i, vector Td of L subjected to ^{Te i i Te i x (i} = 1, ..., L)
(Te ⁱ , Td ⁱ ) that gives the value closest to x is selected from the following, and a high-efficiency coding method using a set (Te ⁱ , i) of a k-dimensional vector Te ^ix and a number i as a coded output In the configuration of the L mn-order square matrices A ⁱ (i = 1,..., L), a result of applying a square window matching the small block x of the original image to L prepared in advance The result of applying a two-dimensional FIR filter H ⁱ (i = 1,..., L) to each pixel of x is A ⁱ x
A high efficiency coding method characterized by being given as: