JPH0313186A - Encoder - Google Patents

Abstract

PURPOSE:To obtain an encoder with high universality by providing a converter whose output characteristic is set at the distribution characteristic of Laplace distribution or Gaussian distribution at the front stage of a neural net input layer. CONSTITUTION:A neural net 5 based on a back propagation method comprising of three layers of an input layer 6, an intermediate layer 7, and an output layer 8, orthogonal transform encoder, the converter 9 whose output characteristic is set at the distribution characteristic of Laplace distribution or Gaussian distribution, and a reverse converter 10 consisting of a decoder with the characteristic opposite to that of the converter 9 are provided. And image information is supplied to the input layer after converting to the information with the distribution characteristic of Laplace distribution or Gaussian distribution with the converter 9. Thereby, it is possible to obtain the encoder with high universality using the neural net in which little change occurs in the information supplied to the input layer even when the image information is changed, and various kinds of compression encoding of an image can be performed based on a taught compression rule.

Description

[Detailed description of the invention] [Industrial application field] The present invention uses a neural network to capture still images.

The present invention relates to an encoding device that compresses image information of a moving image.

[Conventional technology]

BACKGROUND ART Conventionally, in the field of digital transmission of a single image and digital recording/reproduction, it has been considered to compress and decode image information using a neural network.

-t, rInformation Processing J tD Vol, 29, A
511 of 5 (Information Processing Society of Japan, published May 1988)
This page uses a neural network using the backglobe method, and generates two still images (digitized photos) using an encoding device consisting of an input layer of this net and an intermediate layer with a smaller number of neurons (units) than the input layer. ) is described to compress image information.

It is also described that the compressed information of the intermediate layer is decoded by a decoding device composed of the intermediate layer and the output layer to reproduce one image information.

In the above document, the net has a three-layer configuration with one intermediate layer, and ``the number of units in the input layer and output layer is 8 (two-dimensional) in 8 layers according to the division elements of image information; When the number of units is 16, the propagation teaching is repeated and the programming is set based on the ratio compression rule.
Compression with good reproducibility when unit quantization bits of the intermediate layer are 8 to 5.

It has been reported that decryption is possible.

In this case, if the image information as the original signal is 8-bit information of the divided element optical signal, the number of bits of the input layer (8x8x8=
512), the number of bits in the middle layer is (8 to 5) x 16
= 80 to 128 bits, so compression of about 1/64 to 1/4 can be performed by the encoding device.

[Problem to be solved by the invention]

In the case of the conventional encoding device, the image information is supplied to the input layer as is, and the 8EW3 rule is set to match the specific image taught. There are five problems with this method: it cannot be used to compress images different from a specific image, and its versatility is low.

An object of the present invention is to provide a highly versatile encoding device using a neural network.

[Means to solve the problem]

In order to achieve the above object, the encoding device of the present invention includes a converter whose output characteristics are set to Lagras distribution or Gaussian distribution before the input layer of the neural net, and converts image information into the distribution characteristics of the Lagras distribution or Gaussian distribution. A technical measure is taken to convert it into characteristic information and supply it to the input layer.

[For production]

In the case of the encoding device configured as described above, even if the image information changes, the output information of the converter is standardized to a Laplace distribution or a Gaussian distribution and does not change much.

One variety of images can be compressed based on the set compression rules.

〔Example〕

An embodiment will be described below with reference to FIG.

In FIG. 1, (1) is an input terminal for image information as an original signal, (2) is an encoding device, and (3) is a decoding device.

(4) is an output terminal for reproduced image information.

(5) is a well-known neural net using a pack propagation method having a three-layer structure of an input layer (6), an intermediate layer (7), and an output layer (8), and is formed using a computer or the like. (9) is a converter provided before the input layer (6), and an inverse converter provided after the a*rt output layer (8).

In addition, each layer (6) to (8) of neural net (5)
The circle indicates a neuron, that is, a unit.

The converter (9) is, for example, Hadamard transform, l1l
It consists of an orthogonal transform encoder that performs orthogonal transform such as lI scattered cosine transform, and the output characteristics are set to Laplace distribution or Gaussian distribution distribution characteristics.

Further, the inverse transformer is composed of a decoder having an inverse characteristic to that of the transformer (9).

Then, the image information (luminance information) of the still image or video that is digitized is 1, for example, 8 samples) v (pixel point)
It is divided into ×8 lines from the input terminal (1) to the converter (9
), and is converted into information on the distribution characteristics of the Lagras distribution or Gaussian distribution by two-dimensional orthogonal transformation in this converter (9).

Furthermore, the output information of the transformer (9) is fed to the input layer (6) of the neural net (5).

In this net (5), the input layer (6) and output layer (8) are composed of 8x8 units, and the middle layer (7) is composed of 16 units, depending on the division element of input information. There is.

Furthermore, during learning before use, the output information of the converter (9) based on the image information of a specific image or the Laplace distribution input layer (
6) Ill rules for the intermediate layer (7) network and screen playback rules for the intermediate layer (7) and output layer (8) networks are set by programming.

At this time, the compression rule 1 screen reproduction rule is set according to information on the Lagras distribution or Gaussian distribution depending on the output characteristics of the converter (9).

Then, the information supplied to the input layer (6) is compressed according to the Ei rule, and the compressed information is sent to the encoding device (2).
is transmitted to the decoding device (3) as output information.

On the other hand, in the decoding device (3), the compressed information of the intermediate layer (7) is decoded and reproduced according to the screen reproduction rule, and the reproduced information is supplied from the output layer (8) to the inverse transformer αO. .

At this time, like the information supplied to the input layer (6), the reproduced information becomes information with distribution characteristics of Laplace distribution or Gaussian distribution.

Furthermore, the reproduced information is returned to the same information as the image information of the input terminal (1) by the inverse transformation of the inverse transformer QO, and is outputted from the output terminal (4).

Then, to make sure that the probability distribution of the information supplied to the neural net (5) does not change significantly no matter what kind of image image information is supplied to the input terminal (1), the taught compression rule 1-screen playback is performed. Appropriate compression and decoding can be performed based on rules.

By the way, the DC component of the 2-image information has a Laplace distribution.

This direct current component does not follow the Gaussian distribution and passes through the neural net (5), resulting in a decrease in compression ratio.

When a reduction in compression ratio based on the DC component becomes a problem, the converter (9) outputs the DC component information independently, and the DC component information is expressed as shown in the dashed line in the figure. In this case, it is sufficient to supply the signal to the inverse transformer αQ without going through the neural net (5).

Further, in the embodiment described above, the transformer (9) and the inverse transformer αq were formed by an orthogonal transform encoder and its decoder.

It may be formed by a predictive encoder and its decoder, and in this case, the configurations of the transformer (9) and the inverse transformer 00 are simplified.

Furthermore, when compressing and decoding image information of a moving image, the image information at the input terminal (1) may be used as intra-field, inter-field, or inter-frame difference information, and the converter (9) is used as a normal encoder. and one or more encoders for the difference information, and each encoder may be adaptively switched and used.

Furthermore, the configuration of the neural net is not limited to the embodiment. For example, a neural net may have an input layer (
6), between the intermediate layer (7) and between the intermediate layer (7) and the output layer (8).
) can also be applied to the case where there is an auxiliary intermediate layer between them.

〔Effect of the invention〕

Since the present invention is configured as described above, it achieves the effects described below.

A converter is provided before the input layer of the neural net, and this converter converts one image information into information on the distribution characteristics of Laplace distribution or Gaussian distribution and supplies it to the input layer, so that one image information is changed. It is also possible to provide a highly versatile encoding device using a neural network, which has little change in the information supplied to the input layer, can compress and encode various images based on the taught compression rules, and can.

BRIEF DESCRIPTION OF THE DRAWINGS The drawing is a block diagram of one embodiment of the encoding device of the present invention. <5) New full net, (6) Input layer, (7) Middle layer, (9) Converter.

Claims (1)

[Claims] (1) In an encoding device comprising an input layer of a neural net and an intermediate layer having a smaller number of neurons than the input layer, and compressing image information and outputting it from the intermediate layer, an output characteristic is provided at a stage before the input layer. An encoding device comprising: a converter set to a distribution characteristic of a Laplace distribution or a Gaussian distribution, and converting the image information into information having the distribution characteristic and supplying the information to the input layer.