KR970019636A - Improved Neural Network Data Compressor and Data Processing Method Using the Same - Google Patents

Abstract

The present invention provides accurate data restoration at the output of the decoder by adding additional enhancement nodes using instantaneous information to the encoder stage or decoder stage for accurate data restoration while maintaining the same number of hidden layer node stages as conventional neural network data compression. The present invention relates to an improved neural network data compressor and a data processing method using the same.

In the three-layer neural network data compressor including the encoder 1, the hidden layer 30, and the decoder 20, an additional enhancement node 5 using instantaneous information at the input terminal or the decoder input terminal of the encoder unit is further included. It is characterized by optimizing the weight between the encoder stage and the hidden layer node stage.

Description

Improved Neural Network Data Compressor and Data Processing Method Using the Same

Since this is an open matter, no full text was included.

5 is a compressor configuration diagram in which an additional enhancement node using instantaneous information is added to an encoder input terminal of a neural network in accordance with the present invention.

FIG. 6 is a schematic diagram showing a simplified form of FIG.

7 is a block diagram of a decoder portion of a neural network data compressor using instant information of data according to the present invention.

Claims (10)

In the 30-layer neural network data compressor composed of an encoder 1, a hidden layer node end 30 and a decoder 20, further comprising an additional enhancement node 5 using instantaneous information at the input end of the encoder portion, An improved neural network data compressor characterized by optimizing weights between the encoder input stage and the hidden layer node stage. 2. The improved neural network data compressor of claim 1, wherein the instantaneous information is a vector of information obtained from a previous frame group. In a three-layer neural network data compressor comprising an encoder (1), a hidden layer node stage (30), and a decoder (20), the decoder input stage further includes an additional enhancement node (25) using instantaneous information. An improved neural network data compressor characterized by optimizing weights between hidden layer nodes. 4. The improved neural network data compressor of claim 3, wherein the instantaneous information is a vector of information obtained from a previous frame group. In a three-layer neural network data compressor comprising an encoder (1), a hidden layer node stage (30), and a decoder (20), the encoder further comprises an additional enhancement node (5) using instantaneous information at the input of the encoder section. An improved neural network characterized by optimizing the weights between the input and the hidden layer nodes, and further comprising an additional enhancement node 25 using the instantaneous information of the decoder inputs. Network Data Compressor. 6. The improved neural network data compressor according to claim 5, wherein the instantaneous information is a vector of information obtained from a previous frame group. In the three-layer neural network data compressor including the encoder 1, the hidden layer node stage 30, and the decoder 20, adding an additional enhancement node 5 using the instantaneous information to an input terminal of the encoder stage; ; By supplying the basic video vector X and the instantaneous information vector U to the basic video node and the instantaneous information node 5, the decoder output terminal 21 reproduces the X vector as much as possible (X = Obtain encoder weight (3) and decoder weight (23), and store pairs (X, U, Y) of hidden layer node-end output vectors Y corresponding to each basic input vector X and instant vector U. An improved neural network data processing method comprising the first learning step. In the three-layer neural network data compressor comprising the encoder 1, the hidden layer node stage 30 and the decoder 20, adding an additional enhancement node 25 using the instantaneous information to the input terminal of the decoder stage; Input the Y vector, the output value of the hidden layer node, as the decoder stage input value, set the instantaneous information U stored at the encoder stage as the input vector V of the instantaneous information node 25, and use a pair with the output target value X. Re-learn the weights of the decoder section An improved neural network data processing method comprising the second learning step of restoring the value of to be closer to the value of X. A thirty-layer neural network data compressor comprising an encoder (1), a hidden layer node stage (30), and a decoder (20), comprising: adding an additional enhancement node (5) using instantaneous information to an input stage of the encoder stage; The output part of the encoder is quantized by vector to be divided into one vector, and only its representative code is transmitted. Improved neural network data processing method, characterized in that. A three-layer neural network data compressor comprising an encoder (1), a hidden layer node stage (30) and a decoder (20), comprising: adding an additional enhancement node (5) using instantaneous information to an input terminal of the decoder stage; The output part of the encoder is quantized by vector to be divided into several vectors, and only the representative code is transmitted.The decoder restores the transmitted representative code to an estimated vector using a pre-learned codebook and then restores the data input to the encoder Improved neural network data processing method, characterized in that. ※ Note: The disclosure is based on the initial application.