High-precision activation function for CNN model image classification task

Abstract

The invention discloses a high-precision activation function for a CNN (Convolutional Neural Network) model image classification task. A right shift linear correction unit (RsReLU) comprises an RsReLU function mathematical model, an RsReLU function gradient function mathematical model, an RsReLU function image and an RsReLU function gradient function image; according to the RsReLU function, negative value data which is forcibly converted into zero by the ReLU originally are changed into smaller positive values to participate in feature extraction of the convolutional neural network by introducing an attention mechanism idea, so that the problems of neuronal necrosis and lower classification accuracy during image classification in the traditional ReLU are effectively solved. The method is used in a convolutional neural network image classification model, and the effect of improving the model classification precision is effectively achieved.

Description

technical field [0001] The invention belongs to the field of deep learning, and relates to a high-precision activation function used for a CNN model image classification task. Background technique [0002] Convolutional Neural Network (CNN) is a kind of neural network specially used to process data with similar network structure. Since the convolutional neural network was proposed, it has been widely used in the field of computer vision. As well as object detection and other aspects have achieved excellent results. The basic unit of convolutional neural network is usually composed of feature extraction, nonlinear activation function and downsampling, and then the basic unit constitutes a trainable multi-layer network structure. In these constituent units, the activation function can retain the features extracted by the convolution layer and remove redundant data, and map the feature data through nonlinear functions. Its existence is decisive for the ability of the network ...

AI Technical Summary

Problems solved by technology

[0005] The embodiment of the present invention provides a high-precision activation function for CNN model image classification tasks, which has solved the problems of low classification accuracy and poor model effect existing in the current activation function

Images