Image classification method based on convolution neural network

Abstract

The invention relates to an image classification method based on a convolution neural network. The convolution neural network structure is improved. A feature filtering layer is added in a hidden layer. A large number of features are filtered by extracting the convolution network. A part of features with more noise are removed. The convolution network training efficiency is improved. The training time is reduced. The memory use requirement is reduced. A variety of training techniques are gathered, so that the convolution neural network training is converged a better solution to prevent training parameters from falling into a local minimum area. By reducing the number of network parameters, the over-fitting possibility of the neural network is reduced. The image classification accuracy and efficiency can be effectively improved.

Description

technical field [0001] The invention relates to a picture classification method based on a convolutional neural network, belonging to the technical field of image classification. Background technique [0002] In recent years, with the emergence of millions of labeled training sets and the emergence of GPU-based training algorithms, it is no longer a luxury to train complex convolutional network models. Convolutional neural networks have gradually developed and attracted widespread attention as an efficient recognition method. A large number of models based on convolutional neural networks have achieved good results in handwriting recognition and classification tests of the ImageNet library. Many papers have used convolutional neural networks to achieve good results in visual classification tasks. [0003] The basic structure of the convolutional neural network includes two layers, one is the feature extraction layer, the input of each neuron is connected to the local recept...

AI Technical Summary

Problems solved by technology

[0004] Due to the huge parameters of the convolutional neural network, in the field of face recognition, the number of network parameters ranges from several megabytes to hundreds of megabytes, which leads to long neural network training time, consumes a lot of memory space, and has relatively high requirements for the number of training samples. Prone to overfitting or convergence to a local minimum area

In addition, the characteristics obtained during the training and analysis of the network model are not easy to observe and difficult to analyze to adjust the network

Finally, the training method of the neural network is the gradient descent algorithm, which is not only susceptible to the influence of the initial value, but also cannot reflect the invariance ability of neurons

Images