Convolutional neural network structure simplification and image classification method based on evolutionary strategy

Abstract

The invention provides a convolutional neural network structure simplification and image classification method based on an evolutionary strategy. The structure between a convolution model feature extraction part and a full connection layer is simplified. The method comprises the following steps: firstly, pre-training a convolution model to fix a model weight, and performing feature extraction on an image data set by using a convolution part; then, randomly generating a plurality of binary sequences to serve as an initial population, wherein each binary sequence corresponds to a structure between a feature extraction part and a full connection layer; then, under each structure of the current population, using the pre-trained weight to classify the feature vectors of the test set, using theclassification accuracy as the fitness of the corresponding structure, obtaining a new-generation structure population through crossover, variation and selection operations, and continuously iteratingto be equal to the number of iterations; and finally, obtaining a structure between the simplified feature extraction part and the full connection layer, and performing fine adjustment on the pre-training weight by applying the structure.

Description

technical field [0001] The invention belongs to the technical field of image feature extraction and classification methods, in particular to a method for automatically searching for an optimal structure between a feature extraction part and a fully connected layer during image feature extraction and classification, and reducing the number of parameters. Background technique [0002] Convolutional neural networks (CNNs) have already occupied a place in the field of image recognition. Because of its characteristics of weight sharing, local receptive field, downsampling, and translation invariance to image features, it has good feature extraction capabilities. It can effectively reduce a large number of preprocessing processes in traditional image recognition methods, and can achieve good recognition accuracy. [0003] At present, a lot of research has been done on the convolutional neural network. Since AlexNet in 2012, a batch of classic models have been produced, such as VGG...

AI Technical Summary

Problems solved by technology

On the other hand, there are also many model clipping algorithms that can sparse the model parameters without significantly reducing the recognition accuracy to reduce the number of parameters. However, the process of model clipping and model training is carried out simultaneously, which aggravates the Computational pressure during training

Images