Research On Image Recognition Methods Based On Evolutionary Optimized Convolutional Neural Network

Convolutional Neural Network(CNN)is one of the important models in the field of deep learning.It has been widely used in scenes such as image and video processing in recent years.In image recognition,CNN's invariance to image features and good feature extraction capabilities make it possible to obtain better recognition accuracy without tedious preprocessing.In practical applications,the recognition accuracy is mainly affected by the connecting structure and parameters of model,so it is of great significance to study the structure and parameter optimization methods of convolutional neural networks.In this thesis,the basic principles and structure of typical CNNs,existing problems and related improvement methods and results at home and abroad are analyzed and summarized.On the one hand,typical CNNs have many redundant connections between the input and output feature channels of the convolutional layer,which is not only detrimental to the feature extraction of the convolutional layer,but also increases the complexity of the model.The existing improved connection structure needs to be designed according to experience and is not universal.On the other hand,the back propagation algorithm needs to be set up many hyperparameters,resulting in incomplete training of model parameters.Therefore,the following work has been carried out:1.A method for optimizing CNN feature channel connecting structure using adaptive genetic algorithm.First,the CNN was pretrained and the parameters were saved.Then the adaptive genetic algorithm was used to optimize the feature channel connecting structure of the convolutional layer of pretrained CNN.The backpropagation algorithm finetunes the parameters of the CNN with optimal connection structure to adapt the parameters to the new connection structure.The optimized CNN is used for the image recognition of FashionMNIST dataset and CIFAR-10 dataset.The experimental results show that the optimal connecting structure searched by this method can significantly improve the image recognition effect of CNN and speed up the training of the model.2.A CNN parameter optimization method based on differential evolution algorithm is proposed.After the CNN parameters are trained by the back propagation algorithm,the differential evolution algorithm is used to make the model parameters further approximate the optimal solution,and the model parameters are learned more thoroughly to achieve a better recognition effect than when using the back propagation algorithm alone.This method is applied to the self-collection handwritten Chinese character data set based on HCL2000.Experiments show that this method is effective when using different backpropagation optimizers.The highest recognition accuracy rates for simple characters,complex characters and similar characters reached 98.2%,99.6% and 97.2% respectively.