Research And Application Of Convolutional Neural Networks Based On CMA-ES Algorithm

The rapid development of Internet technology,mobile media applications such as WeChat and Weibo have also emerged,and Internet users can upload or view images quickly and easily.However,in real life,a large number of unmarked images are difficult to search and process.At the same time,the general image classification method has been unsatisfactory in the classification and recognition of these images.Especially when dealing with some complex natural images,these methods are insufficient.The emergence of deep learning,which provides a new powerful image classification and recognition methods.Convolutional Neural Network(CNN)is a key research content of deep learning,and it is also the most widely used deep learning method in image processing.The most common method of training is the Back Propagation(BP)algorithm.The core of the algorithm is the gradient descent algorithm,but the gradient descent algorithm is easy to fall into the local extremum.The more the parameters,the more difficult it is to find the global optimal neural network.Covariance Matrix Adaptive Evolution Strategy(CMA-ES)is an excellent intelligent algorithm with good performance.The paper uses the CMA-ES algorithm to optimize CNN and mitigate the local optimality of the network model.Due to the limitations of computer hardware conditions,the paper selected the LeNet-5 network model with relatively simple network structure as the research object.The paper mainly expounds the method of optimizing the parameters of convolutional neural network using CMA-ES algorithm,aiming at improving the accuracy of CNN model,and proposes CMA-ES-CNN algorithm.Although the CMA-ES algorithm has excellent search effect,its mathematical expression is cumbersome,complicated in calculation,and has many parameters,which will certainly consume a large amount of memory resources.In order to reduce the consumption of memory resources,the paper simplified the CMA-ES algorithm and proposed the Simplify Covariance Matrix Adaptive Evolution Strategy(SCMA-ES),which abandoned one of evolution paths.and use the square root of the covariance matrix to replace the covariance matrix itself,which reduces the complexity of the algorithm,reduces the memory resources occupied by the matrix decomposition,uses SCMA-ES algorithm to optimize the CNN parameters,and proposes SCMA-ES-CNN algorithm;finally compare SCMA-ES-CNN algorithm,CMA-ES-CNN algorithm and Genetic Algorithm(GA)convolutional neural network on MNIST(Mixed National Institute of Standards and Technology,MNIST)data set Performance;collect several students' handwriting as a general data set,compare the correct rate of SCMA-ES-CNN and CNN,and verify the generalization performance of the improved algorithm.