Image Feature Separation Algorithm Based On Evolutionary Neural Network

At present,deep neural networks(DNN)are widely used in various fields of computer vision and have achieved remarkable results.However,there are still two important problems:(1)The large computing scale required is difficult to deploy to lightweight devices;(2)The end-to-end prediction model rarely requires manual intervention,resulting in lack of interpretability,and it is difficult to be trusted in important fields(such as the medical field).This paper proposes a new evolutionary neural network algorithm to study these two problems.The goal is to reduce the computational requirements of the DNN model and improve its interpretation ability.This paper uses the excellent random search ability of the evolutionary neural network algorithm to gradually compress the structure from the known network structure,and still maintain the original classification performance.This paper presents a method of dynamic pruning,which can effectively control the degree of model compression when dealing with different problems(data),and avoid the situation of sharp performance degradation of the network.In addition,This paper uses the evolutionary neural network algorithm to divide the features in the network according to the categories.The goal is to divide the network into independent feature sub-networks.In the future,these feature sub-networks can be combined into a model with high interpretation ability similar to a decision tree.In order to verify whether these feature sub-networks really divide the features form the original network according to categories,we took advantage of such as feature visualization and image activation heat map.We verify the above two problems on CIFAR-10 and CIFAR-100 datasets.For reducing the size of the network,we reduced the calculation of the ResNet-110 model to a maximum of 20.17%,the amount of parameters was only 19.17%,and the accuracy decreased by only 0.41%.For the second question,we visually verify that our method really separates the network features by feature visualization and other methods,and by testing the degree of accuracy reduction,we can find that our method controls the change in accuracy to-1% ~ 1 %.