Research On Visualization Of Convolutional Neural Network Based On Gradient Analysis

In recent years,deep learning has become a hot topic in the field of computer science and other fields of learning and scientific research,although humans have used convolutional neural network models to conduct many scientific research related to deep learning,but the convolutional neural network model is still like a "black box".Therefore,in order to recognize and understand the internal working mechanism of the model,it is one of the key tasks in the field of deep learning to classify the features obtained from the internal convolutional computing kernel and convolutional layer,and to understand the image feature information obtained from each level of the convolutional neural network.Through the results of feature visualization,the network structure is appropriately adjusted to optimize the network,avoid blind parameter adjustment,and then optimize the network characteristic information at a faster rate.This paper mainly uses the visualization method based on gradient analysis to study the visualization effect of different convolutional neural network models,and carries out research work from two categories: single-target and multi-objective according to the image data contained in the image dataset.The work of this article is as follows:(1)Based on deconvolution technology,the features extracted by different layers of the convolutional neural network model were visually analyzed.Finally,the experimental results show that the lower layer of the convolutional neural network mainly extracts simple features such as color,outline and texture of the input image.The upper layer extracts more complex abstract features such as eyes,mouths,and wings of the input image,and as the number of network layers deepens,the features extracted inside the network are closer to the conceptual features of the physical object.(2)A feature visualization method based on gradient-class activation is proposed.The experimental results show that the head features of animals are an important basis for the network model to make decisions.Based on this method,an improved Grad-CAM++ method is proposed for feature visualization of multi-target objects,which mainly visualizes the features of images containing multiple homogeneous targets by updating the calculation method of the last layer of the weight of the image and combining the target selection gradient.Experimental results show that compared with other visualization methods,this method performs better in the visualization of multi-target images,and the generated heat map contains more similar target information.(3)The effectiveness of visualization algorithms is generally evaluated through visual coherence,visual resolution,and multi-objective visualization.In order to better evaluate the visualization effect,a visualization evaluation method is proposed,which uses Gaussian noise to generate an adversarial image that is indistinguishable from the original input image to test the anti-spoofing of the visualization method,so as to evaluate its robustness.Experimental results show that the generalization ability of this visualization evaluation method is better than that of other visualization evaluation methods,and can be used for the effect evaluation of other visualization methods.(4)A software system was designed to visualize the visualization results.The system can visually display the visualization effect of different visualization methods,and is simple to operate and easy to use.