Interpretation Techniques For Convolutional Neural Networks Based On Class Activation Mapping

With the rapid development of artificial intelligence,in some specific fields,the decision-making performance of artificial intelligence has surpassed the decision-making ability of human beings.However,the problem of the black-box model in the decisionmaking process of artificial intelligence leads to correct results of the model,but cannot be reasonably explained,hindering the further application of artificial intelligence.In order to improve the interpretability of AI,the deep learning decision-making process needs to be transparent to build a connection and trust between humans and AI.Currently,there are more and more applications of deep convolutional neural networks in artificial intelligence and deep learning,making the interpretability of deep convolutional neural networks attract much attention.In the field of computer vision,visualization is one of the most popular ways to explain deep convolutional neural networks.Among them,the method based on class activation map can highlight the main part of the input image and reveal the salient feature regions related to the decision of the neural network.However,there are many problems in the generation of saliency maps by the basic activation map-like methods,including the generated images contain a lot of meaningless information,such as noise or feature localization range is too large,and the identification of important features is not accurate.Furthermore,generating saliency maps for some models requires extensive training and long computation time.Therefore,this thesis conducts an experimental analysis of the Score-CAM model.In-depth discussion of the principle of the Score-CAM model to identify important features of images and the shortcomings of the model itself,and to improve it.First of all,on the basis of Score-CAM,this thesis proposes to improve the Score-CAM model by using Gaussian filter-based perturbation transformation,in order to optimize the input image of the convolutional neural network to have less noise and less meaningless information interference,and can reflect more important features in the image.Secondly,considering that the use of bilinear interpolation upsampling in the ScoreCAM model may cause certain interference and errors in the upsampled image,this thesis proposes to use the integral gradient to process the bilinear interpolation upsampled image,aiming to reduce the upsampling There are problems such as bad factors and errors in the image,and it also lays the foundation for the Gaussian filter to optimize the Score-CAM model.Finally,the method of improving the Score-CAM model in this thesis is evaluated on the ILSVR2012 data set,and the two proposed methods are experimentally analyzed and compared with the experimental results of the original model.The experimental results show that the improved Score-CAM model method proposed in this thesis has more advantages in the accuracy of image feature recognition and the ability to recognize multiple targets in the image than the existing related methods.In the visualization results,the overall information recognition ability of important targets in the input image has also been improved,with better performance.