SIS-CAM: An Enhanced Integrated Score-Weighted Method Combined with Gradient Optimization for Interpreting Convolutional Neural Networks

Abstract The opacity of deep convolutional neural network(CNN) models has hindered their performance enhancement across various domains, posing challenges in understanding their internal mechanisms. To address this, computer vision has developed approaches to assess CNN interpretability via visualization. However, existing techniques often encounter noise during gradient calculation and may produce rough, blurry saliency maps, leading to the localization of meaningless information. This paper proposes SIS-CAM, optimizing gradients using squared values during backpropagation and integrating the initial saliency map with the input image via feature fusion. The image is iteratively integrated with a masked approach, averaged, and linearly combined with the initial saliency map. This approach refines gradients through squaring, enhancing visual features of neuron activation and improving the saliency map’s effectiveness in capturing information. The improved gradients are integrated with feature mappings to derive preliminary masks, which are merged with the input image to derive secondary masks for accurate delineation of boundary features. Integration operations on the secondary masks compute average scores of masked input images, which are then amalgamated with the initial saliency map to generate the final map. The proposed method undergoes qualitative and quantitative evaluation, including Deletion tests, Insertion tests, Average Drop, Average Insertion tests, Class Discriminative Visualization, and sanity checks on 2000 images from the ILSVRC2012val dataset. Experimental findings show that SIS-CAM effectively reduces noise in saliency maps, accurately captures target boundary characteristics, and exhibits superior visual performance compared to the baseline model.