Gradient - Based Convolutional Neural Network Feature Visualization

This paper proposes a novel method for evaluating the robustness of visualization. In terms of effectiveness, the paper assesses visual coherence, visual resolution, and multi-target visualization. Regarding the robustness of visualization methods, a new evaluation approach is introduced by applying the model's adversarial resistance assessment from image classification to visualization methods. The paper incorporates deconvolution techniques to showcase image features extracted by different convolutional layers of Convolutional Neural Networks (CNN) and employs class activation techniques to visualize critical regions influencing CNN decisions. Additionally, for multi-target input images, a gradient-class activation-based visualization method is proposed and compared across different models, demonstrating its applicability to any model and its superior ability to retain target information compared to other visualization methods. Finally, the paper introduces an evaluation method for visualization effects, focusing on both effectiveness and robustness. In practical applications, this work provides valuable insights for the interpretability of deep learning models and handling complex scenarios.