Visualizing Deep Neural Networks with Double-Sliding Window
Chao Ma, Jianming Wang, Jiaming Liu · 2024
Deep neural networks are widely used in target recognition, image classification, speech signal processing. However, it is quite difficult to understand how deep neural networks complete the tasks from different fields. This paper proposes a method to analyze the principle of deep neural networks for target recognition by applying double-sliding window and dropping features to the model input image samples. First, the input image is divided into several blocks. Then, part of features in each block are removed in turn to observe the effect of the classification results. Finally, the average of the effect from each feature in the input sample on the classification results is calculated. Experimental results prove that the effects of features from the input image on the decisions of classification results are performed well by the method proposed. It can help realize the visualization analysis of target recognition with deep neural networks