Interpretation and Optimization of Shallow Convolutional Filters

Convolutional Neural Networks (CNNs) have become a cornerstone technology in deep learning, achieving remarkable success in computer vision in recent years. Despite CNNs' impressive performance in tasks such as image classification and object detection, their internal mechanisms remain opaque, particularly regarding the principles behind feature extraction by convolutional filters, which are often viewed as a “black box.” Shallow convolutional filters in CNNs play a crucial role in feature extraction, effectively capturing fundamental attributes such as color, texture, and edges. This study aims to elucidate the underlying logic of feature extraction by analyzing the weight distributions of shallow convolutional filters, providing a precise classification of these filters. Furthermore, using Bayesian optimization to determine the optimal configuration of different types of filters and eliminating redundant ones, we conducted experiments across diverse datasets and network architectures. The results demonstrate that the optimized networks achieve notable improvements in both accuracy and computational efficiency, with a reduction in computational overhead. These findings offer valuable theoretical insights for guiding CNN design and application.