A Plug-and-Play Structurally Reparameterized Convolutional Sparse Coding Module for Convolutional Neural Network

The powerful representation learning ability of convolutional neural networks (CNNs) is primarily attributed to their convolutional layer operations. However, from another perspective, these operations also pose challenges in interpreting CNNs. In contrast, convolutional sparse coding (CSC) is generally considered to be interpretable. Nevertheless, existing CSC research primarily focuses on either improving optimization solutions or network mapping, and its application to a broader range of image recognition problems remains limited. In this paper, we propose an efficient CSC module, termed RepCSC, considering the impact of initialization and dictionary design on the performance of the CSC model. RepCSC uses the previous layer of output for initialization, reducing reconstruction errors and achieving a single-layer residual structure. Simultaneously, it introduces the re-parameterization technique into dictionary design to expedite inference speed. Additionally, RepCSC serves as a plug-and-play architectural unit that can directly replace standard convolutions in various convolutional neural networks. Experiments conducted on Mini ImageNet and downstream tasks in medicine and agriculture demonstrate that the CNN incorporating RepCSC (RepCSC-Net) exhibits strong representation learning capabilities, high classification accuracy, and no increase in model parameters.