A Lightweight Adaptive Filtering Network for Efficient Image Dehazing

In recent years, seeking image dehazing solutions in the frequency domain has gained significant attention. However, current methods are insufficient in flexibly selecting the most informative frequency components for restoration, and considering all frequency domain components introduces redundant information from irrelevant regions. In this paper, we propose a lightweight adaptive filtering network (AFNet), which primarily includes an Adaptive Filtering Module (AFM) designed to select the most informative frequency components while removing feature redundancy in both the frequency and channel domains. The AFM consists of two parts: an Adaptive Frequency Filtering Module (AFFM) and a Channel Refinement Feedforward Module (CRFM). Specifically, the AFFM employs a dual-branch adaptive computation, where one branch selects the most informative components in the frequency domain and removes redundant information and noise interactions, while the other branch ensures the network receives sufficient information flow to learn discriminative representations. Simultaneously, the designed CRFM incorporates channel attention and gating mechanisms into the feedforward network to eliminate feature redundancy in the channels, thereby enhancing the restoration of clear latent images by utilizing important channels. Experimental results on commonly used benchmark datasets demonstrate the effectiveness of the designed AFFM and CRFM. For example, AFNet achieves a PSNR of 41.68 dB on the SOTS indoor dataset with only 2.42 million parameters.