CasDyF-Net: Image Dehazing via Cascaded Dynamic Filters

TL;DR

This work proposes using cascaded dynamic filters to create a multi-branch network by dynamically generating filter kernels based on feature map distribution and introduces a residual multiscale block (RMB), combining different receptive fields, to better handle branch features.

Abstract

Image dehazing aims to restore image clarity and visual quality by reducing atmospheric scattering and absorption effects. While deep learning has made significant strides in this area, more and more methods are constrained by network depth. Consequently, lots of approaches have adopted parallel branching strategies. however, they often prioritize aspects such as resolution, receptive field, or frequency domain segmentation without dynamically partitioning branches based on the distribution of input features. Inspired by dynamic filtering, we propose using cascaded dynamic filters to create a multi-branch network by dynamically generating filter kernels based on feature map distribution. To better handle branch features, we propose a residual multiscale block (RMB), combining different receptive fields. Furthermore, we also introduce a dynamic convolution-based local fusion method to merge features from adjacent branches. Experiments on RESIDE, Haze4K, and O-Haze datasets validate our method's effectiveness, with our model achieving a PSNR of 43.21dB on the RESIDE-Indoor dataset. The code is available at https://github.com/dauing/CasDyF-Net.

Figures (9)

• Figure 1: (a) Parameters vs. PSNR on the Haze4K dataset. (b) GFLOPs vs. PSNR on the SOTS-Indoor dataset. Our model achieves excellent performance with low computational overhead.
• Figure 2: Comparison of Several Convolutional Blocks in Convolutional Neural Networks: (a) residual block (RB) in SRResNet b28, (b) memory block (MB) in MemNet b29, (c) dense block (DB) in SRDenseNet b30, (d) residual dense block (RDB) in RDN b31, (e) proposed residual multiscale block (RMB).
• Figure 3: The Proposed CasDyF-Net Network Architecture.(a) CasDyF-Net employs a popular U-shape structure, where the CasDyF-Block is our proposed Cascade Dynamic Filtering block.(b) The proposed CasDyF-Block consists of three processes: Dynamic Segmentation, Local Fusion, and Global Fusion. Dynamic Segmentation includes Dynamic Filtering and our proposed RMB (Residual Multiscale Convolution).(c) DFS (Dynamic Filtering and Segmentation) divides the feature maps into two parts using dynamic filtering.(d) The proposed Local Fusion Module utilizes dynamic 1 convolutions to fuse three adjacent feature branches into the current branch, with a residual connection added to the current branch..
• Figure 4: Visual Comparison of Image Dehazing Effects on the SOTS-Indoor Dataset.
• Figure 5: Visual Comparison on the Hzae4K Dataset.