TSNet:A Two-stage Network for Image Dehazing with Multi-scale Fusion and Adaptive Learning

TL;DR

TSNet introduces a two-stage dehazing framework that enhances generalization and reduces artifacts by combining a Multi-scale Fusion Module with an Adaptive Learning Module. The MSFM expands receptive fields and fuses frequency information, while the ALM uses deformable convolution to preserve texture details. A key novelty is the learning objective shift to opposite fog maps, coupled with a second-stage refinement that further mitigates residual artifacts and color distortion. Across synthetic and real-world datasets, TSNet achieves state-of-the-art performance, with TSNet-L delivering top PSNR/SSIM on multiple benchmarks and the two-stage design providing tangible gains beyond deeper single-stage networks.

Abstract

Image dehazing has been a popular topic of research for a long time. Previous deep learning-based image dehazing methods have failed to achieve satisfactory dehazing effects on both synthetic datasets and real-world datasets, exhibiting poor generalization. Moreover, single-stage networks often result in many regions with artifacts and color distortion in output images. To address these issues, this paper proposes a two-stage image dehazing network called TSNet, mainly consisting of the multi-scale fusion module (MSFM) and the adaptive learning module (ALM). Specifically, MSFM and ALM enhance the generalization of TSNet. The MSFM can obtain large receptive fields at multiple scales and integrate features at different frequencies to reduce the differences between inputs and learning objectives. The ALM can actively learn of regions of interest in images and restore texture details more effectively. Additionally, TSNet is designed as a two-stage network, where the first-stage network performs image dehazing, and the second-stage network is employed to improve issues such as artifacts and color distortion present in the results of the first-stage network. We also change the learning objective from ground truth images to opposite fog maps, which improves the learning efficiency of TSNet. Extensive experiments demonstrate that TSNet exhibits superior dehazing performance on both synthetic and real-world datasets compared to previous state-of-the-art methods.

Figures (10)

• Figure 1: The results of TSNet compared with other state-of-the-art methods on the RESIDE-IN dataset. The size of the circle represents #Param
• Figure 2: The internal structure of TSNet. The dashed box represents the process of changing the learning objective, 'a' is the hazy image, 'b' is the opposite fog map, 'c' is the dehazed image obtained from the first-stage network, and 'd' is the final dehazed image of TSNet
• Figure 3: (a) is the structure of the multi-scale fusion module, mainly consisting of a multi-scale parallel convolutional kernel module and an implicit frequency feature enhancement module. (b) is the internal structure of the implicit frequency feature enhancement module
• Figure 4: (a) Sampling points of regular convolution (b) Sampling points of deformable convolution
• Figure 5: The internal structure of the ALM. DCN 3×3 is a deformable convolution with 9 sampling points