Improving Skip Connection in U-Net Through Fusion Perspective With Mamba for Image Dehazing

Under hazy weather condition, images captured by electronic imaging devices frequently encounter a number of issues, such as blurring of image details and the poorly defined edges. Image dehazing technique is the most intuitive way to address the above issues and shoulders a key role for final presentation in mobile devices. However, existing U-Net based dehazing models still have some room for improvement, due to the fact that they either focus on purely improving the encoder/decoder or simply employing convolutional neural network (CNN) to enhance the performance of skip connections. To this end, a fusion Mamba (S6) based U-Net image dehazing method (FMamba), that has linear complexity for long-range modeling, is proposed. Specifically, considering that features at different scales have different abstract semantics, we firstly designed a Multi-scale Progressive Fusion module (MPFusion) to mitigate the semantic gap across multiple network layers, and then embedded S6 into the MPFusion to leverage its global context modeling capabilities. Moreover, a Channel Spatial Fusion module (CSFusion) is further developed in the decoder stage to produce a higher-quality output. Qualitative and quantitative results on real-world and synthetic datasets indicate that our FMamba achieves a leading performance against state-of-the-art alternatives.