CPSNet: Channel Pixel Spatial Feature Fusion Network for Image Dehazing

Single image dehazing in low-level computer vision is a challenging problem. In this chapter, a multi-branch dehazing network referred to as “Channel-Pixel-Spatial Feature Fusion Network” (CPSNet) was proposed, which was inspired by the human visual system (HVS) for removing real-world dense haze. The authors, first, introduced three attention modules such as channel, pixel, and spatial in parallel for simulating “retina ganglion cells” (RGCs) independently using a simple block architecture. Second, to improve the network performance, the authors incorporated a “supervised attention module” (SAM). Third, traditional residual connection was replaced with “mixed skip connection” (MSC) with a variable that can be learned by increasing the adaptability of the learning process. At last, the authors optimized the network by a novel loss function rather than using common loss functions. The experimental results of the proposed network achieved better results competitively on dehazing datasets, which also boost the performance of high-level computer vision applications.