DMCGF Dehazing Neural Network Design for Edge-AI Implementation
Kuo-Yi Chang, Kunlin Li, Ming‐Hwa Sheu, Szu‐Hong Wang · 2024
In many applications such as autonomous driving, outdoor surveillance, and aerial photography, the haze has become one of the main environmental factors affecting the performance of visual systems. This paper proposes a lightweight dehazing neural network, called DMCGF (Dual Multi-scale Channel Gate Fusion), which includes 4 modules: 1. Multi-scale feature extraction module for capturing local details of haze and the distribution of haze information. 2. Gate fusion block for dynamic weight adjustment fusion and addressing the up-sampling error problem. 3. Feature enhancement module to refine and enhance the fused features, further improving the quality of the dehazed image. 4. Reformulated ASM module to generate dehazed images using the hazy image and the output feature map. From experiments on the RESIDE outdoor dataset with (640x480) images, the proposed DMCGF has the efficient performance with 25.13 PSNR, 0.9401 structural similarity index, 2.204G FLOPs and 0.252M parameters, which are better than the latest works [6]–[8]. Also, the DMCGF neural network has been implemented on an Edge-AI platform to achieve 24.3 FPS for real-time applications.