MSCG-DehazeNet: A Trustworthy Multi-scale CNN and Graph Attention Fusion Network for Single Image Dehazing

Abstract

Dehazing is a basic image restoration problem that is used to restore the scene information from degraded images caused by atmospheric scattering and haze. Existing image dehazing techniques have been shown to be poor performers in natural image denoising in the presence of haze or limited for generalization to other haze situations. Recently deep learning based methods have shown a great deal of progress, but many current convolution neural network (CNN) models are not capable of modeling long-range spatial dependencies and lose out on structural information. To overcome this drawback, this paper introduces a Multi-Scale CNN-GAT Fusion Network consisting of two parts, including a CNN for extracting features and a GAT for capturing the context, to provide more comprehensive context representation. The proposed framework uses multi-scale convolution blocks to learn the local image features, and graph attention to model the non-local relation between image regions. The CNN and graph based representations are fused in a feature fusion module to produce high-quality dehazed images. The experiments were performed using the RESIDE-6K dataset with Adam optimiser with learning rate of 0.001 for 100 epochs of training.In Peak Signal-to-Noise Ration (PSNR) and Structural Similarity Index Measure (SSIM) have been used for performance evaluation. The experimental results show that the proposed method can obtain a PSNR of 23.40 dB and an SSIM of 0.9218, which is better than the traditional CNN, ResCNN, GAT, CNN-GNN and AOD-Net methods.