Gated residual feature attention network for real-time Dehazing

Images captured under complicated weather conditions, such as haze, often suffer from a noticeable degradation and hamper its practical application. Traditional dehazing methods use various hand-crafted priors to get a clear image; in such cases, the performance is limited owing to unconstrained environment. In order to restore the haze-free image directly, we propose an end-to-end Gated Residual Feature Attention Network (GRFA-Net) that leverages the haze representations through feature restacking and propagation. We design a Feature Attention Residual Block (FARB) as the core of feature extraction, which employs the residual block to extract hierarchical features, and followed by a novel Feature Attention Module (FAM) that adaptively captures the inter-dependencies from channel- and spatial-wise perspectives. Furthermore, we utilize a group structure (GS) to enlarge the receptive field and merge different multi-level features via the gate fusion module (GFM), respectively. Extensive experiments demonstrate that our GRFA-Net can obtain results that are comparable or even better than previous state-of-the-art methods in terms of quantitative and qualitative evaluation metrics. Furthermore, we reduce the computational complexity considerably and obtain a real-time FPS. The code is available: https://github.com/leandepk/GRFA-Net.