MSAFF-Net: Multiscale Attention Feature Fusion Networks for Single Image Dehazing and Beyond

Single image dehazing is a critical problem in computer vision. However, most recently proposed learning-based dehazing methods achieve unsatisfactory quality with dehazed images due to inaccurate parametric estimation. The size of these models is also large to be applied with mobile devices' limited resources. Last, most models are tailored to image dehazing, achieving poor migration. Thus, we propose a compact multiscale attention feature fusion network with a model size of 2 MB called MSAFF-Net to perform end-to-end single image dehazing. In the proposed model, we design a simple and powerful feature extraction module to extract complex features from hazy images. We use a channel attention module and a multiscale spatial attention module to consider the regions with haze-relevant features. To our knowledge, this study is the first to directly apply the attention mechanism rather than to embed it into certain modules for single image dehazing. We compare MSAFF-Net with other approaches on the NTIRE18, RESIDE, and Middlebury Stereo datasets. We show that MSAFF-Net achieves comparable or better performance than other models. We also extend MSAFF-Net to single image deraining, and various experiments demonstrate its effectiveness. Results suggest that MSAFF-Net can directly restore clear images using channels with the most useful haze- or rain-relevant features and spatial locations.