| As cameras on various mobile devices become readily-available and more widespread,many people recently begin to show an interest in taking pictures in their daily life.Images acquired from low-light environments are often degraded due to complicated lighting conditions.Such low-light photos suffer from extremely dark zones,unexpected noise and blurred details.Not only affects human visual perception,but also significantly affects the performance of many advanced computer vision applications.Based on the deep learning theory,this thesis uses attention mechanism and multi-scale model to explore the low-light image enhancement methods.It aims to restore an image captured in the low-light condition to a normal one,where visibility,contrast,and noise are expected to be improved,stretched,and suppressed,respectively.The main work includes the following aspects:1.Aiming at the problem of insufficient correlation constraint of RGB channels in low-light image enhancement,this thesis proposes a color channel fusion enhancement network.The proposed network consists of three stages.Among them,channel correlation network maintains the interaction relationship between RGB channels by modeling channel interdependencies,which effectively improves the brightness of low-light images.Multi-scale fusion network utilizes channel shuffle strategy to strengthen the exchange of multi-scale information,and adaptively fuses these information to obtain brightly colored images.Finally,detail enhancement network is used to further enrich the texture details.Experimental results show that the proposed algorithm significantly improves the performance of low-light image enhancement compared to state-of-the-art algorithms.2.To make full use of the multi-scale information in pyramid network to recover the detail,this thesis proposes a bi-directional feature-aware pyramid fusion network.The proposed network consists of three stages to progressively enhance images.Different stages use the feature pathway with a weighted attention mechanism to transfer the shallow/deep information features to deep/shallow layers,which realize the interaction of multi-scale and multi-depth feature information.The multi-scale features are further refined and fully fused to obtain enhanced images with natural colors.Experimental results demonstrate that the proposed algorithm can effectively enhance the brightness of low-light images,while maintaining the image details and naturalness better. |
