| Images serve as an essential way for humans to obtain information from the external world.During the image acquisition process,various uncontrollable factors can lead to low brightness,low contrast,noise,and color distortion in the captured images.These lowvisibility images not only fail to meet people’s visual perception and demand but also bring significant challenges to computer vision tasks such as image segmentation and target detection.Therefore,it is essential to design practical algorithms to recover low-light images.In this paper,based on the existing algorithms for low-light images,a new research idea is proposed,and the main work is divided into two parts:(1)A low-light image enhancement framework combining attention mechanism and Retinex model is proposed(CA&R Net),which exploits attention information and the complementarity between reflectance and illumination to enhance low-light images.CA&R Net consists of three parts: information extraction,reflectance recovery,and illumination adjustment.Unlike other Retinex model-based approaches,CA&R Net introduces an attention mechanism to emphasize important information about the processed object and suppress irrelevant detail information,providing guidance for reflectance recovery and illumination adjustment: extracting an attention map to assess the degree of image underexposure and guiding reflectance recovery in a region-adaptive manner to focus more on underexposed regions during recovery,thereby better enhancing these regions and avoiding over-enhancing normally exposed regions;using the recovered reflectance and low illumination to jointly predict the illumination layer of the image,this joint prediction not only uses the semantic information contained in the reflectance itself but also extracts attentional information from it,allowing the illumination adjustment to learn more detailed content.Experimental results show that CA&R Net outperforms other algorithms in both qualitative and quantitative metrics.(2)A decoupled semi-supervised network framework SSNF to enhance low-light images is proposed.SSNF decouples the low-light image enhancement task into two phases,visibility enhancement,and fidelity restoration,and builds two subnetworks Net1 and Net2 for these two phases to model them separately.Net1 uses an approach based on information entropy and Retinex model to improve the visibility of the input image.It is worth mentioning that it is a lightweight self-supervised network that only requires input low-light images and minute-level training to achieve luminance enhancement,and its enhancement results can be considered low-quality images under normal lighting conditions.Net2 uses UNet and residual networks to eliminate the noise and degradation problems present in these low-quality images,thus enhancing the visual properties of the enhanced images.This design approach,which decomposes the low-light image enhancement task into two subtasks,sequentially boosts the image brightness and solves the image degradation problem,breaking the limitations of merged processing.Experimental results show that SSNF exhibits better performance,shows better advantages in recovering details,and has better robustness and generalizability. |
