Research And Implementation Of Low Light Enhancement Algorithm Based On GAN Unsupervised Learning And Global Self-Attention Mechanism

Due to factors such as outdated equipment hardware,insufficient light,and short exposure times,the quality of images captured under these conditions often suffers from a lack of light signals.The missing details and contrast of these insufficiently illuminated images not only cause unpleasant subjective feelings,but also degrade the performance of many high-level computer vision systems.To make details visible in low light images,recover the features of low light images as much as possible and suppress noise generation,and improve the performance of computer vision systems,this paper focuses on the major difficulties faced by current research and proposes an unsupervised learning based dual-branch fusion low-light image enhancement network based on GAN to enhance low-light images using unpaired datasets and generative adversarial networks.At the same time,indepth analysis of the characteristics of nighttime images is conducted.In order to further enhance the global luminance consistency of enhanced images,this paper conducts in-depth analysis of the characteristics of nighttime images and proposes an image enhancement network based on lightweight Transformer global self-attention,which optimizes the parameter redundancy and high network complexity problems of current low light enhancement algorithms.Specifically,the work in this paper is summarized as follows:(1)In order to solve the problems of noise amplification and image quality degradation during the enhancement process,this paper proposes a GAN-based unsupervised learning two-way fused low-light enhancement network,which can learn from unpaired low light and normal light datasets to map low light images to normal light images.The generator network for performing enhancement steps on the input image is composed of two branches,the upper branch is a refinement branch focusing on noise suppression,and the lower branch is a U-Net-like global reconstruction branch for high-quality image generation based on the attention mechanism.The discrimination network adopts the multi-scale discrimination structure of feature pyramid to enhance the global consistency and avoid local overexposure.The loss function is also improved,and a new fidelity cycle consistency loss is introduced to further improve the quality of image texture information recovery.Through comparative analysis of many qualitative and quantitative experimental results,it has been proven that the method proposed in this paper can effectively suppress the generation of enhanced image artifacts and noise amplification while enhancing the visual effect of the image(2)To address the problems of redundant algorithm parameters,complex network structure that leads to the inability to output enhanced images in real time and the inconsistent global brightness of enhanced images,this paper proposes a lightweight Transformer-based global self-attention image enhancement network.Most of the current low-light enhancement algorithms are for global image enhancement,which leads to local overexposure problems after the enhancement of unevenly lit images.Therefore,this paper improves the Transformer self-attentive mechanism to not only fully utilize its advantages in global image feature correlation,but also avoid the problem of increasing computational complexity caused by the redundancy of parameters.Experiments have shown that this algorithm can achieve a good balance between outputting high-quality images and performance,and can also consider global brightness consistency when processing images with uneven shading distribution.