Research On Low-light Image Enhancement Based On Deep Learning

Low-light image enhancement aims to solve a series of degradation problems of low-quality images captured under dim lighting conditions.Under poor shooting environments,the influence of unprofessional photographers and equipment,captured images often suffer from problems such as uneven lighting,severe noise,significant loss of details,and extreme low-light conditions.Improving the quality of low-light images by addressing degradation problems can enhance the performance of subsequent high-level visual tasks such as image recognition and also improve the performance of some practical applications such as intelligent systems like autonomous driving.However,there are still some issues in enhancing real low-light images.Typically,real low-light images captured under low-light conditions are severely affected by noise.Most existing methods adopt additional denoising methods as pre-processing or post-processing,but the former can cause blurring,and the latter can cause noise amplification.Moreover,there is also a severe loss of details during the low-light image enhancement process.To address these issues,this thesis conducted extensive research on methods for low-light image enhancement focusing on denoising and detail restoration.The specific research contents are as follows:1.Research on low-light image enhancement method based on Retinex theory.In order to effectively suppress noise and restore details,a new supervised convolutional network model based on Retinex theory is proposed.The model has three sub-networks,specifically,a decomposition network consisting of multi-scale U-shaped encoder-decoder networks is proposed,which decomposes the input image into illumination and reflection images with the reflection image of normal light image as a constraint.In addition,a new reflection reconstruction network is designed,which can reconstruct the reflection image based on the information of the illumination image,and uses Joint Attention Residual Units consisting of spatial attention mechanism,channel attention mechanism,residual units and dilated convolutions to achieve noise reduction and contrast enhancement simultaneously.An illumination adjustment network is also constructed to adjust the illumination image of low-light images,and combined with the reflection reconstruction network to obtain higher quality normal light images.Experimental results demonstrate that this method can effectively reduce the influence of noise and color distortion,and improve the effect of image enhancement.2.Research on low-light image enhancement method based on Multi-Attention generative adversarial network.A new unsupervised enhancement method based on dual discriminator GAN is proposed to enhance low-light images by integrating global and local features,overcoming the dependence on paired data in supervised methods.A multi-scale U-shaped encoder-decoder network generator guided by multiple attention mechanisms is proposed,and the global discriminator and local discriminator jointly guide the image discrimination,with the global discriminator focusing on the image structure and illumination enhancement,and the local discriminator focusing on the details and illumination enhancement in the region.The illumination channel is introduced as input,and the Joint Attention Module is used to further enhance the network’s attention guidance ability,specifically capturing important information in the image.Meanwhile,the Nonlocal module is introduced in the encoder-decoder network to preserve more detailed information.The perceptual loss and adversarial loss are jointly used for training.Experimental results demonstrate that this method achieves good low-light image enhancement performance.