Research On Denoising And Enhancement Algorithms For Low-light Image Based On Deep Learning

In recent years,smartphone cameras have become one of the essential tools in people's daily life,and obtaining high-quality photos has gradually become a goal pursued by people.However,in low-light scenes such as night or dim lighting,the images captured by the camera often suffer from high noise,low contrast,and a lot of loss of detail and color.This situation will not only seriously affect the visual effect of the human eye,but also affect many computer vision-related applications,such as face recognition and security monitoring.The purpose of this dissertation is to study the denoising algorithms and contrast enhancement algorithms for low-light images.Based on deep learning algorithms and some classic low-light image enhancement algorithms,a series of researches are launched by this dissertation.The main content of the dissertation is as follows:(1)Research on image denoising algorithm.With a joint attention mechanism and dilated convolution,a CNN blind image denoising algorithm is proposed.First,a lightweight Triplet attention module and a non-local attention module that can capture non-local information are introduced in this algorithm.At the same time,the dilated convolution is used to expand the receptive field of the network without increasing the model's parameters.Then,the idea of residual network is used to design the overall network structure so that the model can quickly converge during training.Experimental comparison proves that on the SIDD dataset and DND dataset,the proposed image denoising algorithm has improved both PSNR and SSIM metrics.(2)Research on low-light image enhancement algorithm.Inspired by Retinex theory,a new low-light image enhancement algorithm is proposed.First,the decomposition module is designed by this algorithm,which imitates the method of multi-scale Retinex to estimate the illuminance map,and decomposes the image into a reflection map and an illuminance map.Then,the decomposition results are processed by the denoising module and the enhancement module respectively.Finally,the color loss is introduced into the loss function design of the network to ensure color consistency.Experimental comparison proves that the proposed algorithm can not only enhance image contrast,but also remove most of the noise while maintaining details.At the same time,the comprehensive results of multiple metrics verify the superiority of this algorithm.(3)Image denoising and enhancement model fusion and model compression.Based on the image denoising and enhancement algorithm proposed in(1)and(2),a set of algorithm flow for a specific camera to achieve low-light image enhancement is designed.Taking the Honor20 mobile phone as an example,more than 10,000 images of extremely dark scenes were collected,and the SIDD algorithm flow was used as post-processing to create an extremely dark scene dataset.Next,the clean image synthesized from multiple short-exposure images is used as the intermediate reference image,and the long-exposure image is used as the final reference image,so as to achieve the fusion of image denoising and image enhancement models.Finally,model compression techniques such as knowledge distillation and model pruning are used to compress the size of the model and reduce the floating point operations.Experimental comparison proves that compared with the independent enhancement model,the fusion model proposed in this dissertation improves the PSNR and SSIM metrics by 4.64 d B and 0.067 respectively,and the model compression technology reduces parameters by about 10 times while ensuring the performance indicators.