Research On Low Light Image Enhancement Algorithm Based On Deep Feature Fusion

Low light images refer to images or photographs taken in relatively complex environments such as low light,uneven light,backlighting,etc.Low light image enhancement is the process of enhancing the effects of these complex environments such as extreme darkness,noise,colour shift and backlighting by means of traditional model mapping or deep learning methods.Low-light image enhancement has received unprecedented attention in recent years.High-quality images are often required as the basis for advanced computer tasks such as target detection,target tracking and recognition,where the reality is that low image quality can often have a serious impact.Due to the rapid development of smartphones,mobile phone brands now focus on night photography algorithms,low-light image enhancement is developing rapidly,but still faces great challenges,such as noise generated by the signal conversion of the imaging process equipment,colour bias,image enhancement after the brightness of unnatural details blurred problem.Generally speaking the pixel composition of low-light images is extremely low contrast and difficult to distinguish under human eye observation,for computers the differences are even slight enough to be recognised and thus the image is accurately enhanced.For these elements this paper uses deep learning methods from unsupervised(zero-shot learning)and supervised respectively to specifically implement:1.Currently in low-light image enhancement,the majority of the data set is achieved by adjusting the exposure due to the small number of pairs of low-light images.Supervised learning is costly,and a Zero-Shot learning network model is suitable for this scenario,as it does not require a large amount of training time and only requires a network model combined with a loss function to achieve low-light image enhancement.In this paper,a Zero-Shot learning low-light image enhancement algorithm is improved by using the Retinex theoretical model to decompose low-light images into light,reflection and noise components and constructing a multi-scale feature extraction network combined with texture enhancement loss and histogram balance loss as the decomposition network.A luminance enhancement network is constructed jointly with light regularization loss to enhance the luminance of the light map,and Retinex theory is used to remove the noise.A zero-shot learning network model is designed to improve the problem of insignificant luminance enhancement and low contrast while effectively suppressing noise to enhance texture details.2.The lack of control information and paired learning data for reference-free networks makes it difficult to obtain high quality images with significant contrast,high brightness and no significant noise.Many recent image enhancement methods have focused on building multi-branch network models,which greatly increase the computational cost.This paper proposes a network model that deeply integrates Transformer network with CNN based on this basis.By fully exploiting the advantages of the Transformer for global control and the CNN for detailed control,a U-shaped Transformer network is constructed by combining the Swin-Transformer,which has three layers of 24,48 and 96 respectively,while the Sk-Fusion fusion module is used to fuse the feature information extracted by the CNN network module with the The fusion of the feature information extracted by the CNN network module with that extracted by the Transformer is then used to construct a soft reconstruction module in line with the Retinex theory to obtain an enhanced image.It greatly improves the problem that CNN loses a large amount of valuable information due to the fixed convolutional kernel pooling and ignores the correlation between local and global,which may make the convergence result converge to the local minimum of the domain instead of the global minimum.The Transformer network is used to supplement the global information through the fusion module,and a new network architecture is constructed.3.The system uses PyQt5 to build a GUI page,Python as the underlying algorithm language,and calls Zero-DCE low-light image processing network model,Faster R-CNN pedestrian search network model and YOLOv5 as an aid to extract the target character to be detected.In summary,this paper has designed two enhancement methods for different domains in unsupervised(Zero-Shot learning)and supervised models for low light image enhancement,both of which are able to enhance contrast while suppressing noise and increasing detail.Comparative experiments were also done on a large dataset using PSNR,SSIM,NIQE,Entropy,MSE and other metrics to evaluate the results and both have advantages over the latest methods.