Research On Low-light Image Enhancement Technology Based On Generative Adversarial Network

The images collected by image acquisition devices in low-light environments are often low-quality images with low brightness,missing features,and insufficient image information.Through image enhancement processing,the overall brightness of low-quality images can be effectively improved,and detailed features in the image can be highlighted,thereby improving image quality and providing quality assurance for subsequent image processing.With the continuous maturity of deep learning technology,more and more scholars use neural network to solve the problem of low-light image enhancement,and have achieved certain results.However,limited by the lack of reference and no reference data,the existing models are prone to poor generalization ability and overfitting problems;and the increasing complexity of related models also poses more challenges to scene deployment and real-time performance.This thesis makes relevant research work on the above problems,which are summarized as follows:(1)Aiming at the weak generalization ability of image enhancement methods due to insufficient paired reference data in supervised learning,this thesis proposes a few-sample generative adversarial network low-light image enhancement model.The model makes a series of improvements on the basis of maintaining the original generative adversarial nets: first,the Diff Augment is embedded in the network structure to expand the data.Secondly,in order to extract features more effectively and improve the training effect,the generation network adds a residual structure and sc SE(Concurrent Spatial and Channel ’Squeeze & Excitation’)attention mechanism.In addition,in order to improve the stability of model training,the network optimization process uses the huber function to smooth the loss value.Experiments show that the proposed method has better indicators and stronger generalization performance.(2)Aiming at the problem that the model structure is difficult to apply in supervised learning,this thesis proposes a lightweight image enhancement method combining neural network and wavelet transform.First,for the low-frequency part,because the existing wavelet base is easy to damage the image brightness structure,bilinear interpolation and bilateral filtering are used to simulate the low-frequency part of the image,and input the simulated lowfrequency information to generative adversarial network;for the high-frequency part,in order to meet the detail information under normal illumination,Through color coordinate system conversion,dual thresholds are designed to individually enhance the brightness component.Secondly,a multi-scale Retinex with color is introduced to adjust the reconstructed brightness,and bilateral filtering is used to smooth the noise existing in the image.The experimental results show that the proposed new method can not only achieve the weight reduction of the network but also maintain the effect of image quality.Compared with most traditional methods,there is a higher improvement in various evaluation indicators;compared with deep learning methods,while greatly reducing the weight of the network,it has also achieved better evaluation results.(3)In view of the lack of reference data in unsupervised learning and the consideration of lightweight models,this thesis proposes a Zero-DCE no-reference low-light image enhancement method combined with Cycle GAN.Firstly,a double-discrimination Cycle GAN network is designed to generate low-light images of the required scene from the overall and detailed considerations,so as to directly expand the non-reference data.Secondly,from the perspective of unsupervised learning and light weight,based on the Zero-DCE zero-reference network,the low-light images generated by Cycle GAN are enhanced.This method is not only simple in structure and convenient for deployment,but also greatly saves manpower and simulation costs.Experimental results show that the proposed method has obvious advantages in various evaluation indicators compared with similar unsupervised learning methods.