Research On Low-Light Image Enhancement Method Under Low-Light Conditions

Along with the quick development of Internet technology,AI has set off an irreversible wave of development.As one of the important carriers of human information delivery,images carry a large amount of important research information and are an important prerequisite for the stable work of some AI systems.However,in the real scenes where images are acquired,there is inevitably interference from the low-light imaging environment,which leads to acquired images with low brightness,low contrast,high noise,a large amount of detailed information that cannot be represented,difficulty in distinguishing image colors and target structures,and other disturbances,which seriously affect the stability of the relevant AI systems.Therefore,research on low-light image enhancement algorithms with high performance and versatility is essential.This paper conducts an in-depth investigation of current low-light image enhancement algorithms,analyzes the model principles of different existing methods,and analyzes their advantages and shortcomings.We find that there are two main challenging problems in enhancing real low-light images,one is the problem of spatially independent image low-light;the other is the problem of non-uniform noise and color distortion.Based on the natural scene image’s own pixel information,this paper combines deep learning methods to obtain a low-light image enhancement model by training and learning.The research content and main contributions of this article are summarized as follows:(1)To address the issue of non-uniform noise distribution in low-light images,the darker the image,the higher the noise level in the area.Additionally,the contribution values generated by different channels of color information in the image during image restoration also have different proportional relationships.This paper proposes a low-light image enhancement network model based on fusion attention mechanism and contextual information.The network model is mainly composed of encoding/decoding modules,mixed attention modules,cross scale context modules,and fusion modules.Among them,the mixed attention module mainly consists of a spatial attention module and a channel attention module.The spatial attention module is used to calibrate different noise positions in the image through weight values,and the channel attention module assigns different weights to different color channel information in the image.The cross scale context module mainly aggregates information on semantic features of different depths in the network model,effectively reducing the problem of information loss caused by other reasons in the network model.Through a large number of subjective and objective experimental comparisons,it is shown that the network model proposed in this paper can effectively improve the brightness of the image to a certain extent,and the enhanced image noise suppression effect is obvious,and the color information of the image is well preserved.In terms of objective image evaluation indicators such as PSNR,SSIM,and LPIPS,the network model proposed in this article has improved by 0.74 dB,0.153,and 0.172 compared to the optimal values of the comparison method,respectively.(2)In the process of low-light image enhancement,there are issues such as how to correctly improve the brightness of different areas of the image,suppress noise,and maintain the consistency of image texture details.This paper also proposes a multi head self-attention U-shaped network model based on shift window.The network model is based on self-attention mechanism and mainly consists of encoder,decoder and jump connection.By utilizing the advantage of self-attention mechanism in obtaining global dependencies in the process of image feature extraction,the global features of the image are better obtained.Not only did it increase the accuracy of the network model in improving image brightness,but it also achieved better results in image detail restoration.Finally,the network model proposed in this article is compared with the comparison method through a large number of experiments.The results indicate that the network model proposed in this article has achieved good results in both subjective and objective aspects.