Low-Light Image Enhancement Via CNN-Transformer Dual-Stream Feature Extraction And Convolutional Dictionary Learning

The purpose of low-light image enhancement(LIE)task is to improve the brightness and contrast of input images,to make information hidden in dark areas visible,and to improve the perception or interpretability of images.LIE has attracted a lot of attention in several emerging computer vision fields.This thesis summarizes and analyzes the state of the art in low-light image enhancement tasks,and proposes a low-light image enhancement algorithm via CNN and Transformer twostream feature extraction and a low-light image enhancement algorithm via convolutional dictionary learning and attention mechanism,and the practical applicability of the two algorithms is also verified.1.Low-light Image Enhancement via CNN-Transformer Two-stream Feature Extraction(LoCTTs)In order to make full use of local and global feature information,and therefore to improve the performance of low-light image enhancement tasks,and to limit the computational cost to a reasonable range,a low-light image enhancement algorithm via CNN-Transformer twostream feature extraction,namely LoCTTs,is proposed.The algorithm completes image enhancement in two stages.In the first stage,a dual-stream heterogeneous network based on CNN and H-Transformer is designed at the encoder side to extract local and global features respectively.In the "jumping section",a local information processing branch based on Res B module in image spatial domain and a global information processing branch based on FFT-FRCB module in the frequency domain are proposed to achieve multi-view and fine-grained information enhancement of the input feature maps,and then the AFF modules are used for cross-branch feature fusion,and to provide additional rich information for the decoder.At the decoder side,E-Transformer that introduces a convolution operator to model the long-term dependencies of feature information while combining the advantages of local contextual information for detail enhancement is used for level-by-level decoding,and an initial enhanced image is achieved.In the second stage,a conditional generative adversarial network is constructed.With unpaired high visual quality images as input conditions,the initial enhancement images from the first stage are coloradjusted and more detailed information can be recovered through image perception qualityguided adversarial learning,and the final enhanced images with high visual quality can be output.Based on the PSNR and SSIM evaluation metrics,the effectiveness of the proposed LoCTTs algorithm has been demonstrated by performing comparison experiments with ten comparative algorithms in recent years on three data sets including LOL,LOL-Real and LSRW.2.Low-light Image Enhancement via Convolutional Dictionary Learning and Attention Mechanism(LoCDA)In order to effectively correct the color and brightness of the input low-light image while reconstructing image texture structure,and to get a more lightweight model,a low-light image enhancement algorithm via convolutional dictionary learning and attention mechanism,namely LoCDA,is proposed.The algorithm consists of two coupling parallel branches:(1)the branch based on convolutional dictionary learning,and(2)the branch based on attention mechanism learning.The first branch not only adaptively learns dictionary D from input images to obtain the morphologically richer atom bases for better matching of image structures,but also learns the prior knowledge of the dictionary D and the coefficient matrix X based on powerful representation learning ability of deep neural networks,which therefore increases the flexibility of the model.In this branch,image details and texture information can be restored in an iterative manner,and the initial enhanced image can be achieved.The second branch adaptively learns the color transformation matrix and Gamma value from the input low-light images,and acts on the initial enhanced result to adjust the color and brightness of the sample image so as to further improve image perceptual quality and finally achieve high quality enhancement of input lowlight images.Based on the PSNR and SSIM evaluation metrics,the effectiveness of the proposed LoCDA algorithm has been demonstrated by performing comparison experiments with ten comparative algorithms in recent years on three data sets including LOL,LOL-Real and LSRW.3.Practical application of the algorithmThe practical applicability of the proposed two algorithms is verified in two real-world scenarios.The application of LoCTTs and LoCDA for real indoor and outdoor low-light image enhancement demonstrated the effectiveness of both algorithms in daily life image processing.Comparative experiments conducted on low-light image object detection based on Cascade RCNN model show that both LoCTTs and LoCDA can effectively improve the performance of advanced visual tasks when used for input image preprocessing in indoor and outdoor dark environments.