Research And Application Of Low-light Image Enhancement Based On Deep Neural Network

Images are important source for computer vision systems to perceive the information of the outside world,whose quality has a great impact on the accuracy of subsequent computer vision tasks.However,images captured in the low-light environment(e.g.,overcast weather and nighttime)usually suffer from image degradation problems such as low brightness and contrast,loss of detailed textures,and strong noise.To improve the quality of low-light images,existing researches built the low-light image enhancement models based on traditional priors or deep convolutional neural networks,which achieved great success in quality improvement.However,these methods still face the following problems:1)most existing methods ignore the restoration of local details,resulting in the texture distortion of enhanced images;2)the color information of images are not effectively leveraged,resulting in the color distortion,3)the noises of the image are amplified during the enhancement process,resulting in a more complex noise distribution for the enhanced image,4)most existing methods mainly focus on global brightness and contrast enhancement,which makes the enhanced results contain over/under-exposed regions,resulting in nonuniform and unnatural illumination.This dissertation is devoted to developing effective low-light image enhancement methods to address the above problems.The main research contents are listed as follows:Firstly,aiming at the problems of textural detail degradation and color distortion,we propose a structure-texture aware network.By analyzing the differences between the global structure and local texture features of the low-light images,we design a guided filter that uses a finer contour map as the guided map to decompose a low-light image into global structure maps that retain large-scale structural features and local texture maps that contain fine details.Then,we construct a structure-texture aware module to separately process the extracted structure and texture information,which makes the proposed network can retain abundant detailed information when extracting the global structure information.Moreover,a color loss function is designed to alleviate the color distortion in the enhanced image.Extensive experiments demonstrate that the proposed method can effectively improve the visual quality of images and achieves enhanced results with abundant detailed information and high color naturalness.Secondly,aiming at the problem of noise amplification,we propose an amplified noise map guided network that continuously estimates the amplified noise map during the enhancement process by adopting the residual connection,and then we use the extracted amplified noise map to guide the denoising network to remove the amplified noise,which simultaneously achieves the low-light enhancement and noise removal.Extensive experimental results demonstrate that the proposed method can achieve a significant effect on noise suppression and produce high-quality enhancement results with low noise.Thirdly,aiming at the problems of the nonuniform illumination and damaged details,we propose a hierarchical feature mining network.By exploring the frequency distribution relationships between crucial image features(illumination and edge detail features)and hierarchical features,a feature mining attention module combined with a hierarchical supervised loss is introduced to mine crucial features in appropriate network layers,which makes the proposed network achieve uniform illumination with clear edge details when enhancing the brightness and contrast.In addition,an unpaired adversarial loss is introduced to prevent overfitting caused by deep supervision.Extensive experiments and analysis demonstrate that the proposed method can obtain enhanced results with uniform illumination and rich details,and outperform other methods in terms of image quality,increasing the PSNR and SSIM by 12.5%and 3.5%,respectively.At last,a series of applied researches of low-light image enhancement are carried out for practical scene tasks.By combining the proposed image enhancement methods with the night-time face detection and drone-based night-time pedestrian detection tasks,we can achieve a significant increase in detection accuracy and obtain better visual perception quality.In summary,in this dissertation,we propose a series of more effective low-light image enhancement methods based on convolutional neural networks,which address some problems of existing image enhancement models,contributing to improving the visual quality of low-light images.Extensive experiments on some public datasets demonstrate the superiority of the proposed methods.Moreover,we apply the proposed methods to night-time face detection and drone-based night-time pedestrian detection tasks,demonstrating the effectiveness of the proposed methods in practical applications.