Research On Low Illumination Image Enhancement Based On Deep Learning

In recent years,with the wide application of computer vision technology,more and more research pay attention to image enhancement technology,which is a basic research work in the field of computer vision.Meanwhile,Low illumination image enhancement is one of the main research topics in image enhancement.Low illumination image enhancement refers to the recovery of degraded images caused by visual sensors in low illumination environments by using the technology of image processing and pattern recognition.At present,many researchers have studied different low-light image enhancement problems based on different theoretical foundations.Various deblurring,denoising,and brightness adjustment techniques have been proposed.However,there are still many challenges.Besides,the recovery of the image has met many difficulties due to image degradation and lack of detail caused by low illumination image.In order to solve the problem of noise reduction,color distortion,and the recovery of image details in low light environments,a deep learning algorithm combined with Retinex theory is proposed in this thesis,which is compared on two real datasets.The proposed algorithm has significant improvement in visual effects and quantitative indicators.Especially,the proposed algorithm can enhance the ability of noise removal and color deviation correction.Our proposed algorithm has certain theoretical guiding significance and application value for the research of image enhancement in a low light environment.The main contributions are presented as follows:1)An outdoor multi-exposure low-illumination image dataset was established.In this paper(thesis),the existing low-lightness datasets are analyzed.We collected a lot of low illumination images simultaneously through multiple cameras and multiple shooting methods.Each image has Image RAW data and RGB data,which increase the diversity of images in real scenes,and also ensure the robustness of subsequent algorithms.The dataset of this paper provides reliable assurance and support for the study of subsequent related issues.2)An end-to-end dual-channel autoencoder network framework is proposed.Taking into account the challenges of noise removal,color distortion,and local detail recovery of low-light image enhancement,the frame encodes the image through a two-way convolution autoencoder network to obtain the reflection feature coding and the illuminance hidden coding.The brightness of the image is automatically adjusted by illuminance hidden coding,and the decoder can recover and reconstruct image content.3)A joint loss function is designed to constrain the network.Due to the image decomposition based on Retinex theory is an ill-posed problem,it is no reasonable constraint on the output.The loss function of this paper is divided into three parts,namely,reconstruction loss,color loss,and reflection loss.The reflection characteristics based on the low-illuminance image and the normal exposure image should have the same structure and detailed.Based on the prior knowledge,the reflection loss constraint is designed,including the detailed constraint and the structural similarity constraint.Then,in order to correct the color deviation of the reconstructed image and to ensure image content,color loss and reconstruction loss are designed.The experimental results verify the effectiveness and advancement of the proposed algorithm.