Research On Low-light Image Enhancement Algorithm Based On Multi-path Convolutional Network

Image quality is closely related to the performance of many computer vision related technologies.High-quality images can convey more information.Many factors in daily life can directly or indirectly affect image quality,and low-light condition is one of them.If the low-light images can be enhanced in advance to improve the image quality,the performance of some computer vision tasks can be improved,such as recognizing or segmenting object of interest at night or deep sea,and other low light conditions.Therefore,researches on low-light image enhancement have important practical and theoretical value.This paper focuses on the image enhancement algorithm in low-light conditions.Many existing low-light image enhancement deep learning methods are based on Retinex theory.These methods usually need to train an additional decomposition network to separate luminance and reflectance.It will increase the number of model parameters and calculations.The low-light image enhancement method based on multi-path convolutional neural network proposed in this paper can avoid these defects.This thesis first introduces some basic theories of image processing and deep learning,reviews and discusses the related works of low-light image enhancement methods,and then presents the research methods in our thesis.This thesis describes the motivations,assumptions and results of our proposed multi-path convolutional structures,and systematically designs experiments to verify them.The main contributions of this thesis include: 1)For low-light image enhancement tasks,this thesis proposes two enhancement methods based on multipath convolutional neural network architecture.2)Considering that the low-light environment has different effects on the luminance and chroma of the image,this thesis designs a special network structure and loss functions and verifies their effects.3)The impact of global information to low-light image enhancement model is carefully considered in this thesis.The experiments show that introducing global information can improve the model's effectiveness.The low-light image enhancement model proposed in this thesis shows distinctive advantages by comparing with the state-of-the-art low-light image enhancement algorithms in public data sets using common indicators.