Research On Image Enhancement Method In Low-light Condition

With the rapid development of modern information technology,images have be-come one of the important sources for humans to obtain information.However,due to low-light,rain,snow,fog and other complex external environmental conditions,the qual-ity of the obtained images is often unsatisfactory,which brings great challenges to the wide application of images.For example,low-light images usually have problems such as low visibility,low contrast,and color degradation,which affects the presentation of image visual effects and the performance of subsequent visual tasks.In order to extract and use the useful information in the low-light image and improve the usability of the image,it is necessary to carry out the research on the image enhancement in the low-light environment,so as to restore the low-light image quality and improve the target detection performance in the low-light environment.In response to the above problems,the main research results of this paper are as follows:(1)To address the problems of low visibility in low-light images,a progressive net-work model based on Retinex decomposition is proposed.The low-light images and reference images are decomposed into illumination map and reflection map.On this basis,the iterative pixel dynamic range adjustment of the illumination map is carried out through the progressive illumination enhancement network.In order to maintain the structure of the low-light image reflection map during the enhancement process,an au-toencoder based on U-Net is designed to reconstruct the structure.In addition,in order to solve the problem of color distortion in the reflection map during structural reconstruc-tion,the L₁loss is introduced in the HSV color space to restore the color of the reflection image,so that the final fusion result can reach the color saturation under normal lighting conditions.Finally,the effectiveness of the algorithm is verified on the low-light data set;(2)To address the deterioration of target detection performance in low-light environ-ment,an end-to-end image enhancement method for target detection is proposed.The back propagation of the target detection network is used to optimize the front image enhancement network,so as to realize the joint training of the image enhancement net-work and the target detection network to improve the target detection performance in the low-light environment.In addition,in order to constrain the enhancement network in the end-to-end network training so that it tends to optimize the direction of brightness enhancement,a reference-free brightness enhancement constraint loss that conforms to human visual perception is introduced.The results of target detection in the low-light image data set show that this method can effectively improve the performance of target detection in the low-light environment.