Research On Low-light Image Enhancement Under Complex Illumination

As an important carrier of information,images always affect human daily life.Low-light images are often of low quality due to the influence of natural environment such as overcast,night and twilight,as well as objective factors such as visual occlusion,different standards of image acquisition equipment and uneven shooting techniques.On the one hand,it affects human visual experience.On the other hand,it limits the performance of computer vision system.Therefore,the researches on the enhancement of low-light images aim to solve the problems of low contrast,loss of texture details,color distortion,high noise level and fuzzy contour under complex illumination.It is of great theoretical and practical value to provide data quality assurance for target recognition and tracking,video post-processing and other tasks.This paper first introduces the basic theory of low-light image enhancement and convolutional neural network,and then analyzes the advantages and disadvantages of the existing low-light image enhancement methods based on traditional methods and deep-learning methods.On the basis of the existing methods,aiming at their shortcomings,three methods for low-light image enhancement are proposed.The researches of this paper are as follows:(1)Aiming at the problems of over-enhancement and color distortion of low-light images,a new method of image enhancement based on HSV space is proposed.The three color channels in RGB space have a strong correlation with brightness.With the change of brightness,the proportion of color components will change in the processing of image enhancement,which easily leads to color distortion.HSV space is more in line with the characteristics of human vision,and a new brightness adjustment model is designed based on its theory.In order to solve the problem of over-enhancement,an adaptive brightness adjustment function is designed to enhance the brightness of the dark areas of the low-light images adaptively,while the other areas with sufficient brightness are kept linearly.The experimental results show that the combination of the new brightness model and the adaptive function can effectively improve the phenomenon of image over-enhancement,and has good effect in maintaining the color saturation.In addition,it can not only solve the problems of low-light image enhancement,but also take into account the natural images.(2)In view of the limited numbers of images that can be selected by traditional algorithms,the poor generalization ability of parameters,and the poor noise suppression performance of images in extremely low illumination environments,this paper proposes a low-light image enhancement algorithm based on convolutional neural network,and constructs a new network architecture integrating decomposition,restoration and adjustment.The image is decomposed into a reflectance map containing object attributes and an illumination map containing illumination information.Many existing low-light image enhancement algorithms based on convolutional neural network often enhance the brightness of the decomposed illumination maps,and then reconstruct the processed illumination maps and reflectance maps.Such algorithms are easy to be affected by the original illumination of the images,resulting in the images of the enhancement being too dark or too bright,and the images tend to lack authenticity.This paper processes the reflectance maps and the illumination maps obtained by decomposition at the same time.The noise of the reflectance maps is suppressed,and the illumination maps adopt the idea of brightness adaptive adjustment in(1).The reflectance components of highlight images are used as the reference of suppressing the noise.The U-Net network is used to suppress the noise on the reflectance maps,and depth separable convolution is used to replace parts of the traditional convolution in U-Net,which can reduce the calculation cost and effectively suppress the noise,and the color saturation module is added to retain the color and other details in the process of image restoration to the greatest extent.Users can adjust the brightness of the illumination maps according to their own preferences,which shows the universality of the algorithm.Experiments show that this method can effectively suppress noise,improve image quality and avoid color distortion.(3)Most of the algorithms can only deal with the low-light images under the specific illumination,and lack of the low-light image enhancement algorithm which can efficiently solve the problems of complex illumination.In view of this phenomenon,this paper proposes a low-light enhancement method under complex illumination.Using the large framework in(2),we decompose the images with UNet++ which is more excellent in segmentation.We can extract more feature information of the original low-light image by taking the original low-light image as the input of the recovery network.In the recovery network,we design a UNet++ generator guided by attention mechanism to repair the low light defects,and design a color loss function to maintain the colors of the images.In the illumination map,an adaptive adjustment network is designed to transform the low illumination condition into the high-light condition,and a network composed of full convolutions for feature extraction is designed with the idea of maintaining the position relationship,so that the operation is focused on the detail texture and color enhancement rather than the scale change.The experimental results show that this method can enhance the low-light images significantly,appropriate brightness enhancement amplitude,high color saturation,clear texture details,fine edge details,and perfect performance of noise suppression.At the same time,it can also deal with low-light images under complex illunination.