Research On Deep Learning-based Non-uniform Illuminated And Nighttime Low-light Image Enhancement Algorithm

As a fundamental computer vision and image processing task,low-light image enhancement plays an important role in both life and industry.Its main purpose is to selectively enhance the underexposed areas in the image,or suppress the local overexposed areas in the image,so that the image matches the visual response characteristics;the key point is how to balance the overall and local illumination.At the same time restore the original contrast and color information of the image.Researchers have proposed many strategies in this long-term exploration,which have solved some problems in this field to a certain extent,but the results are still not completely satisfactory.Therefore,this paper delves into low-light image enhancement from the following two aspects:1.In order to solve the problem that the current algorithm cannot deal with the underexposed and overexposed areas of the image at the same time.This paper proposes an iterative enhancement fusion network algorithm.The main contributions are as follows:(1)This paper proposes a non-uniform brightness low-light image synthesis algorithm based on the principle of local random brightness distribution.Different from the existing methods that directly adjust the global brightness value of the image,the method in this paper assigns different brightness values to different local areas of the image in a relatively random way,while retaining a certain relationship between illumination information and semantic information.This enables the low-light image dataset synthesized in this paper to cover most of the illumination distribution in the real world,which further enhances the generalization performance of the model.Comparative experiments show that the network model trained on the dataset in this paper can better deal with backlit images,and the ability to restore image colors is also stronger.(2)According to the idea of progressive enhancement,this paper proposes an iterative enhancement fusion network framework.This paper splits the entire augmentation process into two sub-stages.The initial enhancement network in the first stage performs initial enhancement on the input image through a lightweight stretch coefficient estimation network,and the fusion network based on the encoder-decoder structure in the second stage combines the initial enhancement result with the local optimal value of the input image.The regions are fused at feature level to obtain the final enhancement result of this round.Through multiple rounds of iterations to gradually optimize the intermediate results,the method in this paper can obtain enhanced results with suitable brightness and rich details at a very short running speed.The experimental results show that the algorithm in this paper basically achieves the best results on both synthetic datasets and real datasets.2.To solve the problem that it is difficult for existing algorithms to overcome the problem of image enhancement at very dark nights,this paper proposes a light-guided brightness enhancement and denoising network framework.The main contributions are as follows:(1)By analyzing a large number of real night images,this paper first proposes a strategy for synthesizing night images.Nighttime images tend to be characterized by low overall light intensity,low contrast,and high levels of noise.The synthesis strategy in this paper simulates the illumination distribution of real night images by separately counting the pixel value information of bright and dark areas in real night images.This paper additionally adds synthetic noise to the image to simulate the real situation.The experimental data show that,compared with other datasets,the synthetic dataset in this paper is closer to the visual perception of the human eye for extremely dark environments in terms of visual effects.(2)With the idea of illumination guidance,this paper proposes a network framework for restoring brightness,contrast and denoising.The light map reflects the brightness relationship between a low-light image and its corresponding normal exposure image,and it abstracts the lighting information as a guide for subsequent enhancements.The light map obtained from the input image serves as an attention mechanism to guide the U-Net-based enhancement network to restore the brightness and contrast of the input image in a more targeted manner.The intensity of signal-related noise in low-light images is often related to pixel values.By combining the light map,the denoising network can remove noise in a more targeted manner,resulting in a cleaner and smoother enhanced image.Extensive comparative experimental results show that,compared with other methods,the enhancement results obtained by the proposed algorithm can better recover lighting from darkness,enhance details,and remove noise.