Research On Low Light Visual Information Enhancement Method In Narrow Space Based On Deep Learning

Limited by environmental lighting conditions,narrow spaces are usually closed or weakly light,which makes the Hyper-Redundant Manipulator’s vision system unable to complete such environmental perception,foreign object recognition,and 3D reconstruction tasks in narrow spaces.Therefore,this paper fully analyzes the visual information characteristics of the low-light environment in narrow space and researches the enhancement method of low-light visual information in narrow space based on deep learning to improve the quality of low-light visual information.To ensure the smooth development of tasks related to the Hyper-Redundant Manipulator.The research carried out and experimental results obtained in this paper include:(1)Aiming at the problem that the visual information characteristics of the existing public low-light image and video datasets are inconsistent with the characteristics of the low-light environment in narrow spaces,this paper collects the narrow spaces’ low-light image and video datasets according to the characteristics of the low-light narrow environment to provide data support for the subsequent research of this paper.(2)Aiming at the problem that the traditional low-light image enhancement method has weak perception ability for the narrow spaces’ low-light images and the effect of the enhanced images are poor,this paper proposes a low-light image enhancement method based on image brightness estimation to further overcome the problem of inaccurate estimation of image brightness in the traditional method.This method proposes an encoder-decoder network and a loss function based on the image brightness formula,so that the model can accurately estimate the brightness difference between the normal image and the low illumination image when generating the enhanced low illumination image,and improve the brightness and color accuracy of the output image.The experimental results show that the proposed method can effectively enhance the brightness of low-light images and suppress local over-exposure in subjective visual perception,and the Peak Signal-to-Noise Ratio(PSNR)is improved by 4.2%compared with the GLADNet method.(3)In order to meet the requirements of single image processing speed in the vision system of the Hyper-Redundant Manipulator,this paper proposes a fast enhancement method for low-light images in narrow space using global frequency domain filtering.This method proposes a lightweight encoder-decoder network consisting of depthwise separable and transposed convolution and a feature screening and fusion method based on global frequency domain filtering.On the premise of improving the inference speed of the enhancement network,it further improves the quality of enhanced low-light images in narrow space.Experimental results show that the proposed method improves the inference speed by 25%and has better performance in subjective perception such as brightness enhancement and texture preservation.Compared with the GLADNet and Kin D method,the proposed method has 4.5% and 8.8% improvements in Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity(SSIM)respectively.(4)In order to satisfy the requirements of the vision system of the Hyper-Redundant Manipulator in real-time execution,this paper proposes a real-time enhancement method for narrow space low-light video via temporal consistency learning.Based on A fast low-light image enhancement network using global frequency domain filtering,a low-light video enhancement method is proposed to learn the temporal consistency between video frames through consistency loss,so as to avoid the calculation of optical flow and other prior information resulting in increased computational overhead.By learning the consistency between video frames,the proposed method eliminates the flicker and noise superposition in the output video and improves the quality of the enhanced video.Experimental results show that compared with the extended image model,the proposed method has better performance in the output video brightness flicker suppression and the output video frame noise superposition.In terms of objective indicators,the proposed method has the same reasoning speed as the fast enhancement method for low-light images in narrow space using global frequency domain filtering and has a 15.9% improvement in Peak Signal-to-Noise Ratio(PSNR)and a 77% improvement in Average Brightness Variance(ABV).Finally,the overall experimental results show that the proposed low-light image enhancement method in narrow space is superior to the comparison methods in terms of subjective visual perception and objective evaluation indicators,and the proposed low-light video enhancement method has a significant improvement in enhancing the temporal consistency of low-light video compared with the extended image method.Therefore,the lowlight image and video enhancement method proposed in this paper can make the HyperRedundant Manipulator’s vision system complete various visual tasks in narrow low-light environments more effectively.