Noise Reduction Algorithm Of Video And Image Under Low Light Environment

Low-light-level night vision is a technology that uses photoelectric devices and image processing algorithms to convert weak light signals into video or images in low-light-level conditions.It has a wide range of applications in military night reconnaissance,public safety monitoring,and biomedical imaging.However,the video images captured in low-light-level environments often come with significant noise,which can affect subsequent video analysis and human observation.With the development of computational photography and digital image processing in related fields,using video image denoising algorithms to denoise the large noisy video images captured in low-light-level environments is of great significance for the development of low-light-level night vision technology.With the development of physical imaging devices and advancements in video and image processing technology,video and image denoising has become one of the important topics in low level computer vision research.Imaging in low-light-level environments often results in low signal-to-noise ratio,low contrast,image blur,and complex noise in the collected video images due to the extremely weak signal intensity.Therefore,this article combines computer image processing technology and computational photography knowledge to first analyze the noise in data collected in low-light-level environments,and then design noise generation models or denoising algorithms based on the characteristics of the noise.This article proposes the following methods to solve the problems of video and image acquisition in low-light-level environments:(1)Proposed a spatiotemporal EMCCD video denoising algorithm based on pixel similarity-weighted frame averaging.This paper analyzes the noise characteristics of EMCCD(Electron-multiplying charge-coupled device)video in low-light-level environments due to high gain settings,resulting in high noise and blurry images.A spatiotemporal joint video denoising algorithm is designed accordingly.The denoising algorithm combines improved Kalman filtering in the temporal domain and pixel similarity-weighted frame averaging for motion estimation and inter-frame noise elimination.In the spatial domain,bilateral filtering,histogram equalization,and slant stripe removal algorithms are performed for detail enhancement.The hardware implementation of this algorithm on an FPGA only occupies 30%of LUT and 36% of I/O resources,and the operating frequency can reach 22 MHz,which meets the requirements of processing video sequences at a rate of 25 fps.(2)Proposed a video denoising algorithm based on multi-innovation Kalman filtering.This algorithm is another solution to the problems mentioned in(1).The improved algorithm introduces the theory of multi-innovation to the original Kalman video denoising algorithm.The improved algorithm models the problem of denoising low-light-level videos,using the results of the current frame after guided filtering processing as the observation value of the current state of the system,and optimizes the denoising results by combining the prediction value of the multi-innovation Kalman filter.Through experiments comparing the Kalman video denoising algorithm before and after introducing the multi-innovation theory,the multi-innovation Kalman filtering algorithm shows a 4dB improvement in PSNR compared to the improved Kalman filtering algorithm.(3)Proposed a Raw domain image denoising algorithm based on noise modeling.This paper proposes a method of using a noise model to generate noise data from clear images,addressing the problem of difficulty in obtaining real paired data in low-light-level environments.First,the characteristics of CMOS imaging sensors in low-light-level environments are analyzed.Secondly,based on the noise characteristics of Raw domain images,a noise model is established.Then,a network architecture based on Transformer is used to calibrate the noise model parameters,and the calibrated noise model is used to synthesize a large amount of data for denoising network training.Finally,the feasibility of the proposed noise model is demonstrated through experiments.The experiments show that the method proposed in this paper outperforms other algorithms both in objective metrics and subjective visual evaluation.