低照度场景下的时空频域视频去噪算法

Spatio-Temporal Frequency Domain Video Denoising Algorithm in Low Light Scene

随着移动端图像以及视频拍摄技术的广泛应用,低照度场景成为重要的成像场景之一。然而以夜晚、室内等为代表的低照度场景下,图像传感器采集到的视频图像普遍存在信噪比低和视觉效果差等问题。针对低照度场景下视频图像的质量问题,本文提出一种提升去噪效果的时空频域结合(TSF)视频去噪算法。所提出的TSF算法首先建立包含数字增益的低照度图像噪声模型,然后针对原始视频中的相邻前后两帧,基于光流法完成图像配准得到位移图,在时空域对高频图像进行去噪和重构,进而结合后续噪声图像实现无限脉冲响应处理,同时在空域进行降噪,输出后一帧去噪图像。在0.32lux到1.5lux的低照度环境下,基于当前手机中主流4800万像素图像传感器采集到的数据集,所提出的TSF算法与谷歌Pixel2所采用的实时视频去噪算法相比,具有更好的图像质量和纹理保持效果,同时与VBM4D去噪算法相比,有效降低运行时间和复杂度。

With the widespread application of mobile image and video shooting technologies,low-light scenes have become one of the important imaging scenes. However,in low-illumination scenes represented by night and indoors,the video images collected by the image sensor generally have problems such as low signal-to-noise ratio and poor visual effects. Aiming at the problem of video image quality in low-illumination scenes,this paper proposes a spatio-temporal frequency domain combined( TSF) video denoising algorithm to improve the denoising effect. The proposed TSF algorithm first establishes a low-illuminance image noise model by introducing digital gain,and then for the adjacent two frames of the original video,the image registration is completed based on the optical flow method to obtain the displacement map,and the high-frequency image is performed in the space-time domain with denoising and reconstruction,and then combined with the subsequent frame noise image to achieve infinite impulse response processing,while reducing noise in the spatial domain,output the next frame of denoised image. In a low illumination environment from 0. 32 lux to 1. 5 lux,based on the data set collected by the mainstream48 million pixel image sensor in the current mobile phone,the proposed TSF algorithm have better image quality and texture preservation compared with the real-time video denoising algorithm used on Google pixel 2,and compared with the VBM4 D denoising algorithm,it effectively reduces the computing time and complexity.