【作者】 王伟；

【导师】 岳涛；

【作者基本信息】 南京大学 ， 电子与通信工程（专业学位）， 2020， 硕士

【摘要】 弱光成像技术是成像领域中一个重要研究方向,在国防军事、安防监控、环境监测和野生动物保护等领域有着广泛的应用需求,同时高质量的成像技术有利于后续的视频分割检测识别等任务。然而,目前硬件实现的弱光成像技术成本高昂,难以实现商业级的应用,同时软件实现的弱光成像技术信噪比低并且缺乏颜色信息,所得图像与人眼感知特性差别巨大。因此,如何利用现有的成像设备,通过计算摄像学技术来经济有效地实现弱光成像,成为研究的热点课题。弱光成像技术的研究难点在于如何有效去除低照度下相机的高增益设置引起的大噪声,以及如何在夜间光照不均衡的条件下保持动态范围的均衡。本文对弱光成像技术的难点进行分析,研究提出了两种关键算法:视频去噪算法和高动态范围增强算法。本文提出了一种基于实际相机噪声建模的视频去噪算法。该算法通过探索数字相机在高增益下噪声的物理成因,建立数学模型,标定噪声参数以及制作符合实际的噪声视频训练集,训练基于卷积LSTM的视频去噪增强的网络架构。在主观视觉和客观指标上都取得了很大的性能提升,尤其是在实际噪声视频上效果显著。本文提出了一种基于深度学习的视频高动态范围增强算法。该算法在网络架构中加入最新的注意力模块来处理运动区域的对齐问题,同时加入了感知损失模块来进一步增强网络的HDR细节重建的能力,在HDR数据集上获得了明暗区域细节丰富的视觉效果。更多还原

【Abstract】 Low-light imaging technology is an important research direction in imaging field,which has a wide range of applications in national defense and military,security monitoring,environmental monitoring and conservation of wildlife,etc..Meanwhile,highquality imaging is conducive to the following tasks such as video segmentation,detection and recognition.However,the current hardware implementation of low-light imaging technology is expensive and difficult to commercial applications.The low signal to noise ratio(SNR)and the lack of color information in the computational lowlight imaging technology make the image different from the human eye perception.Therefore,how to make use of the existing imaging equipment to realize low-light imaging economically and effectively through computational photography algorithm has become a hot issue of research.The difficulty of low-light imaging is how to effectively remove the large noise caused by the camera’s high sensitivity setting under low illumination and how to maintain the balance of dynamic range under the condition of uneven illumination at night.This paper analyzes the difficulties of low-light imaging technology and proposes two key algorithms: video denoising algorithm and high-dynamic range enhancement algorithm.This paper proposes a video denoising algorithm based on real camera noise modeling.In this paper,we explore the physical origins of the practical high sensitivity noise in digital cameras,model them mathematically,calibrate the model parameters and generate the training data,and propose to enhance the low light videos based on the noise model by using an LSTM-based neural network.The performance of the al-gorithm has been greatly improved in both subjective vision and objective indicators,especially in the real noise videos,the effectiveness is remarkable.This paper proposes a high dynamic range enhancement algorithm based on deep learning.The algorithm adds the latest attention module to the network architecture to deal with the alignment of the motion region,and the perception loss module to further enhance the HDR detail reconstruction capability of the network,detailed visual effects in the light and dark regions are obtained on the HDR data set.更多还原