【作者】 杨群；

【导师】 吴裕斌；

【作者基本信息】 华中科技大学 ， 光电信息工程， 2020， 硕士

【摘要】 近些年来,机器视觉技术被广泛地应用于自动驾驶、机器人导航等领域。在这些领域中,视觉算法经常面对例如林荫道、隧道等明暗对比强烈的高动态范围场景。如何对类似的高动态范围场景进行成像是一个值得研究的问题。针对上述问题,本文设计了一种基于深度学习的高动态范围成像算法,主要包含以下工作:(1)研究了一种不良曝光像素数量最小化的自动曝光算法。在高动态范围场景下,该算法能够自动优化相机的曝光时间,减少不良曝光像素数量。(2)研究了一个用于相机图像信号处理和单曝光高动态范围成像的卷积神经网络HDR-Net。HDR-Net由U-Net改进而来,其计算量更小,同时加入全局特征模块以处理大面积弱纹理区域。HDR-Net对图像传感器输出的12位Raw图像进行去马赛克、去噪和增强操作,并输出8位RGB图像。并且输出图像能直接被后续图像识别算法使用。为了训练HDR-Net,本文采集制作了Raw12数据集,该数据集包含60个场景,每个场景有多张不同曝光时间的12位Raw图像和一张使用多曝光高动态范围成像技术合成的参考图像。(3)搭建了实验平台,在多个典型场景和数据集上对上述算法进行了功能测试,并对算法代码做了性能调优。实验表明,本文的自动曝光算法采集的图像中的不良曝光像素数量小于通用算法采集的图像。本文的HDR-Net在Raw12测试集上的SSIM和PSNR分别达到0.95和23.55。在高动态范围场景下,使用上述两个算法模块构建的整体算法的输出图像质量高,适合用于后续的图像识别,并且分辨率为2048×2448时运行速度可以达到18FPS。更多还原

【Abstract】 In recent years,machine vision is widely used in fields such as autonomous driving and robot navigation.In these fields,vision algorithms often face high-dynamic-range(HDR)scenes,such as tree-lined roads and tunnels.How to image such HDR scenes is a valuable research topic.Aiming at above problems,we design a high-dynamic-range imaging(HDRI)algorithm based on deep learning,which mainly includes the following works:(1)An auto exposure algorithm that can minimize the number of bad pixel caused by severe poor exposure is proposed.In HDR scenes,the algorithm can automatically optimize the camera’s exposure time to minimize the number of bad pixel in the image.(2)A CNN called HDR-Net for camera image signal processing(ISP)and single-exposure HDRI is designed.HDR-Net is smaller than naive U-Net,and have a global feature block to deal with large textureless regions.HDR-Net performs demosaicing,denoising and enhancement on 12-bit Raw images and outputs an 8-bit RGB images,which can be directly used by subsequent image recognition algorithms.For training HDR-Net,we collect the Raw12 dataset,which contains 60 scenes,each scene has multiple 12-bit Raw images acquired in different exposure times and a reference image synthesized using multi-exposure HDRI technology.(3)The experimental platform is built.Our algorithm is tested on several typical scenarios and Raw12 datasets.Experiments show that the area of extremely poor exposure area in the image acquired by our automatic exposure algorithm is smaller than that acquired by the common algorithm.HDR-Net’s SSIM and PSNR on the Raw12 test set can reached 0.95 and 23.55 respectively.In the HDR scene,the overall algorithm constructed using the above two algorithm modules has high output image quality and is suitable for subsequent image recognition,and the running speed can reach 18 FPS when the input resolution is 2048 × 2448.更多还原