采用主成分分析与梯度金字塔的高动态范围图像生成方法

摘要

针对由多曝光低动态范围(LDR)图像融合生成高动态范围(HDR)图像对采集设备要求高且不适用于动态场景的问题,提出一种采用主成分分析(PCA)与梯度金字塔的HDR图像生成(HDR-PGG)方法.首先用改进的S曲线全局映射和Retinex局部色调映射对PCA变换后的亮度图像进行处理,以避免光晕和泛灰,并结合PCA逆变换获取多曝光LDR图像;其次利用多曝光LDR图像的对比度、饱和度、曝光度和亮度信息构造标量权重图,确保生成的HDR图像亮暗区域分明且细节清晰;最后,利用梯度金字塔生成HDR图像.实验结果表明,当视觉差异大于75％时,由HDR-PGG方法生成的HDR图像的平均像素分布概率比拉普拉斯金字塔和梯度金字塔算法降低了10.84％和30.75％,其平均相对熵与噪声程度比值分别提高了0.542 1和0.508 9.%A generating method (HDR-PGG) of high dynamic range (HDR) images using principal component analysis (PCA) and gradient pyramid is proposed to solve the problem that HDR image generated by the fusion of multi-exposure low dynamic range (LDR) images has high requirements for sampling equipment and is not applicable to dynamic scenarios.Firstly,the improved S-curve global mapping algorithm and the Retinex tone mapping algorithm are used to process a PCA transformed luminance image so that halo and graying are avoided,and the PCA inverse transformation is used to obtain multi-exposure LDR images.Secondly,the information of contrast,saturation,exposure and luminance of multi-exposure LDR images are used to construct scalar weight maps to ensure that the bright and dark areas of the generated HDR image are distinct,and the details in the HDR image are clear.Finally,the HDR image is fused using the gradient pyramid.Experiments are performed and the HDR images generated by HDR-PGG are compared with those by the Laplacian pyramid algorithm and the gradient pyramid algorithm.The results show that the average pixel distribution probability of HDR images generated by HDR-PGG method reduce by 10.84％ and 30.75％ when the visual difference is over 75％,respectively,and the average relative entropy and the noise ratio increases by 0.542 1 and 0.508 9,respectively,compared with those of the Laplacian pyramid algorithm and the gradient pyramid algorithm.