摘要
图像容易受外界照明条件的影响或相机参数条件的限制,导致图像整体偏暗、视觉效果不佳,降低了下游视觉任务的性能,从而引发安全问题。以人脸识别任务为驱动,提出了一种基于对比学习范式的非成对低光照图像增强算法Low-FaceNet。Low-FaceNet主干采用基于U-Net结构的图像增强网络,引入特征保持、语义分割和人脸识别3个子网络辅助图像增强网络的训练。使用对比学习范式可以使得真实世界大量非成对的低光照和正常光照图像作为负/正样本,提高了真实场景的泛化能力;融入高阶语义信息,可以指导低阶图像增强网络更高质量地增强图像;任务驱动可以增强图像的同时提升人脸识别的准确率。在多个公开数据集上进行验证,可视化与量化结果均表明,Low-FaceNet能在增强图像亮度的同时保持图像中各种细节特征,并有效地提升低光照条件下人脸识别的准确率。
Images are susceptible to external lighting conditions or camera parameters, resulting in overall darkness and poor visualization, which can degrade the performance of downstream vision tasks and thus lead to security issues. In this paper, a contrastive learning-based unpaired low-light image enhancement method termed Low-FaceNet was proposed for face recognition tasks. The backbone of Low-FaceNet was in the form of an image enhancement network based on the U-Net structure, introducing three sub-networks, i.e., feature retention network, semantic segmentation network, and face recognition network, thereby assisting the training of the image enhancement network. The contrastive learning paradigm enabled a large number of real-world unpaired low-light and normal-light images to be used as negative/positive samples, improving the generalization ability of the proposed model in the wild scenarios. The incorporation of high-level semantic information could guide the low-level image enhancement network to enhance images with higher quality. In addition, the task-driven approach made it possible to enhance images and improve the accuracy of face recognition simultaneously. Validated on several publicly available datasets, both visualization and quantification results show that Low-FaceNet can effectively improve the accuracy of face recognition under low-light conditions by enhancing the brightness of images while maintaining various detailed features of the images.
作者
范溢华
王永振
燕雪峰
宫丽娜
郭延文
魏明强
FAN Yi-hua;WANG Yong-zhen;YAN Xue-feng;GONG Li-na;GUO Yan-wen;WEI Ming-qiang(Institute of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing Jiangsu 210016,China;State Key Laboratory of New Computer Software Technology,Nanjing University,Nanjing Jiangsu 210023,China)
出处
《图学学报》
CSCD
北大核心
2022年第6期1170-1181,共12页
Journal of Graphics
基金
国家自然科学基金项目(62172218,62032011)。
关键词
低光照图像增强
人脸识别
对比学习
任务驱动
语义分割
low-light image enhancement
face recognition
contrastive learning
task-driven
semantic segmentation
