基于超分辨率重建的低分辨率人脸检测算法
Low-resolution Face Detection Algorithm Based on Super-resolution Reconstruction
王国辉 1陈健美1
作者信息
- 1. 江苏大学计算机科学与通信工程学院 镇江 212013
- 折叠
摘要
低分辨率人脸检测在视频监控等领域中有重要应用,而当前人脸检测算法在低分辨率人脸检测上效果并不理想,针对这一问题,论文提出了基于超分辨率重建的低分辨率人脸检测算法.首先,通过前置基础人脸检测器S3FD检测出大部分常规人脸,接着通过降低类别置信阈值,将可能含有人脸的候选区域送到基于GAN改进的超分辨率重建网络(MGAN)来进一步完成人脸检测任务,最后汇总人脸区域,并采用非极大值抑制算法得出最后检测结果.实验结果表明在WIDERFACE数据集上,论文算法相比S3FD等主流人脸检测算法检测精度更高,在hard子集上提升较明显.
Abstract
Low-resolution face detection has important applications in video surveillance and other fields.However,current face detection algorithms are not ideal in low-resolution face detection.For this problem,the paper proposes a low-resolution face detection algorithm based on super-resolution reconstruction.First,most of the normal faces can be detected by the prepositioned basic face detector.Secondly,by lowering the category confidence threshold.the regions proposal that may contain faces are sent to the super-resolution reconstruction network(MGAN)based on the improved GAN to further complete the face detection task.Final-ly,the face regions are summarized and the non-maximum suppression algorithm is used to obtain the final detection results.The ex-perimental results show that in the WIDERFACE data set,compared with the mainstream face detection algorithms such as S3FD,the proposed algorithms have higher detection accuracy,and the improvement is obvious in the hard subset.
关键词
超分辨率重建/生成对抗网络/人脸检测/低分辨率Key words
super-resolution/generative adversarial net/face detection/low-resolution引用本文复制引用
基金项目
国家自然科学基金项目(61702229)
江苏省自然科学基础研究计划基金项目(BK20150531)
出版年
2024
国家科技期刊平台
NSTL
万方数据