电子测量与仪器学报

赖欣,王储,陈航.低照度下人脸检测 MSRCR 光频分段滤波增强算法[J].电子测量与仪器学报,2022,36(2):96-106

低照度下人脸检测 MSRCR 光频分段滤波增强算法

MSRCR optical frequency segmented filter enhancementalgorithm in low-light face detection

DOI：

英文关键词:face detection MTCNN multi-scale retinex with color restoration guided filtering image enhancement

基金项目:油气藏地质及开发工程国家重点实验室项目（PLN2020 10）、四川省科技厅应用基础研究面上项目（2019YJ0311）资助

作者	单位
赖欣	1. 西南石油大学机电工程学院，2. 石油天然气装备技术四川省科技资源共享服务平台
王储	1. 西南石油大学机电工程学院
陈航	1. 西南石油大学机电工程学院

Author	Institution
Lai Xin	1. School of Mechanical and Electrical Engineering, Southwest Petroleum University,2. Oil and Gas Equipment Technology Sharing and Service Platform of Sichuan Province
Wang Chu	1. School of Mechanical and Electrical Engineering, Southwest Petroleum University
Chen Hang	1. School of Mechanical and Electrical Engineering, Southwest Petroleum University

摘要点击次数: 1040

全文下载次数: 974

中文摘要:

在不受约束的环境下,人脸检测由于光照、遮挡和表情的不同具有一定的挑战性,低照度环境下多任务级联卷积神经网络(MTCNN)人脸检测器准确率下降。为提高低照度环境下人脸检测的准确率,提出了一种基于 MSRCR 光频分段滤波增强算法(3CGF-MSRCR)。利用 MTCNN 进行人脸检测,采用一种 RGB 三通道分解引导滤波(GF) 方法对多尺度视网膜增强算法 (MSRCR)进行改进。首先对人脸图像进行 MSRCR 增强并分解 RGB 三通道,得到三通道的图像权重,利用 GF 方法对各个通道分别进行滤波,更新三通道权重,最后重构人脸图像。在实际低照度场景人脸数据集 Dark Face 与公开的标准人脸数据集 CelebA 上进行了训练与测试,并在实际路灯场景下进行了测试,对比了算法的运行时间。测试结果显示:本文所提出的方法能有效抑制 MSRCR 的高频噪点,并保留亮度增强效果,提高了准确率,且算法运算速度较快。

英文摘要:

In an unconstrained environment, face detection is challenging due to differences in light, occlusion, and expressions. The accuracy of the multi-task cascaded convolutional neural network ( MTCNN) face detector is reduced in low-light environments. To improve the accuracy of face detection in low-light environment, a MSRCR-based optical frequency segmented filtering enhancement algorithm ( 3CGF-MSRCR) is proposed. This paper uses MTCNN for face detection, and uses a RGB three-channel decomposition guided filtering (GF) method to improve the multi-scale retinex with color restoration. Firstly, face images are enhanced by MSRCR and decomposed into RGB three channels to obtain the image weights of RGB. Then the GF method is used to filter each channel separately and update the weights of the RGB images. Finally, we reconstruct the face image. Training and testing are conducted on the actual lowlight scene face dataset: Dark Face and the public standard face dataset CelebA. Meanwhile, the running time of the proposed algorithm is compared with other enhancement algorithm. The results show that 3CGF-MSRCR can effectively suppress the high-frequency noise of MSRCR, retain the brightness enhancement effect, and improve the accuracy. Meanwhile, 3CGF-MSRCR has a faster running speed.

查看全文查看/发表评论下载PDF阅读器