高低频通道特征交叉融合的低光人脸检测算法  

Low-light face detection method based on the cross fusion of high-and low-frequency channel features

作　　者：许皓 钱宇华 王克琪[1] 刘畅 李俊霞 XU Hao;QIAN Yuhua;WANG Keqi;LIU Chang;LI Junxia(Institute of Big Data Science and Industry,Shanxi University,Taiyuan 030006,China;School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学大数据科学与产业研究院,山西太原030006 [2]山西大学计算机与信息技术学院,山西太原030006 [3]山西大学计算智能与中文信息处理教育部重点实验室,山西太原030006

出　　处：《智能系统学报》2024年第2期472-481,共10页CAAI Transactions on Intelligent Systems

基　　金：科技创新2030-“新一代人工智能”重大项目(2021-ZD0112400);国家自然科学基金项目(62136005);山西省揭榜挂帅项目(202201020101006)。

摘　　要：低光条件下捕获的人脸图像存在着噪声严重、对比度低等不足,极大影响了现有人脸检测器的准确性,另外,现有的低光图像检测算法欠缺小区域人脸信息的提取能力。为此,提出一种基于深度学习的两阶段人脸检测算法,即利用现有的低光图像增强算法对人脸图像进行增强后再进行检测。为提升检测算法对人脸信息的提取能力,设计一种新型的高低频通道特征交叉融合方法,该方法首先使用高低频通道特征可分离模块分离出不同尺度特征的高低频信息,然后对上述信息进行交叉融合,提升网络提取高频细节信息和低频色域信息的能力,进而提高检测网络的性能。对比试验和消融试验验证了该研究方法的有效性,试验结果表明该方法优于基准方法4.0%mAP。Face images captured under low-light conditions suffer from significant noise and low contrast,which negatively impact the accuracy of existing face detection systems.In addition,existing low-light image detection algorithms struggle to extract information from small facial areas.To tackle these issues,this paper proposes a two-stage face detection algorithm based on deep learning.This algorithm enhances low-light images before initiating the detection process using an established low-light image enhancement method.The objective is to enhance the ability of the detection method to extract facial information.Thus,a new cross-fusion method of high-and low-frequency channel features is designed.The first step involves using a separable module for high-and low-frequency channel features,enabling the separation of different scale features.These separated features are then cross-fused to improve the ability of the network to extract high-frequency details and low-frequency color information.This,in turn,improves the performance of the detection network.The comparative and ablation experiments validate the effectiveness of the proposed method.The experimental results demonstrate that our method surpasses the baseline method by 4.0%mAP.

关 键 词：低光人脸检测 高低频通道特征 低光增强 多尺度特征融合 计算机视觉 图像处理 深度学习 频率域分析 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]