一种改进的多任务级联卷积神经网络人脸检测算法  

An improved face detection algorithm based on multi-task cascaded convolutional neural network

作　　者：刘彩云[1] 李雅雯 刘倩 LIU Caiyun;LI Yawen;LIU Qian(School of Information and Mathematics,Yangtze University,Jingzhou 434023,Hubei;Electronics and Information School,Yangtze University,Jingzhou 434023,Hubei)

机构地区：[1]长江大学信息与数学学院,湖北荆州434023 [2]长江大学电子信息学院,湖北荆州434023

出　　处：《长江大学学报（自然科学版）》2021年第6期111-118,共8页Journal of Yangtze University(Natural Science Edition)

基　　金：湖北省教育厅科学技术研究项目“基于小波变换的位场数据处理与解释方法研究”(B2016034)。

摘　　要：人脸识别是人工智能的重要应用领域之一,人脸检测是人脸识别的关键步骤。由于姿势变化、外物遮挡以及光源方向等多方面因素的影响,人脸检测的准确率不高,并且对于多人图片,往往很难准确地识别出所有人脸。提出了一种改进的多任务级联卷积神经网络的人脸检测算法(IMTCNN):对R-Net层网络集成图片信息卷积残差模块,通过扩大特征图的感受野来获取更多人脸信息,以提升R-Net层网络对人脸目标的检测鲁棒性,并且通过加入反卷积层和最大池化层解决特征融合时维度不一致问题;对O-Net层网络集成图片信息卷积残差模块,进一步提升对多人图片的人脸检测性能,降低人脸检测过程中受外部条件影响产生的误差,同时添加2个卷积池化层使特征融合时维度一致。通过改进R-Net层和O-Net层集成图片信息卷积残差模块,扩大特征图的感受野,对图片进行人脸候选框定、选区筛选以及人脸关键点定位,最终实现人脸检测。试验结果表明,该算法速度快,准确性高,并且可以一次性检测多张人脸,为后续人脸识别打下了良好的基础。Face recognition is one of the important application areas of artificial intelligence,and face detection is a key step of face recognition.Due to the influence of various factors such as posture changes,object occlusion,and the direction of the light source,the accuracy of face detection is not high.And for multi-person pictures,it is often difficult to accurately identify all faces.In this article,an improved face detection algorithm based on multi-task convolutional neural network(IMTCNN)was proposed:the image information convolution residual module was integrated into the R-Net layer network.To improve the robustness of R-NET layer network to face target detection,more face information was obtained by enlarging the receptive field of feature image.The problem of dimensional inconsistency during feature fusion was solved by adding deconvolution layer and maximum pooling layer.The convolution residual module of image information was integrated into the O-Net layer network to further improve the face detection performance of multi-person images and reduce the errors caused by external conditions in the face detection process.At the same time,two convolution pooling layers were added to make the dimension of feature fusion consistent.By improving the R-Net layer and the O-Net layer to integrate the image information convolution residual module,the receptive field of the feature image was expanded,and face candidate framing,selection area screening and face key point positioning were performed on the images to finally realize face detection.The test results show that the algorithm is fast and accurate,and can detect multiple faces at once,laying a good foundation for subsequent face recognition.

关 键 词：级联卷积神经网络 图片信息卷积残差模块 P-Net层 R-Net层 O-Net层 人脸检测 人脸关键点 定位 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]