一种基于感受野增强的人脸检测方法  

A face detection method based on perceptual field enhancement

作　　者：董春峰 杨春金 周万珍[1] DONG Chunfeng;YANG Chunjin;ZHOU Wanzhen(School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang,Hebei 050018,China;Hebei Taihang Machinery Industries Company Limited,Shijiazhuang,Hebei 052160,China)

机构地区：[1]河北科技大学信息科学与工程学院,河北石家庄050018 [2]河北太行机械工业有限公司,河北石家庄052160

出　　处：《河北工业科技》2022年第6期474-479,共6页Hebei Journal of Industrial Science and Technology

基　　金：河北省自然科学基金(F2018208116)。

摘　　要：为了解决多任务级联卷积神经网络(MTCNN)算法网络模型在小人脸检测方面鲁棒性较低的问题,提出了一种基于感受野增强的网络模型。首先,为MTCNN算法模型中的R-Net网络和O-Net网络添加感受野模块(receptive field blocks,RFB-S)。其次,通过添加批量标准化和全局平均池化,加速网络模型的收敛,防止模型过拟合。最后,调整网络任务的权重,P-Net和R-Net网络用于人脸区域粗筛选,O-Net网络用于人脸区域精筛选以及人脸关键点回归。实验结果表明,与MTCNN算法网络模型相比,所提模型缩小了16%,但检测速度提升了9%,在FDDB数据集上的检测精度提高了2.3%。因此,基于感受野增强的网络模型能有效完成人脸的检测任务,增强对小人脸检测的鲁棒性,可为人脸识别、表情识别等提供技术支持。Aiming at the problem of low robustness of MTCNN(Multi-task convolutional neural network)algorithm network model in small face detection,a network model of MTCNN algorithm based on receptive field enhancement was proposed.First,the Receptive Field Blocks(RFB-S)were added to the R-Net network and O-Net network in the MTCNN algorithm model.Second,the method of batch normalization and global average pooling was used to accelerate the convergence of the network model and prevent the model from overfitting.Finally,the weights of the network tasks were adjusted,the P-Net and R-Net networks were used for coarse screening of face regions,and the O-Net network was used for fine screening of face regions and face key point regression.The experimental results show that compared with the MTCNN algorithm network model,the proposed model size is reduced by 16%,the detection speed is increased by 9%,and the detection accuracy on the FDDB dataset is increased by 2.3%.Therefore,the network model based on perceptual field enhancement can effectively complete the face detection task,and enhance the robustness of small face detection,which provides technical assistance for subsequent tasks such as face recognition and expression recognition.

关 键 词：图像处理 人脸检测 多任务卷积神经网络 RFB-S 全局平均池化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]