一种融合上下文信息特征的改进MTCNN人脸检测算法

Improved MTCNN face detection algorithm fused with context features

在课堂场景下,针对多任务卷积神经网络(multi-task convolutional neural network,MTCNN)人脸检测算法对小人脸检测率较低的问题,提出一种改进的MTCNN算法。首先,对MTCNN算法网络模型的R-Net层网络集成上下文信息卷积模块,扩大特征图感受野获取更多小人脸信息;其次,引入反卷积层与最大池化层,以解决特征融合数据维度不一致问题;最后,对MTCNN算法网络模型的O-Net层网络集成上下文信息卷积模块,进一步提取小人脸特征信息,并引入2个卷积池化层进行特征融合。实验结果表明:与MTCNN算法相比,所提方法在FDDB数据集上检测精度提升3%,在课堂场景数据集上人脸检测召回率与F_(1)分数分别提升8.69%和4.94%。

The multi-task convolutional neural network(MTCNN)face detection algorithm has a low detection rate of small faces in classroom scenes.An improved MTCNN algorithm that integrates context features was thus proposed.Firstly,the context convolution module was integrated with the R-Net layer network of the MTCNN model,and the feature map receptive field was expanded to obtain more small face information.Secondly,the deconvolution layer and the max-pooling layer were introduced to solve the problem of inconsistency of feature fusion data dimensions.Finally,the O-Net layer network of the MTCNN model integrated the context convolution module to further extract the small face feature information,and two convolutional pooling layers were introduced for feature fusion.The experimental results show the detection accuracy increased by 3%on the FDDB data set,the face detection recall rate and the F_(1) score increased by 8.69%and 4.94%respectively on the classroom scene data set compared with those of MTCNN algorithm.