基于改进Faster-RCNN的自然场景人脸检测

Face Detection in Natural Scene Based on Improved Faster-RCNN

为实现对自然场景下小尺度人脸的准确检测,提出一种改进的Faster-RCNN模型。采用ResNet-50提取卷积特征,对不同卷积层的特征图进行多尺度融合,同时将区域建议网络产生的锚框由最初的9个改为15个,以更好地适应小尺度人脸检测场景。在此基础上,利用在线难例挖掘算法优化训练过程,采用软非极大值抑制方法解决漏检重叠人脸的问题,并在训练阶段通过多尺度训练提高模型的泛化能力。实验结果表明,该模型在Wider Face数据集上平均精度为89.0%,较原Faster-RCNN模型提升3.5%,在FDDB数据集上检出率也高达95.6%。

To realize accurate detection of small-scale faces in natural scene,this paper constructs an improved Faster-RCNN model.The model uses ResNet-50 to extract convolution features,and performs multi-scale fusion for feature maps of different convolutional layers.At the same time,the number of Anchors generated by the Regional Proposal Network(RPN)has been changed from 9 to 15 to better adapt to the small-scale face detection scenes.On this basis,the Online Hard Example Mining(OHEM)algorithm is used to optimize the training process.Soft-Non-Maximum Suppression(Soft-NMS)method is used to reduce the missed detection of overlapping faces,and in the training phase the multiscale training method is adopted to improve the generalization ability of the model.Experimental results show that the average precision of the proposed model is 89.0%on the Wider Face dataset,which is 3.5%higher than that of the original Fast-RCNN model.The relevance ratio of the proposed model reaches 95.6%on the FDDB dataset.