边缘优化下的戴口罩人脸实例分割

Face Instance Segmentation with Mask Wearing Under Edge Optimization

传统基于人脸关键点的人脸检测方法,在戴口罩的人脸识别中由于口罩大面积遮挡精度大幅下降,而分割算法在边缘特征上误检率较高。为了更精准地定位出戴口罩人脸的有效区域,本文提出了一种融合多阶段边缘分割算法的戴口罩人脸精细实例分割方法。一方面,在掩模区域卷积网络(Mask Region Convolutional Neural Networks,Mask RCNN)中添加了可拆分注意力的思想,增强了不同尺寸特征间的联系和检测性能;另一方面,使用双线性差分孪生网络搭建遮挡字典结合多阶段边缘优化分割方法优化目标边缘信息,逐步实现精细分割。在RMFD戴口罩人脸数据集上进行训练和测试。实验结果表明,RS_Mask R-CNN+边缘分割模型相比于基准Mask R-CNN的平均像素精确度(Pixel Accuracy,MPA)提高了4.11个百分点,目标检测精度达99%以上,在自遮挡以及不同类型口罩的人脸图像上都能取得不错的分割效果。

The traditional face detection method based on face key points has a significant decrease in accuracy in face recognition of mask-wearing faces due to large mask occlusion,while the segmentation algorithm has a high false detection rate on edge features.In order to locate the effective region of mask-wearing faces more accurately,this paper proposes a fine instance segmentation method for mask-wearing faces that incorporates a multi-stage edge segmentation algorithm.On the one hand,the idea of splittable attention is added to Mask Region Convolutional Neural Networks(Mask RCNN)to enhance the connection and detection performance between features of different sizes;on the other hand,a bilinear difference twin network is used to build a masking dictionary combined with a multi-stage edge optimization segmentation method Optimize the target edge information to achieve fine segmentation step by step.Training and testing are performed on the RMFD mask-wearing face dataset.The experimental results show that the RS_Mask R-CNN+edge segmentation model improves the average pixel accuracy(MPA)by 4.11 percentage points compared with the baseline Mask R-CNN,and the target detection accuracy reaches more than 99%,and achieves good segmentation results on self-masked and different types of masked face images.