【摘要】 目前，很多人出入公共场所仍须佩戴口罩。检测是否佩戴口罩变得尤为重要，而深度学习算法能够大幅度提高检测速度。本文依据ANN注意力机制结合特征网络改进得到PSA(Path Strengthen Integration ANN)多尺度特征融合注意力模块，进而形成最终的PSA-Retina口罩识别网络，其中骨干网络基于ResNet-50融合空间金字塔池化；采用优化的GFocal Loss损失函数；融合GELUs激活函数重做预测器，并在口罩人脸识别数据集RMFD上进行对比实验，融合PSA模块的网络比原网络的平均精度均值mAP高5.24%,每秒传输帧数FPS高3.1 f/s,比YOLOv3网络mAP高3.06%,FPS高0.6 f/s,实验数据表明，多特征融合注意力的PSA-Retina人脸口罩识别网络定位更准，准确率更高，具备在有遮挡或者非标准佩戴等情况下的检测能力，提升口罩识别效率。更多还原

【Abstract】 At present, many people are still required to wear masks when entering and leaving public places. Detecting whether a mask is worn has become particularly important, and deep learning algorithms can greatly improve the detection speed. Based on the ANN attention mechanism and feature network improvement, this paper obtains the Path Strength Integration ANN(PSA) multi-scale feature fusion attention module, and then forms the final PSA-Retina mask recognition network. In the design, the backbone network is based on ResNet-50 fusion spatial pyramid pooling; the optimized GFocal Loss function is used, GELUs activation function is fused to redo the predictor, and a comparative experiment is conducted on the mask face recognition data set RMFD. The simulation shows that the average accuracy mAP of the network added to the PSA module is 5.24% higher than the original network, and the number of frames per second FPS is 3.1 f/s higher, mAP is 3.06% higher and FPS is 0.6 f/s higher than the results of YOLOv3 network. The experimental data demonstrates that the multi-feature fusion attention PSA-Retina face mask recognition network has more accurate positioning and higher accuracy, and has the detection ability in the case of occlusion or non-standard wearing, etc., therefore improves the efficiency of mask recognition.更多还原