提出一个新的基于轻量级注意力机制的网络框架。在YOLOv3主干网络的基础上,使用深度卷积和点卷积代替标准卷积设计特征提取网络,加快模型的训练,提高检测的速度,然后引入注意力机制模块进行模型速度和精度的权衡,最后通过增加多尺度提取更多网络层的特征信息,同时使用K-means++ 聚类算法进一步优化网络参数。实验结果表明,该方法可以显著提高人脸检测模型的性能,在Wider Face数据集上可以达到94.08%的准确率和83.97%的召回率,且平均检测时间只需0.022 s,相比原始YOLOv3算法提高了4.45倍。

This study proposes a new network framework based on a lightweight attention mechanism and the YOLOv3 backbone network. When designing the feature extraction network, the standard convolutions of the YOLOv3 backbone network are replaced using depthwise and pointwise convolutions, thereby accelerating the model training and increasing the detection speed. Next, the speed and accuracy of the model are weighted using an attention mechanism module. Finally, multiple-scale prediction layers are added to extract more feature information; simultaneously, the network parameters are optimized using the K-means++ clustering algorithm. In an experimental evaluation on face-detection performance, this method considerably improved the face-detection performance, achieving 94.08% precision and 83.97% recall on the Wider Face dataset. The average detection time is 0.022 s, which is 4.45 times higher than that of the original YOLOv3 algorithm.