一种改进型RetinaFace的遮挡人脸检测算法

Modified occlusion face detection algorithm for RetinaFace

针对遮挡人脸检测中检测模型较大,检测速度较慢的问题,提出了一种新的改进型RetinaFace人脸检测算法。在RetinaFace模型的基础上,采用卷积神经网络MobilenetV3作为特征提取网络,通过注意力机制CANet和加权双向特征金字塔加强特征提取,优化RetinaFace算法的损失函数使模型更加关注人脸位置。依据Wider-Face和MaskFace数据集建立了实验数据集,并通过实验验证了算法的有效性。

For the problem of large detection model and slow detection speed in occlusion face detection,a novel face detection algorithm for a modified RetinaFace is proposed.Based on the RetinaFace model,the convolutional neural network MobilenetV3 is used as the feature extraction network to strengthen the feature extraction through the attention mechanism CANet and the weighted bidirectional feature pyramid,and optimize the loss function of the RetinaFace algorithm to make the model pay more attention to the face position.The experimental dataset is established based on the Wider-Face and MaskFace datasets,and the experimental results show the effectiveness of the algorithm.