基于注意力机制的多尺度实时人脸检测方法

Multi-Scale Real-Time Face Detection Method Based on Attention Mechanism

在实际工程应用开发中,多数人脸检测算法存在实时性低,模型参数量大等问题。为了提升算法的实用性,提出一种轻量型、多尺度、基于注意力机制、实时的onestage人脸检测方法。通过使用深度可分离卷积将传统的卷积核分为一个深度卷积核和点卷积核,提高算法的检测效率;使用逆残差结构和注意力机制,提升算法的精度;构建特征金字塔进行特征融合获得丰富的上下文信息并可对多尺度目标进行独立检测;通过增加人脸关键点回归任务,提升框体回归的精度。实验结果表明,算法在FDDB数据集上的准确率达97.22%,并能达到实时检测的要求。

In actual engineering application development,most face detection algorithms have problems such as low real-time performance and large model parameters.In order to improve the practicality of the algorithm,a lightweight,multi-scale,attention-based,real-time one-stage face detection method is proposed.By using deep separable convolution,the traditional convolution kernel is divided into a deep convolu⁃tion kernel and point convolution kernel to improve the detection efficiency of the algorithm;using inverse residual structure and attention mechanism to improve the accuracy of the algorithm;building a feature pyramid Perform feature fusion to obtain rich contextual information and independently detect multi-scale targets;increase the accuracy of the frame regression by increasing the face key point regression task.Experimental results show that the accuracy of the algorithm on the FDDB data set is 97.22%,and it can meet the requirements of realtime detection.