针对高性能人脸检测模型参数量大、计算复杂度高,难以在嵌入式设备进行边缘部署的问题,对RetinaFace模型进行轻量化改进,提出一种轻量级人脸检测算法。采用MobileNetV2_0.5×作为特征提取骨干,应用轻量的PANLite对多尺度特征进行双向融合,增强特征表征能力。采用RFBLite实现特征增强,在增大特征感受野的同时合并上下文信息。使用滤波器剪枝算法对训练后的模型进行剪枝处理,再次训练微调网络参数后部署到嵌入式端Nvidia Jetson Nano进行模型推理。实验结果表明,该轻量级模型能够以较少的参数量和较低的计算复杂度实现较高的人脸检测性能,且能在嵌入式平台上进行实时推理。

Aiming at the problem that the high-performance face detection model has a large number of parameters,high computational complexity,and it is difficult to deploy on the edge of embedded devices,a lightweight face detection algorithm is proposed,which is based on the lightweight improvement on RetinaFace.Firstly,MobileNetV2_0.5× is used as the backbone to extract features,and PANLite with few parameters is selected for two-way fusion of multi-scale features to enhance the ability of feature representation.RFBLite is used as the feature enhancement module to incorporate context information while increasing the feature receptive field.Finally,filter pruning algorithm is applied in the trained model.Then the pruned model is retrained for fine-tuning and depolyed to Nvidia Jetson Nano for model inference.The experiment results show that the lightweight model can achieve high face detection performance with a small amount of parameters and low computational complexity,and can perform real-time inference on embedded platforms.