基于深度可分离的多尺度Lw-YOLO轻量化人脸检测网络

MULTI-SCALE LW-YOLO LIGHTWEIGHT FACE DETECTION NETWORK BASED ON DEPTHWISE SEPARABILITY

针对目前神经网络模型计算复杂,在无GPU嵌入式平台的模型检测精度较低的问题,提出一种可在树莓派3B+上检测的轻量化人脸检测网络Lw-YOLO(Lightweight-YOLO)。此网络基于YOLO-LITE模型,利用深度可分离卷积替代传统卷积,有效地减少网络计算量并提升网络深度;增加多尺度预测模块,为预测层提供丰富的语义信息,提高网络精度。实验结果表明,训练得出的网络模型大小只有3.1 MB,在WiderFace人脸数据集上取得77.13%的平均精度,比原模型高23.22%,更适合无GPU的嵌入式平台。

In view of the fact that the current neural network model is computationally complex and its model detection accuracy on the embedded platform without GPU is low, a lightweight face detection network Lw-YOLO(Lightweight-YOLO) which can be detected on the Raspberry Pi 3B+ is proposed. This network was based on the YOLO-LITE model and used depthwise separable convolutions to replace traditional convolutions, which effectively reduced the amount of network calculations and increased the depth of the network. A multi-scale prediction module was added to provide rich semantic information to the prediction layer and improve network accuracy. The experimental results show that the size of the trained network model is only 3.1 MB. The average accuracy of 77.13% on the WiderFace face dataset is 23.22% higher than the original model, which is more suitable for embedded platforms without GPU.