适用于嵌入式平台的E-YOLO人脸检测网络研究

针对现有人脸检测深度学习算法计算量大,难以移植到嵌入式平台,无法满足移动设备实时性和便捷性需求的问题,提出一种基于YOLO(You Only Look Once)算法的适用于嵌入式平台的小型人脸检测网络E-YOLO(Enhance-YOLO)。借鉴YOLO算法的思想,将人脸检测问题转换为回归问题,将待检测的图像均分为S×S个单元格,每个单元格检测落在单元格内的目标。通过修改YOLO网络模型中的卷积神经网络结构,提高其检测的准确性,同时减少网络结构中卷积核的数目,降低模型的大小。实验结果表明,E-YOLO模型大小为43MB,视频的检测帧率为26FPS,在WIDERFACE和FDDB数据集上均有较高的准确率和检测速度,可以实现在嵌入式平台下的实时人脸检测。

E-YOLO FACE DETECTION NETWORK FOR EMBEDDED PLATFORM

Aiming at the problems of the existing deep learning algorithms for face detection,such as the large amount of computation,difficulty in porting to the embedded platform,and inability to meet the real-time and convenient requirements of mobile devices,this paper proposes a small face detection network E-YOLO(Enhance-YOLO)based on the algorithm of YOLO(You Only Look Once),which is suitable for embedded platform.Using for reference of YOLO algorithm,face detection can be regarded as a regression problem.The image to be detected was divided into S×S cells and each cell detected the target fell in them.By modifying the convolutional neural network structure in the YOLO network model,the accuracy of its detection was improved.In the same time,the number of convolution kernels in the network structure was reduced,and the size of the model was reduced.The experimental results show that the size of the E-YOLO model is 43M and the frame rate of video detection is 26 FPS.It has high accuracy and detection speed in both WIDERFACE and FDDB data sets,which can realize the real-time face detection on the embedded platform.