基于多任务卷积神经网络人脸检测网络的优化加速方法

Optimization acceleration method for face detection network based on multi-task convolutional neural network

针对人脸检测网络在安卓设备上速度较慢、检测效果不佳的问题,提出了基于多任务卷积神经网络(MTCNN)的优化加速的方法。一方面,更改Caffe框架的ATLAS数学库为OpenBLAS,利用多核来进行矩阵计算;另一方面,在网络结构上进行优化,通过对特征图进行多尺度的卷积再聚合操作、对不同层的特征图进行特征融合操作来提高网络的学习能力,同时对网络中的卷积核进行分解来减少卷积操作的计算量。在LFW和FDDB数据集上进行测试,实验结果表明:与MTCNN人脸检测算法相比,改进后算法的准确率提高了0.9%,耗时降低了64%,在3288安卓板卡上速度达到了9帧/秒。改进后的算法对遮挡、模糊人脸有更好的鲁棒性,并且在速度上有了很大的提升,在复杂场景下具有出色的检测效果。

Focused on the slow speed and poor detection effect of face detection network on Android devices,an optimization acceleration method based on Multi-Task Convolutional Neural Network(MTCNN)was proposed.On the one hand,the ATLAS math library of the Caffe framework was changed to OpenBLAS,where multiple cores were used for matrix calculation;on the other hand,optimization was performed on the network structure.Multi-scale convolution re-aggregation operations on feature maps and feature fusion operation on feature maps of different layers were used to improve the learning ability of the network,and the convolution kernels in the network were decomposed to reduce the computational complexity of the convolution operation.The experimental results on LFW and FDDB datasets show that compared with the face detection algorithm based on MTCNN,the accuracy of the improved algorithm was improved by 0.9%,and the time consumption was reduced by 64%,and the speed on the 3288 Android board was 9 frames per second.The improved algorithm has better robustness to occlusion and fuzzy faces,and has a great improvement in speed,which has excellent detection effect in complex scenes.