基于嵌入式系统的多任务人脸属性估计算法

针对传统人脸属性估计算法算力大、推理速度慢、精度低,难以完成算法在移动或嵌入式设备上集成应用等问题,提出一种基于嵌入式系统的多任务人脸属性估计算法。首先,采用MobileFaceNet网络中的瓶颈结构融合跨阶段融合网络(cross stage partial network, CSPNet)和空间金字塔网络(spatial pyramid pooling network, SPPNet)设计CSPSPP＿bk结构作为人脸属性估计算法共享网络特征提取模块;然后,在局部属性中增加通道注意力机制,在较困难的全局属性中使用更深、性能更优的网络模型作为Teacher模型指导所设计的轻量级多任务属性网络进行知识蒸馏,采用逐层剪枝的方法对网络模型进行优化,优化后的模型量仅1.8 MB;最后,通过动态类别抑制损失函数进行损失度量,均衡样本数据分布。在公共数据集CelebA和Adience数据集上进行测试比较,性别和眼镜的平均准确率分别为98.89%、99.72%,标准差为3.01%时,年龄估计精度为60.21%,在RK3288开发板上的前传推理速度为138 fps。结果表明：所提方法可广泛应用于嵌入式...

Multi-tasking face attribute estimation algorithm based on embedded system

Traditional face attribute estimation algorithms have large computing power, slow reasoning speed and low accuracy, which make it difficult to integrate the algorithm into mobile or embedded devices. A multi-task face attribute estimation algorithm based on embedded system was proposed. Firstly, the bottleneck structure of MobileFaceNet network was integrated with cross stage partial network (CSPNet) and spatial pyramid pooling network (SPPNet) to design the CSPSPP_BK structure as the face attribute estimation algorithm to share the network feature extraction module. Then, in the local properties increase channel attention mechanism, using the network in a difficult global properties deeper and better performance of network model as a guide for the proposed model, the design of lightweight and multi-tasking attribute network knowledge distillation, the method of layered pruning was adopted to optimize the network model, the optimized model of quantity is only 1.8 MB. Finally, the dynamic category inhibition loss function was used to measure the loss and equalize the distribution of sample data. Compared with CelebA and Adience data sets, the average accuracy of gender and glasses is 98.89% and 99.72%, respectively. When the standard deviation is 3.01%, the accuracy of age classification is 60.21%, and the pretransmission reasoning speed on RK3288 development board is 138 fps. The results show that the proposed method can be widely applied to embedded devices and mobile edge devices.