改进的轻量化人脸识别算法

Improved lightweight face recognition algorithm

嵌入式平台计算资源有限,无法实时运行计算量和参数量巨大的深度学习模型。基于mobilenet v2提出一种改进的轻量化人脸识别算法L-mobilenet v2,首先对原有网络结构进行优化,然后以三元损失函数为主,将传统分类任务中的softmax损失改为Am-softmax作为辅助损失函数,使用10575个人的49万张图片进行联合训练。相比于改进前的模型及训练方法,新模型在LFW测试集和自制数据集的识别准确率达到98.56%和95%,将模型参数量缩减72.3%的同时将识别准确率提高了1.56%和7.1%,在嵌入式平台Jetson nano上的平均识别帧率提升了36.3%。该模型可以在计算资源受限的移动端实时运行。

Due to the limited computing resources,the embedded platform can not run the deep learning model with huge amounts of calculation and parameters in real time.An improved lightweight face recognition algorithm L-mobilenet v2 is proposed based on the mobilenet v2.The algorithm first optimizes the original network structure,then uses the triplet loss function as the main loss function to change the softmax loss in traditional classification task to Am-softmax,which is used as the auxiliary loss function,and uses 490 thousand images of 10575 people for joint training.Compared with the previous model and training method,the recognition accuracy of the new model on LFW test data set and self-made data set has reached 98.56%and 95%,respectively,which increased the recognition accuracy by 1.56%and 7.1%while reducing the number of model parameters by 72.3%.And at the same time,the rate of frame recognition on average on the embedded platform Jetson nano has increased by 36.3%.The model can run in real time on mobile terminals with limited computing resources.