嵌入式人脸识别实验系统的模型压缩及其分析  

Model compression of embedded face recognitionexperimental system and its analysis

作　　者：陈利民[1] 严伟豪 梁音[1] 李春泉[1] CHEN Limin;YAN Weihao;LIANG Yin;LI Chunquan(School of Information Engineering,Nanchang University,Nanchang 330031,China)

机构地区：[1]南昌大学信息工程学院,江西南昌330031

出　　处：《南昌大学学报（理科版）》2021年第6期601-606,共6页Journal of Nanchang University(Natural Science)

基　　金：国家自然科学基金资助项目(61863028);江西省高等学校教学改革研究课题(JXJG-20-1-32)。

摘　　要：现有国产化嵌入式人工智能计算平台运算能力有限,不能满足部署人脸识别模型等高复杂度模型的要求。在DarkNet19网络中引入跳连接、模型剪枝,提出了一种基于深度卷积网络压缩的嵌入式人脸识别模型。采用8比特定点量化压缩,进一步实现了该网络模型的轻量化。在国产化嵌入式人工智能实验平台K210上,分别部署DarkNet19、DarkNet53和MobileNet-v1的压缩模型,对比其识别性能。同时分析了模型剪枝、精细化模型、模型量化三种压缩方式对上述三种模型的影响。实验结果表明,基于跳连接的DarkNet19压缩模型在模型参数规模、准确率和推理耗时上有更好的综合表现。The existing domestic embedded artificial intelligence computing platform has limited computing power and cannot meet the requirements of high-complexity face recognition models.An embedded face recognition model based on deep convolution network compression is proposed by introducing skip-connection and model pruning into DarkNet19 network.The 8-bit point quantization compression further realizes the lightweight of the network model.On the domestic embedded artificial intelligence experimental platform K210,the compression models of DarkNet19,DarkNet53 and MobileNet-v1 were deployed respectively,and their recognition performance was compared and studied.At the same time,the influence of model pruning,refined model and model quantization on the above three models were analyzed.The experimental results show that the DarkNet19 compression model based on skip-connection has a better comprehensive performance in parameter scale,accuracy and time-consuming.

关 键 词：模型压缩 跳连接 模型剪枝 模型量化 人脸识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]