监控场景下基于轻量级卷积神经网络的行人属性识别  

Pedestrian attribute recognition based on lightweight convolutional neural network in surveillance scene

作　　者：黄仝宇 胡斌杰[1] Huang Tongyu;Hu Binjie(South China University of Technology,Guangzhou 510641,China)

机构地区：[1]华南理工大学,广州510641

出　　处：《电子测量技术》2020年第14期118-122,共5页Electronic Measurement Technology

摘　　要：针对面向行人属性识别的传统卷积神经网络模型占用系统资源较多、运行效率较低等问题,通过采用若干个倒置残差模块来构建轻量级卷积神经网络MB-ResNet-Lite的主干网络,使用深度可分离卷积代替标准卷积,以减少模型的计算量;并通过倒置残差模块对特征的维度先扩增后缩减,实现更好地提取特征。为了解决多种行人属性识别的效率问题,所提方法是在共享主干网络之后,采用若干个分支网络进行各自特征的提取,以完成多种行人属性的分类与识别。最后,该算法在自建数据集、单核RK3399平台上进行比对实验。实验结果表明,所提算法MB-ResNet-Lite模型的大小为0.82 M,分别为ResNet-18和MobileNet的1.8%和6.3%,明显减小了模型的存储空间;在运行速度上,所提算法模型处理单张图片耗时为25 ms,分别为ResNet-18和MobileNet的18.4%和64.1%,较好地节省了图像处理时间;在内存使用方面,所提算法模型占用内存为21.56 MB,分别为ResNet-18和MobileNet的6.6%和60.0%,有效节省了系统的内存资源。在算法准确率方面,所提算法模型的平均准确率为89.24%,相比MobileNet提高了1.52%,相比ResNet-18略微下降0.86%。结果表明,所提方法有效地保证行人特征识别的精确度,减少模型的参数量和计算量,确保在低成本硬件平台的运行效率。In order to solve the problem that the traditional convolutional neural network model for pedestrian attribute recognition takes up a lot of system resources and has low efficiency,the algorithm model in this paper uses several inverted residual modules to build the backbone network of the lightweight convolutional neural network called MB-ResNet-Lite,and uses the depthwise separable convolution instead of the standard convolution to reduce the calculation amount of the model;and through the inverted residual block,The dimensions of the features are expanded first and then reduced to achieve better feature extraction.In order to solve the efficiency problem of pedestrian attribute recognition,this paper uses several branch networks to extract their own features after sharing the backbone network,so as to complete the classification and recognition of various pedestrian attributes.Finally,the algorithms are compared on a single core RK3399 platform with a self built data set.The experimental results show that the size of MB-ResNet-Lite is 0.82 M,which is 1.8%and 6.3%of ResNet-18 and MobileNet respectively,which significantly reduces the storage space of the model;in terms of running speed,the processing time of single image of the algorithm model is 25 ms,which is 18.4%and 64.1%of ResNet-18 and MobileNet respectively,which saves image processing time;in terms of memory usage,The memory occupied by the algorithm model in this paper is 21.56 MB,which is 6.6%and 60.0%of ResNet-18 and MobileNet respectively,which effectively saves the memory resources of the system.In terms of algorithm accuracy,the average accuracy of this algorithm model is 89.24%,which is 1.52%higher than MobileNet and just 0.86%lower than ResNet-18.The results show that this method can effectively ensure the accuracy of pedestrian attribute recognition,reduce the amount of parameters and calculation of the model,and ensure the operation efficiency of the low-cost hardware platform.

关 键 词：深度学习 轻量级卷积神经网络 深度可分离卷积 倒置残差 行人属性识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]