Journal of Computer Applications ›› 2023, Vol. 43 ›› Issue (3): 744-751.DOI: 10.11772/j.issn.1001-9081.2022020252

• Artificial intelligence • Previous Articles    

Efficient person search algorithm and optimization with Sophon SC5+ chip architecture

Jie SUN1, Shaoxin WU1, Xuejun WANG2, Jing HUA1()   

  1. 1.School of Computer and Information Engineering,Zhejiang Gongshang University,Hangzhou Zhejiang 310018,China
    2.Shenzhen Xinghuo Electronic Engineering Company,Shenzhen Guangdong 518001,China
  • Received:2022-03-04 Revised:2022-05-27 Accepted:2022-05-30 Online:2022-08-16 Published:2023-03-10
  • Contact: Jing HUA
  • About author:SUN Jie, born in 1985, M. S., intermediate experimentalist. His research interests include image and video processing, visual analysis.
    WU Shaoxin, born in 1997, M. S. candidate. His research interests include image and video processing, machine learning.
    WANG Xuejun, born in 1969. His research interests include machine learning.
  • Supported by:
    National Natural Science Foundation of China(61972353)

基于Sophon SC5+芯片构架的行人搜索算法与优化

孙杰1, 吴绍鑫1, 王学军2, 华璟1()   

  1. 1.浙江工商大学 计算机与信息工程学院,杭州 310018
    2.深圳市星火电子工程公司,广东 深圳 518001
  • 通讯作者: 华璟
  • 作者简介:孙杰(1985—),男,浙江杭州人,中级实验师,硕士,CCF会员,主要研究方向:图像与视频处理、可视化分析
    吴绍鑫(1997—),男,浙江温州人,硕士研究生,主要研究方向:图像与视频处理、机器学习
    王学军(1969—),男,内蒙古赤峰人,主要研究方向:机器学习
    华璟(1974—),男,浙江杭州人,教授,博士,主要研究方向:计算机图形学、数据可视化、医学图像分析。
  • 基金资助:
    国家自然科学基金资助项目(61972353)

Abstract:

The computational costs of traditional deep neural network-based person search algorithms are very high, so that these algorithms are difficult to deploy on devices with limited hardware resources and budgets because of high cost and low speed. Aiming at the above problems, a person detection and person re-identification algorithm based on the high-performance inference chip Sophon SC5+ was proposed to optimize the efficiency of deep learning from the algorithm end to the hardware end in a top-down approach. Firstly, by using the lightweight Ghost module to replace the backbone network of YOLOv5s, the parameters and computational cost of the model were greatly reduced. Secondly, Convolutional Block Attention Module (CBAM) attention mechanism was integrated to enhance the feature learning capability and improve the detection precision of the algorithm. Thirdly, the central loss constraint and Non-local attention mechanism were added to the person re-identification module, and the central constrained triple loss and the additional interval cross-entropy loss were combined to optimize the model and improve the performance of the person re-identification algorithm. Finally, based on Sophon SC+, person detection model and person re-identification model were quantized and the final inference model was generated. Experimental results on Market-1501 and DukeMTMC-ReID datasets show that, the mean Average Precisions (mAPs) of the person detection and person re-identification algorithms were improved by at least 43.8 and 25.7 percentage points compared with YOLOv4-tiny, Attribute-Complementary Re-ID Net (ACRN), Singular Vector Decomposition Net (SVDNet) and other mainstream algorithms. After the implementation of int8 quantization based on Sophon SC5+ chip, although the proposed algorithm has the mAP decreased by 1.7 percentage points, it has the model size reduced by 74.4%. It can be seen that the proposed algorithm can be used in large-scale, city-level person search systems.

Key words: person re-identification, person search, Ghost module, central loss, Sophon SC5+, attention mechanism

摘要:

传统的基于深度神经网络的行人搜索算法计算量大,在大规模部署时搜索性能低,导致算法在落地应用于硬件和预算有限的终端时面临成本高、速度慢的难题。针对以上问题,提出一种基于Sophon SC5+高性能推理芯片的行人检测与重识别算法,从算法到硬件自上而下地优化深度学习的效率。首先,利用轻量化的Ghost模块替换YOLOv5s的主干网络,从而大幅度降低模型的参数和计算量;其次,融入CBAM注意力机制,以增强算法的特征学习能力,并提高检测精度;然后,将中心损失约束和 Non-local注意力机制加入行人重识别模块,并结合中心约束三元组损失和附加间隔交叉熵损失优化模型,以提升行人重识别算法性能;最后,基于Sophon SC+量化行人检测模型和行人重识别模型并生成最终的推理模型。在Market-1501与DukeMTMC-ReID数据集上的实验结果表明,相较于YOLOv4-tiny、ACRN、SVDNet等主流算法,行人检测算法与行人重识别算法的平均精度均值(mAP)至少提高了43.8和25.7个百分点。基于Sophon SC5+芯片实现int8量化后,所提算法的mAP虽然减小了1.7个百分点,但模型大小减小了74.4%,能够在大规模、城市级行人搜索系统中落地使用。

关键词: 行人重识别, 行人搜索, Ghost模块, 中心损失, Sophon SC5+, 注意力机制

CLC Number: