面向图像检索的sgemv算法嵌入式优化技术  

Embedded Optimization Technology of Sgemv Algorithm for Image Retrieval

作　　者：郑恩 张翰成 周俊鹏 白林亭 文鹏程[1] ZHENG En;ZHANG Han-cheng;ZHOU Jun-peng;BAI Lin-ting;WEN Peng-cheng(Xi′an Aeronautics Computing Technique Research Institute,AVIC,Xi′an 710000,China)

机构地区：[1]航空工业西安航空计算技术研究所,陕西西安710000

出　　处：《航空计算技术》2024年第1期62-65,共4页Aeronautical Computing Technique

基　　金：航空科学基金项目资助(2022Z071031001)。

摘　　要：行人重识别主要解决跨摄像头跨场景下行人的识别与检索,是继人脸识别之后又一针对“人”的视觉任务,主要任务是针对一个特定的行人在多摄像头输入的大规模图片集合中找出相同的人。如何在靠近摄像头的边缘端把特定的行人从大量行人库中快速检索出来是行人重识别研究的一个重要问题,由于边缘端嵌入式平台算力有限,提出一种面向图像检索的sgemv算法嵌入式优化技术,在边缘端对sgemv算法采用循环展开、OpenMP、Neon等技术进行加速优化,在飞腾D2000嵌入式平台、银河麒麟系统进行实验验证。结果表明,优化后比优化前提升速度达5.2倍,方法有效地提升了边缘端图像检索效率。Pedestrian re identification mainly solves the recognition and retrieval of pedestrians in cross camera and cross scene situations.It is another visual task for"people"after face recognition.The main task is to find the same person for a specific pedestrian in a large scale image collection input by multiple cameras.How to quickly retrieve specific pedestrians from a large number of pedestrian databases at the edge near the camera is an important issue in pedestrian re identification research.Due to the limited computing power of the embedded platform at the edge,this paper proposes an image retrieval oriented sgemv algorithm embedding In the edge end,the sgemv algorithm is accelerated and optimized using technologies such as loop unrolling,OpenMP,and Neon,and the experimental verification is carried out on the Phytium D2000 embedded platform and the Galaxy Kirin system.The results show that,the speed after optimization is 5.2 times higher than that before optimization.This method effectively improves the efficiency of image retrieval at the edge end.

关 键 词：行人重识别 图像检索 循环展开 OPENMP NEON 

分 类 号：V19[航空宇航科学与技术—人机与环境工程]