基于昇腾NPU的目标检测加速技术评估  

Evaluation of Object Detection Acceleration Technology Based on Ascend NPU

作　　者：刘秋发 

机构地区：[1]江苏自动化研究所,江苏连云港222000

出　　处：《工业控制计算机》2025年第4期108-108,136,共2页Industrial Control Computer

摘　　要：针对目标检测模型部署的实际要求,评估国产智能芯片作为加速处理单元的可行性。选择昇腾310智能芯片和YOLOv4模型,在不改变神经网络模型结构以保证检测精度的前提下,将目标检测处理流程转移到昇腾神经处理单元上进行,以提高检测速度。结果表明,借助昇腾310的高性能算力与图像处理模块,可有效降低单张图片平均检测耗时,满足实时检测要求。Aiming to the actual requirement of the deployment of the object detection model,the feasibility of the domestic intelligent chip as an accelerated processing unit is evaluated in this paper.Without changing the neural network model structure to ensure the detection accuracy,the Ascend 310 intelligent chip and YOLOv4 model are selected.In order to improve the detection speed,the object detection process is transferred to the Ascend neural processing unit.The result shows that with the help of the high-performance computing power and digital vision processing module,the average detection cost of a single image is reduced effectively and the real-time detection requirement can be met.

关 键 词：目标检测 智能芯片 检测加速 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]