基于昇腾AI处理器的轻量化MNG-YOLO模型研究  

Research on Lightweight MNG-YOLO Model Based on Ascent AI Processor

作　　者：赵月爱[1] 沈帅杰 王智瑜 王玲[2] ZHAO Yueai;SHEN Shuaijie;WANG Zhiyu;WANG Ling(School of Computer Science and Technology,Taiyuan Normal University,Jinzhong Shanxi 030602,China;School of Automation and Software,Shanxi University,Taiyuan Shanxi 030006,China)

机构地区：[1]太原师范学院计算机科学与技术学院,山西晋中030602 [2]山西大学自动化与软件学院,山西太原030006

出　　处：《电子器件》2024年第5期1193-1200,共8页Chinese Journal of Electron Devices

基　　金：国家社科基金项目(20BJL080);山西省重点研发计划项目(201803D121088);太原师范学院研究生教育教学改革研究课题项目(SYYJSJG-2153);太原师范学院研究生教育创新项目(SYYJSYC-2398)。

摘　　要：随着目标检测神经网络算法精度不断提升,算法的参数量以及计算量都有着较高的增长,导致模型实际应用部署困难,因此对神经网络模型进行轻量化,减少模型的参数量和计算量对模型部署在边缘设备上是至关重要的。昇腾AI处理器是华为推出的一款专用于神经网络加速的芯片,为充分发挥昇腾AI处理器的优势并解决算法模型较为庞大的问题,基于此平台提出一种轻量化目标检测模型MNG-YOLO,对YOLO模型采用轻量级主干网络和Ghost卷积以减小模型大小,添加NAM注意力模块和Mish激活函数提升模型准确率。实验结果表明,MNG-YOLO模型相比于原始模型参数量以及计算量均减少约75%,参数量从7 015 519个减少至1 739 799个,计算量从15.8 GFLOPs减少至3.5 GFLOPs,模型精确度也由95.9%提升至97.5%。同时,在昇腾AI处理器上的推理速度达到205 FPS,远超实时性检测的速度要求。With the continuous improvement of the accuracy of object detection neural network algorithms, the number of parameters and computations of the algorithm has increased rapidly, which makes it difficult to deploy the model in practical applications. Therefore, it is crucial to make the neural network model lightweight and reduce the number of parameters and computations of the model to be deployed on edge devices. The Ascend AI processor is a chip dedicated to neural network acceleration launched by Huawei, in order to give full play to the advantages of the Ascend AI processor and solve the problem of large algorithm models, the MNG-YOLO object detection lightweight model is improved based on this platform, a lightweight backbone network and Ghost convolution are used to reduce the size of the YOLO model, and the NAM attention module and the Mish activation function are improved by adding the accuracy of the model. The experimental results show that compared with the original model, the MNG-YOLO model reduces the number of parameters and computations by about 75%,the number of parameters is reduced from 7 015 519 to 1 739 799,the number of computations is reduced from 15.8 GFLOPs to 3.5 GFLOPs, and the accuracy of the model is also improved from 95.9% to 97.5%. At the same time, the inference speed on the Ascend AI processor reaches 205 FPS,far exceeding the speed requirements of real-time detection.

关 键 词：目标检测 YOLO模型 昇腾AI处理器 模型轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]