A lightweight convolutional neural network model optimization method for ARM platform

With the optimization and upgrading of the ARM architecture and the popularization of mobile smart devices, the application requirements of deep learning technology in real life are increasing. It is a big problem to run the integrated neural network model on the ARM platform. In this paper, aiming...

Full description

Saved in:
Bibliographic Details
Published in2022 International Conference on Computer Engineering and Artificial Intelligence (ICCEAI) pp. 498 - 502
Main Authors Wang, Xuqiang, Zheng, Jian, Jin, Yao, Yang, Yifan, Zheng, Yang
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.07.2022
Subjects
Analytical models
ARM
Computational modeling
Convolution
lightweight network
Object detection
Performance evaluation
Program processors
Quantization (signal)
YOLOv3
Online AccessGet full text

Cover

More Information
Summary:With the optimization and upgrading of the ARM architecture and the popularization of mobile smart devices, the application requirements of deep learning technology in real life are increasing. It is a big problem to run the integrated neural network model on the ARM platform. In this paper, aiming at the problems existing in the deployment of the target detection lightweight network on the ARM mobile terminal, the network model is improved and the convolution parallel computing is optimized, and an effective inference acceleration scheme is proposed. In this paper, the YOLOv3 target detection algorithm based on one-stage is used for multi-scale target feature detection, and the YOLOv3 network model is deeply studied and improved. The theoretical acceleration effect of the depthwise separable convolution structure adopted in the classic lightweight network Mobile Net V1 is analyzed. In order to reduce the amount of network parameters, an improved Mobile Net_YOLOv3 model is proposed, Mobile Net V1 is selected to replace the backbone network Darknet53 in the original algorithm, and this model is accelerated for forward reasoning and model quantization deployment.
DOI:10.1109/ICCEAI55464.2022.00109