吉林大学学报(信息科学版) ›› 2023, Vol. 41 ›› Issue (6): 1063-1071.

• • 上一篇    下一篇

资源受限 MCU 的轻量化部署策略和实现

吴 薇1a,1b , 阮 星1a,1b , 蔡闯华1a,1b , 刘长勇1a,1b , 刘彦秀2 , 王宜怀3   

  1. 1. 武夷学院 a. 数学与计算机学院; b. 认知计算与智能信息处理福建省高校重点实验室, 福建 武夷山 354300; 2. 山东女子学院 数据科学与计算机学院, 济南 252000; 3. 苏州大学 计算机科学与技术学院, 江苏 苏州 215006
  • 收稿日期:2022-11-23 出版日期:2023-11-30 发布日期:2023-12-01
  • 通讯作者: 王宜怀(1962— ), 男, 福建建阳人, 苏州大学教授, 主要从事嵌入式物联网研究, (Tel)86-512-65113107 E-mail:yihuaiw@ suda. edu. cn
  • 作者简介:吴薇(1978— ), 女, 福州人, 武夷学院副教授, 主要从事嵌入式物联网、 人工智能研究, (Tel)86-18650667699 (E-mail) amyweiwu@ 163. com
  • 基金资助:
    福建省自然科学基金计划资助项目(2022J011202); 福建省中青年教师教育科研基金资助项目(JAT190782); 认知计算与 智能信息处理福建省高校重点实验室开放基金资助项目(KLCCIIP2018104); 福建省社会科学基金资助项目(FJ2022X016)

Lightweight Deployment Strategy and Implementation of Resource-Constrained MCUs

WU Wei 1a,1b , RUAN Xing 1a,1b , CAI Chuanghua 1a,1b , LIU Changyong 1a,1b , LIU Yanxiu 2 , WANG Yihuai 3   

  1. 1a. College of Mathematics and Computer Science; 1b. Key Laboratory of Cognitive Computing and Intelligent Information Processing of Fujian Education Institutions, Wuyi University, Wuyishan 354300, China; 2. College of Data Science and Computing, Shandong Women ' s University, Jinan 252000, China; 3. School of Computer Science and Technology, Soochow University, Suzhou 215006, China
  • Received:2022-11-23 Online:2023-11-30 Published:2023-12-01

摘要: 为实现低资源嵌入式设备的图像分类识别, 针对能实现简单图像识别任务、对图像识别准确率要求不 高, 且要求低成本的场景, 将卷积神经网络(CNN: Convolutional NeuralNetwork)部署到资源受限的微控制器单元(MCU: Microcontroller Units)上。 首先提出一种在资源受限 MCU 上的轻量化部署策略: 为降低模型的参数量, 提出一种轻量化的神经网络算法; 为保证模型大小能适应有限的随机存取存储器(RAM: Random Access Memory), 提出了一种基于闪存(FLASH: Flash Memory)扇区的替存储算法。 其次, 在资源受限的嵌入式设备上部署该策略。 针对采集图像的质量和采集速度不匹配问题, 设计了摄像头外围电路; 对采集图像进行基于高斯分布的自适应阈值二值化处理并对图像样本完整性进行校验。 实验结果表明, 该系统取得大约 80% ~ 89% 的 识别准确率。 虽然该准确率低于训练精度 10% 左右, 但在上述对精度要求不高的实际场景中可以较好地应用。

关键词: 嵌入式系统, 资源受限 MCU, 图像识别, 深度学习

Abstract: This work aims to deploy a CNN onto resource-constrained MCUs(Microcontroller Units) to achieve image classification and recognition for scenarios that require simple image recognition tasks, low image recognition accuracy, and low cost. Firstly, a lightweight deployment strategy on resource-constrained MCUs is proposed. To reduce the number of model parameters, a lightweight neural network algorithm is proposed. To ensure that the model size can fit into limited RAM(Random Access Memory), a storage replacement algorithm is presented based on FLASH ( Flash Memory ) sectors. Secondly, the strategy on embedded devices is implemented. The camera peripheral circuit is designed for image quality, but the acquisition speed does not match. The collected images are binarized by an adaptive threshold based on Gaussian distribution and the integrity of image samples is verified. Experimental results show that the system can achieve better image classification and recognition accuracy when applied in the above practical scenarios. 

Key words: embedded system, resource-constrained microcontroller units(MCU), image discrimination, deep learning

中图分类号: 

  • TP391