资源受限MCU的轻量化部署策略和实现  

Lightweight Deployment Strategy and Implementation of Resource-Constrained MCUs

作　　者：吴薇 阮星[1,2] 蔡闯华 刘长勇[1,2] 刘彦秀 王宜怀 WU Wei;RUAN Xing;CAI Chuanghua;LIU Changyong;LIU Yanxiu;WANG Yihuai(College of Mathematics and Computer Science,Wuyi University,Wuyishan 354300,China;Key Laboratory of Cognitive Computing and Intelligent Information Processing of Fujian Education Institutions,Wuyi University,Wuyishan 354300,China;College of Data Science and Computing,Shandong Women's University,Jinan 252000,China;School of Computer Science and Technology,Soochow University,Suzhou 215006,China)

机构地区：[1]武夷学院数学与计算机学院,福建武夷山354300 [2]武夷学院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 [3]山东女子学院数据科学与计算机学院,济南252000 [4]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《吉林大学学报（信息科学版）》2023年第6期1063-1071,共9页Journal of Jilin University（Information Science Edition）

基　　金：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)。

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。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.

关 键 词：嵌入式系统 资源受限MCU 图像识别 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]