嵌入式智能物体认知系统设计与实现  

Design and implementation of embedded intelligent object recognition system

作　　者：张建 张海飞[3] 史洪玮 ZHANG Jian;ZHANG Haifei;SHI Hongwei(Wenzheng College of Soochow University,Suzhou 215104,China;Soochow College,Soochow University,Suzhou 215006,China;School of Computer and Information Engineering,Nantong Institute of Technology,Nantong 226002,China;Suqian College,Suqian 223800,China)

机构地区：[1]苏州大学文正学院,江苏苏州215104 [2]苏州大学东吴学院,江苏苏州215006 [3]南通理工学院计算机与信息工程学院,江苏南通226002 [4]宿迁学院,江苏宿迁223800

出　　处：《现代电子技术》2022年第16期155-160,共6页Modern Electronics Technique

基　　金：国家自然科学基金项目:面向群智感知的高可靠数据收集与筛选关键技术研究(61672369);苏州大学文正学院教学改革课题(20WZJG0007,20WZJG0033,20WZJG0032)。

摘　　要：目前嵌入式图像识别系统对资源要求普遍偏高,且针对低配置资源的终端研究较少。为此,文中将深度学习中的卷积神经网络算法引入到嵌入式终端,提出一种嵌入式智能物体认知系统。在硬件设计上,通过优化摄像头采集图像流程、图像归一化处理来缩短终端对图像采集的时间。在软件实现上,采用双峰均值算法对图像进行预处理,将卷积神经网络算法训练的参数下发到终端,使终端能够高效地识别采集到的图像;再采用分散存储方式,将每层参数以常量数组的方式分别存储,从而大幅度降低推理过程中对资源的占用。通过对数字、字母及实物进行识别测试得出,文中系统识别准确率高于90%,表明在资源受限的终端也能对物体进行较好地识别。As the embedded image recognition system generally has high requirements for resources,and there is less research on the terminal with low configuration resources,the CNN(convolutional neural network)algorithm in deep learning is introduced into the embedded terminal,and an embedded intelligent object recognition system is proposed.In the hardware design,the time of image acquisition by the terminal is shorten by optimizing the camera image acquisition process and image normalization processing.In the software implementation,the bimodal mean algorithm is used to preprocess the image,and the training parameters of convolution neural network algorithm are sent to the terminal,so that the terminal can recognize the collected image efficiently.The decentralized storage method is used to store the parameters of each layer in the form of constant array,so as to greatly reduce the resource occupation in the reasoning process.The recognition test of numbers,letters and physical objects show that the recognition accuracy of the system can reach more than 90%,which indicates that the system can better recognize objects in the terminal with limited resources.

关 键 词：图像识别 神经网络 深度学习 双峰均值 图像采集 物体认知 人工智能 

分 类 号：TN915-34[电子电信—通信与信息系统]