深度学习的移动端可视耳标检测识别研究  

Research on mobile visual ear tag detection and recognition based on deep learning

作　　者：李硕 陈春雨[1] 刘文龙 LI Shuo;CHEN Chunyu;LIU Wenlong(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001

出　　处：《应用科技》2022年第2期70-74,共5页Applied Science and Technology

基　　金：国家自然科学基金项目(61871142);民品横向项目(KY10800180032);中央高校基本科研业务费项目(3072020CFT0803)。

摘　　要：随着产业数字化发展,智能养殖技术也是其中重要一环。为了防控各类疾病、改良品种质量、肉制品质量溯源以及改善农业虚假假保险索赔现象,可通过移动式或固定式的智能终端设备进行实时、精准的智能化采集与录入大群体的生产数据,实现快速、精准确定猪只个体身份。本系统通过对猪只数据进行数据采集,在树莓派这一优秀的硬件基础之上,以轻量级的深度神经网络推理引擎移动神经网络(MNN)作为支撑,采用速度更快的MobileNet网络对猪只耳标数据进行检测,并且利用ZXing技术识别耳标快速反应(QR)二维码,完成对猪只个体身份的识别,最终完成移动端的猪只耳标检测、识别。同时为了更好地了解神经网络在耳标检测时关注图像的任意区域,使用类激活映射(CAM)的可解释性方法可视化卷积神经网络感兴趣区域。With the development of digital industry, intelligent breeding technology is also an important part. In order to prevent and control various diseases, improve variety quality, meat product quality traceability and improve agricultural fake insurance claims, real-time and accurate intelligent collection and input of large groups of production data through mobile or fixed intelligent terminal equipment, so as to quickly and accurately determine the individual identity of pigs.Based on the excellent hardware of raspberry pie, the system uses the lightweight deep neural network inference engine mobile neural network(MNN) as the support, uses the faster MobileNet to detect the pig ear mark data, and uses ZXing technology to identify the ear mark quick response(QR) two-dimensional code, Complete the identification of individual pig identity, and finally complete the mobile intelligent terminal device for pig ear mark detection and identification. At the same time, in order to better understand which region of the image the neural network pays attention to in ear mark detection, the interpretability method of class activation mapping(CAM) is used to visualize the region of interest of convolution neural network.

关 键 词：深度学习 树莓派 移动神经网络 移动端 QR二维码 身份识别 类激活映射 可解释性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]