基于卷积神经网络的金属检测系统设计

Design of Metal Detection System Based on Convolutional Neural Network

摘要: 为了解决传统金属检测系统受产品效应影响下误报率高、检测灵敏度低，检测失效等问题，本文通过参考经典卷积神经网络，提出一种专用于金属检测的MD-net卷积神经网络二分类模型，并设计了基于该分类模型的金属检测系统。该系统的下位机以FPGA为核心，实现了金属信号的测量和预处理。PC上位机在Windows系统环境下，实现了金属信号的特征提取和卷积神经网络的二分类识别。试验结果表明，即使在较强产品效应影响下，对于铁颗粒的最小检测精度为1.2 mm，检测准确率可达98.92%，该系统对金属异物具有较高的检测灵敏度和良好的检测性能。

Abstract: In order to solve the problems of traditional metal detection systems such as high false alarm rate, low detection sensitivity, and detection failure due to product effects, this paper proposes an MD-net convolutional neural network dedicated to metal detection by referring to the classic convolutional neural network. Based on the MD-net binary classification model, this paper designs a metal detection system for detecting metal foreign objects in food. The lower computer of this system uses FPGA as the core to realize the measurement and preprocessing of metal signals. The PC host computer realizes the feature extraction of metal signals and the two-class recognition of convolutional neural networks in the Windows system environment. The test results show that even under the influence of strong product effects, the minimum detection accuracy for iron particles is 1.2 mm, and the detection accuracy can reach 98.92%. The system has high detection sensitivity and good detection performance for metal foreign matter.