【摘要】 针对核桃生产线的异物检测需求，首先根据现有通用的核桃加工生产线结构特点，设计并搭建了一套核桃异物检测装备，该装备包括设备框架、图像采集系统和恒定光源系统，整体尺寸为470 mm×600 mm×615 mm。然后以浙江省杭州市核桃生产基地的核桃和实际生产加工中出现的树叶、树枝、石子、金属、塑料等异物为检测对象，通过工业相机实时采集生产线上的核桃图像，获取直观的图像信息数据。结合了深度学习与计算机视觉技术，利用基于全卷积神经网络(Fully convolutional networks, FCN)的算法进行图像边缘检测，对核桃生产加工中可能出现的异物进行了检测，并通过试验对其性能加以验证。结果表明，训练集检测准确率为92.75%,验证集准确率为90.35%,检测速率为4.28 f/s,满足生产线运输速度1 m/s的检测要求。该研究即使在样本量较少的情况下，仍然得到了较好的图像分割效果，可以实现核桃生产线的异物实时检测。更多还原

【Abstract】 Aiming to solve the needs of foreign matter detection in walnut production line, a set of walnut impurity detection equipment was designed and built based on the existing universal walnut processing production line, including portable frame, image acquisition system, and constant light source system. The overall size was 470 mm×600 mm×615 mm. Walnuts from Zhejiang Province and impurities, including leaves, stones, paper, screws and fabric were photographed as detection objects by industrial camera above the production line in real time for intuitive image information data. An image segmentation technology combined with deep learning and computer vision, and the fully convolutional network(FCN) algorithm were applied to detect impurities that might occur in walnut production and processing. According to the test, the accuracy for detection and classification of walnut and foreign body was effective, which was 92.75% of training set and 90.35% of testing set. The speed of production line was 1 m/s. The recognition speed of detecting was 4.28 f/s, which can meet the requirements of real-time detecting of impurities. The biggest error was in the “walnut-background”, where original walnut was predicted to be the background. The main reason was that some features in walnuts(such as cracks and lines) were similar to the background. Focusing on the analysis of foreign body error, it showed that impurities were mis-predicted as the “background” much more than the impurities were mis-predicted as the “walnut”. Two main reasons led to this difference. On the one hand, when labelling manually, the pollutants on the conveyor belt were not judged as foreign bodies. On the other hand, because the size of impurities was generally small and the cardinality of pixel points was insufficient, the influence of false prediction was greater, thus amplifying the error. The reliability of the model was good. Even if the artificial labeling error occurred, walnut was mislabeled as impurities, but the trained model could still distinguish walnut and adjacent impurities well. The method proposed was worthy of further study for the online detection of impurities in automatic production of walnut, and it was of great significance to broaden the market of nut food and improve its economic benefits.更多还原