Since small metal cylindrical shell (MCS) is a kind of very important metal object widely used in structure engineering and weapon production, especially in the manufacture of bullets, it is necessary to assure the high-precision surface of MCS. While the detection of MCS is generally done manually. In this paper, a novel automatic defect detection system for MCS is built using transfer learning of Inception-v3 and logistic regression (LR). By using the powerful feature extraction capabilities of Inception-v3 deep convolutional neural network, features are fetched from MCS images firstly and then trained on an LR machine learning classifier to establish a detection model. During the process of detection, five images of one MCS captured by the camera are sent to the computer for detection using the established detection model, with these five images’ composite outcomes representing this MCS’s detection result. Experimental results show that the proposed detection system could reach an accuracy of 97%, meeting the requirements of industrial production.

中文翻译：

---

使用转移学习和逻辑回归自动检测小型金属圆柱壳

由于小型金属圆柱壳（MCS）是一种非常重要的金属物体，广泛用于结构工程和武器生产，特别是在子弹制造中，必须保证MCS的高精度表面。而 MCS 的检测一般是手动完成的。在本文中，使用 Inception-v3 和逻辑回归 (LR) 的迁移学习构建了一种新的 MCS 自动缺陷检测系统。利用Inception-v3深度卷积神经网络强大的特征提取能力，首先从MCS图像中提取特征，然后在LR机器学习分类器上训练建立检测模型。在检测过程中，将摄像头拍摄的一个MCS的五幅图像发送到计算机进行检测，使用建立的检测模型，这五个图像的合成结果代表了这个 MCS 的检测结果。实验结果表明，所提出的检测系统可以达到97%的准确率，满足工业生产的要求。