Core failure inspection is an important issue in die casting. The inspection process is often carried out by manually examining X-ray images. However, human visual inspection suffers from individual biases and eye fatigues. Computer-vision-based automatic inspection, if it can achieve equal to or better than human performance, is favored to assist the inspectors to achieve better quality control. Most existing works are heavily relied on the supervised methods, which require enormous labeling and cannot be deployed quickly and economically. This is particularly difficult for a die casting plant that has many different types of products. Labeling each type of product before applying automated inspection may not be feasible in practice. It is therefore necessary to investigate unsupervised methods for die casting products. In this research, an inspection framework built on top of convolutional autoencoder (CAE) is designed and developed to inspect core failures from real-world die casting X-ray images in an unsupervised manner. Identification of good and scrap product, and localization of the defect are achieved in a single network. The framework is designed to be easily generalized to other image inspection scenarios. The area of interest for inspection is first extracted automatically through the Hough transformation. Then the preprocessed image is inspected by CAE. The noises of the model are removed using edge detection. It achieved an impressive 97.45% classification accuracy on average, and precisely pinpointed the defect regions with a small training set of 30 images.

型芯失效检查是压铸中的一个重要问题。检查过程通常通过手动检查 X 射线图像来执行。然而，人类视觉检查会受到个人偏见和眼睛疲劳的影响。基于计算机视觉的自动检测，如果能够达到等于或优于人类的表现，则更有利于辅助检测人员实现更好的质量控制。大多数现有工作严重依赖于监督方法，这需要大量标记并且无法快速、经济地部署。这对于拥有多种不同类型产品的压铸厂来说尤其困难。在应用自动检查之前标记每种类型的产品在实践中可能不可行。因此，有必要研究压铸产品的无监督方法。在这项研究中，一个建立在卷积自动编码器 (CAE) 之上的检查框架被设计和开发，以无监督的方式从真实世界的压铸 X 射线图像中检查核心故障。在单个网络中实现良品和废品的识别以及缺陷的定位。该框架旨在轻松推广到其他图像检查场景。首先通过霍夫变换自动提取要检查的感兴趣区域。然后由CAE检查预处理的图像。使用边缘检测去除模型的噪声。它平均实现了令人印象深刻的 97.45% 的分类准确率，并用 30 张图像的小型训练集精确定位了缺陷区域。