<p>Self-supervised industrial image anomaly detection holds significant practical value. Although existing anomaly detection models perform well on many public benchmark datasets, their large computational overhead and inability to effectively detect logical anomalies hinder their widespread application in real-world environments. To address this issue, we propose a global–local feature autoencoder (GLAE) for industrial logical anomaly detection, which can be flexibly integrated in parallel into a lightweight teacher–student network. GLAE employs a novel global feature extractor as the encoder (GFencoder), which introduces one-dimensional convolutions in two different directions to extract global semantic information from images, followed by two-dimensional convolutions for local feature constraints. This design significantly reduces computational costs while reconstructing global–local information of images. During the training phase, both the student network and GLAE are trained using only normal samples. During testing, the student and GLAE cannot accurately fit the local or global feature distributions of abnormal samples. By calculating the feature distribution distances among the teacher, student, and autoencoder, local and global anomalies in images are identified. Our method was evaluated on three industrial anomaly detection datasets. On the classic logical anomaly detection dataset MVTec LOCO, GLAE achieves an image-level AUC-ROC of 91.1%, demonstrating state-of-the-art performance. It also achieves excellent results on the MVTec AD and VisA datasets, with scores of 99.2% and 98.2%, respectively. Our method processes anomalies with the latency of less than 7&#xa0;ms on a single NVIDIA RTX 3090 GPU, making it an economically viable solution for industrial applications.</p>