SEAD: Segment Element-Based Anomaly Detection

Anomaly detection in images is crucial for the quality control process in manufacturing. Existing anomaly detection methods have achieved high accuracy on many public datasets, which typically include structural anomalies that affect part of the image. However, in real-world scenarios, in addition to structural anomalies, logical anomalies that affect the entire image can occur, and these methods often struggle to detect them. Therefore, this paper proposes an anomaly detection method called "SEAD", based on the conventional "ComAD", which can detect structural and logical anomalies with higher accuracy while segmenting each element of the image more flexibly. Specifically, SEAD involves annotating each MVTec LOCO dataset category which contains structural and logical anomalies with five images per category. SEAD then employs the few-shot segmentation model "SegGPT" to segment each image into multiple elements. Following this, SEAD constructs the memory bank that stores the color and pixel count of each element in normal images and detects anomalies by measuring the Euclidean distance from the test images. Additionally, SEAD normalises the anomaly scores using the evaluation dataset to align the anomaly score scales of the conventional anomaly detection model with those of the proposed model. Experiments validate the proposed method by comparing it to previous anomaly detection methods on the MVTec LOCO dataset. The experimental results show that the proposed method achieves an AUROC of 91.2% (an improvement of 2.2%), demonstrating its superiority over existing methods.