Scaling anomaly detection with segmentation models

Detecting anomalies in product images is a critical task in industrial quality control, where even subtle defects can have operational and financial impact. However, deploying anomaly detection algorithms in real-world industrial scenarios remains challenging, particularly when products are large, complex, or captured at high resolution. Many existing methods struggle to scale effectively while maintaining precision. This work aims to develop an algorithm that can effectively scale to larger, more complex objects. The method, SADSeM (Scaling Anomaly Detection with Segmentation Models), is based on classic convolutional neural networks for segmentation, such as Mask-RCNN. Thanks to these models’ ability to learn and encode an object's structure, we can design a pipeline that uses both their segmentation maps and feature embeddings to carry out unsupervised anomaly detection. As the segmentation task is effectively solved by these models independently of image size, we scale to higher-resolution images with more effectiveness than competitors, while maintaining competitive results in simpler scenarios.