A two-stage deep learning method for foreign object detection and localization

Foreign object detection and localization is of critical importance in various real-world scenarios. For example, a foreign object in an automatic assembly line could result in severe dangers. In this paper, we propose a two-stage deep learning method to identify and localize foreign objects in a working environment. It includes two major stages, i.e., detection and localization. In the detection stage, an advanced anomaly detection model is first used to identify unknown object candidates. However, the unknown objects detected in this stage might include potential normal classes. Subsequently, we use the working environment data to train a YOLO model to filter out the false positives in the potential unknown objects. In the localization stage, we use K-Means++ to cluster a heatmap generated in the first stage and extract activation points with highest activation scores which are fed into an advanced segmentation model for accurate segmentation and localization of foreign objects. We have conducted experiments to validate the performance of the method in an experimental setup. The developed mode can well adapt to various scenarios in manufacturing automation.