2D Image-Based Burr Classification for Abrasive Tool Selection in Sheet Metal Post-Processing

Accurate characterization of burrs in laser- and plasma-cut sheet metal is essential for selecting suitable abrasive tools during post-processing and finishing. This research presents an automated computer vision pipeline embedded in a salesbot system, designed to classify different types of laser-induced burrs and slag from end-user images captured in real world environments. The system facilitates rapid, data-driven tool recommendations by overcoming key challenges such as variable lighting conditions, reflective metal surfaces, and disparities in smartphone camera hardware. The proposed solution enhances technical sales workflows by bridging the gap between expert knowledge and end-user input. In the first processing step, the customer image is segmented to isolate edge snippets of 224x224 pixel in which laser burr or slag can occur. A prompt-driven FastSAM/MobileSAM ensemble with center bias ensures fast and robust segmentation. To classify burrs in each snippet, a ResNet-50 feature extractor is trained on an image set acquired in controlled lab conditions selected for feature prominence. The validation with real-world images from various types of smartphones achieves an F1-score of 0.86. For increased robustness, the extracted feature vectors of each snippet are aggregated in an attention network which in an adverse condition test provides 100% accuracy for image-level laser burr vs. slag decisions. Moreover, subcategories for laser burrs can be classified by a majority vote on MLP outputs achieving at least 94% accuracy. The proposed architecture thus provides a production-ready solution for complex visual assessment of smartphone images to obtain tooling decisions for varied cutting processes.