Analysis of the Efficiency of Quality Control Algorithms for Modified Surfaces for High-Error Processes Based on 2D Miniatures and Non-visual Information

This study investigates the efficiency of 2D thumbnail-based algorithms, complemented by non-visual information, in enhancing quality control for manufacturing processes characterized by high error rates like rimming. Amidst the challenges faced by traditional 3D technology-based quality control methods, such as extended computing times and the necessity for advanced hardware, this research proposes a different approach leveraging high-resolution 2D images and part metadata to identify defects. By employing Machine Learning techniques, specifically Mask R-CNN models enhanced with Squeeze-and-Excitation Blocks (SEBlocks) for detailed feature detection, and Random Forest classifiers for scale classification and tolerance validation, the study presents a cost-effective yet efficient alternative. The methodology encompasses four primary steps: image acquisition, enhanced hole detection, scale classification, and prediction optimization, concluding with diameter tolerance classification and validation. Employing a dataset of synthetic images annotated for hole locations, the system's performance is evaluated using metrics such as Intersection over Union (IoU), Dice score, and Mean Squared Error (MSE), demonstrating its efficacy in near-real-time defect detection. This research contributes to the field by offering an adaptable, scalable solution for quality control in manufacturing, significantly reducing dependency on costly hardware and paving the way for future innovations in process monitoring and defect detection after laser and machining processing.