Defect Segmentation Strategy Comparison in Industrial Quality Control

This topic presents a project focused on developing and comparing deep learning strategies for industrial defect detection, emphasizing fine-grained segmentation of multiple defect classes in quality control processes.Industrial QC challenges:

• Industrial quality control requires precise and fast detection of subtle surface defects in processes like painting and welding, where deep learning surpasses traditional fixed-rule vision systems;

• Multiple defect classes: Handling multiple distinct and often imbalanced defect classes such as 'paint run' and 'scratch-deep' demands effective segmentation strategies and training approaches;

• Project objective: The project aims to design, implement, and validate an end-to-end AI pipeline to systematically compare monolithic versus modular deep learning segmentation strategies on a multi-class industrial defect dataset;

• Segmentation strategies: Two main approaches are compared: a monolithic model segmenting all classes simultaneously, which simplifies deployment but risks class interference, and a modular hierarchical approach using specialized models for subsets of classes to improve accuracy at the cost of complexity;

• Evaluation criteria: The comparison focuses on segmentation accuracy for fine-grained classes, computational efficiency, and maintainability in real industrial settings;

• Pipeline details: The project includes data acquisition with high-resolution imaging and pixel-level annotation of up to 10 defect classes, advanced data augmentation, loss functions addressing class imbalance, rigorous training with cross-validation, and real-time inference validation using industrial metrics;

• Learning and profile: Participants will gain expertise in end-to-end AI pipelines, advanced segmentation architectures, strategic model training, industrial data handling, and performance benchmarking, requiring strong deep learning and computer vision fundamentals.