Walid GHENAIET

Addressed the critical challenge of data scarcity in automated visual quality control for industrial manufacturing by designing generative pipelines to create high-fidelity synthetic defect data. These datasets enabled the training of supervised models that achieved competitive performance, often surpassing state-of-the-art unsupervised methods. Key Achievements: Built a few-shot anomaly generation pipeline using pre-trained diffusion models, adapted with parameter-efficient fine-tuning (LoRA, text embedding tuning), to generate diverse and realistic synthetic defects from very limited real data. Designed and trained supervised models for anomaly classification and pixel-level segmentation, reaching strong results on industrial benchmarks. Applied Grad-CAM to analyze and interpret model behavior, improving debugging and model explainability. Developed an advanced mask transformation pipeline for defect inpainting, including category-specific strategies (ORB keypoint matching for alignment, Otsu’s thresholding for object isolation), boosting the diversity and semantic realism of generated anomalies. Conducted rigorous benchmarking against leading unsupervised methods and simpler synthesis techniques, demonstrating the superior value of generative approaches for segmentation tasks. Authored a comprehensive thesis comparing generative strategies, identifying their limitations on complex logical anomalies, and proposing future directions for industrial deployment.