Evaluatie van de productconditie met behulp van deep learning voor een circulaire economie

Waste Electrical and Electronic Equipment (WEEE) is among the fastest-growing waste streams globally. Improper disposal not only endangers the environment and human health but also wastes valuable resources. This issue fundamentally hinders the closing of material cycles and the extension of product lifetimes, as targeted in a circular economy strategy. Surface condition evaluation is a critical step. The type and location of defects determine whether the device should be reused, remanufactured, or recycled. However, current condition evaluation relies mainly on manual operations. This method suffers from high subjectivity and a lack of transparency in results. Although deep learning has advanced automation in condition evaluation, existing technologies still face challenges due to the diversity of product models and significant differences in defect size and shape. To address these challenges, this PhD thesis proposes deep learning condition evaluation solutions that integrate computer vision algorithms with image acquisition systems. First, this thesis surveys standards from mainstream reuse and recycling enterprises. On this basis, it refines grading guides for evaluating the condition of laptop surfaces. Combined with deep learning, this thesis explores feasible approaches to achieve transparent, quantitative evaluation results. Next, to address the diversity of defect sizes and shapes, this thesis investigates multimodal color (RGB) and Depth (D) fusion techniques to improve defect detection accuracy. Subsequently, this thesis leverages Multimodal Large Language Models (MLLMs) to evaluate product condition through a semi-automatic, active learning Visual Question Answering (VQA) pipeline. Several combinations of creating prompts with single top-view or multi-angle-view images are explored, as well as the use of Retrieval-Augmented Generation (RAG) with the inclusion of corresponding good-condition images from the Digital Product Passport (DPP), performing Iterative Closest Point-registered point cloud comparison for Anomaly Highlighting (ICPAH) or YOLO11 defect detection to highlight areas in the image included in the prompt. This VQA approach enhances the system's generalization ability across different products. The effectiveness of the proposed methods in this thesis has been validated across different products and datasets. Experimental results demonstrate that these methods significantly improve defect inspection accuracy. Finally, this thesis discusses different solutions for various application scenarios.