This thesis addresses the problem of automated defect inspection in glass manufacturing, a challenging industrial context characterized by high production rates, complex material properties, and strong variability in defect appearance. Due to the transparency, reflectivity, and refractive behavior of glass, defects are often only visible under specific illumination conditions or viewing angles. As a result, modern inspection systems rely on multi-view and multi-modal imaging configurations, generating heterogeneous data that remain difficult to exploit effectively with conventional methods. \\
The main objective of this work is to investigate how deep-learning–based multi-modal and multi-view image fusion can improve the reliability, robustness, and interpretability of automated inspection systems, while satisfying industrial constraints such as real-time processing and incomplete data acquisition. \\
First, this thesis provides a comprehensive state of the art on image fusion, introducing a novel taxonomy that unifies supervised, unsupervised, and task-driven learning paradigms, and distinguishes between mono-category and multi-category fusion methods. This framework highlights the limitations of existing approaches and motivates the need for generalist fusion architectures capable of jointly handling multiple modalities and viewpoints.
The core contribution of this work is the introduction of EDIF (End-to-End Detection-Driven Image Fusion), a task-driven framework designed for multi-view and multi-modal defect detection. EDIF relies on a graph attention neural network to model relationships between keypoints extracted from multiple images, enabling explicit reasoning across views and modalities. The method integrates keypoint detection, feature matching, and object detection into a unified pipeline, allowing consistent grouping and classification of defects across all available observations. A dedicated multi-stage training strategy is proposed to jointly learn correspondence, localization, and classification tasks. To support this research, a novel industrial dataset, MMDOD, is constructed, comprising thousands of annotated images acquired from multiple viewpoints and modalities, providing a realistic benchmark for multi-view multi-modal detection under real production conditions. Building upon EDIF, a second contribution introduces a graph-based multi-view keypoint completion framework, designed to address missing data scenarios frequently encountered in industrial acquisition. By leveraging graph message passing and multi-view consistency constraints, the proposed approach enables robust inference even when some views or modalities are absent.
Finally, this thesis proposes a compact multi-view detection architecture based on Mamba state-space models, specifically designed for the detection of surface wearing marks on returnable glass containers. This type of defect, characterized by progressive abrasion patterns induced by repeated industrial usage cycles, is inherently viewpoint-dependent and particularly challenging to model. The proposed approach combines computational efficiency with strong performance, making it suitable for deployment in high-throughput industrial environments.
This work demonstrates that task-driven, graph-based fusion strategies significantly improve defect detection performance in complex industrial settings. It provides new methodological insights and practical solutions for bridging the gap between advanced image fusion research and real-world deployment in glass inspection systems.