Abstract
In the era of rapid technological advancements, computer vision is a cornerstone in the evolution of intelligent industrial production. The transformative potential of Convolutional Neural Networks (CNNs) in various aspects of product assessment, including classification, object detection, and anomaly detection, has opened new frontiers in quality inspection. However, amidst this progress, computer vision applications for irregular texture products still need to be explored. This thesis embarks on a comprehensive exploration of irregular texture products, delving into their distinctive characteristics and addressing the research gaps in anomaly detection and localization. Irregular texture products lack consistent shapes and patterns, causing unpredictable variations and anomalies and posing significant challenges for existing computer vision methods for anomaly detection and localization tasks.
Our first research question addresses the discrimination challenges inherent in irregular texture products. This thesis proposes a novel Content-Based Image Retrieval (CBIR) method, namely Class-Specific Variational Auto-Encoder (CS-VAE), that is designed to enhance the distinction between normal and abnormal irregular texture products. This contribution is supported by a real-world case study on a dataset of fibrous products, showcasing promising results in simulating authentic production scenarios.
Our second research question focuses on the complexities of precise defect localization in irregular texture products. A critical analysis of existing quantitative metrics reveals a mismatch with qualitative results, prompting the proposal of alternative metrics (Area Under the Precision-Recall Curve (AUPR) and F1-score) for more effective evaluation. Furthermore, a superpixel-based method called Superpixel-based Coupled-hypersphere-based Feature Adaptation (Sp-CFA) is introduced to enhance anomaly localization by addressing the challenges posed by irregular textures.
Lastly, this thesis introduces the Regularized Discriminative Coupled-hypersphere-based Feature Adaptation (RD-CFA) method, focusing on multi-class anomaly detection. This innovative model discriminates among various irregular texture products, simultaneously addressing both research questions with promising results compared to the state-of-the-art techniques.
In essence, this thesis not only explores the intricacies of irregular texture products but also contributes novel methodologies, addressing critical challenges and paving the way for the future application of computer vision in anomaly detection and localization of irregular texture products.
Our first research question addresses the discrimination challenges inherent in irregular texture products. This thesis proposes a novel Content-Based Image Retrieval (CBIR) method, namely Class-Specific Variational Auto-Encoder (CS-VAE), that is designed to enhance the distinction between normal and abnormal irregular texture products. This contribution is supported by a real-world case study on a dataset of fibrous products, showcasing promising results in simulating authentic production scenarios.
Our second research question focuses on the complexities of precise defect localization in irregular texture products. A critical analysis of existing quantitative metrics reveals a mismatch with qualitative results, prompting the proposal of alternative metrics (Area Under the Precision-Recall Curve (AUPR) and F1-score) for more effective evaluation. Furthermore, a superpixel-based method called Superpixel-based Coupled-hypersphere-based Feature Adaptation (Sp-CFA) is introduced to enhance anomaly localization by addressing the challenges posed by irregular textures.
Lastly, this thesis introduces the Regularized Discriminative Coupled-hypersphere-based Feature Adaptation (RD-CFA) method, focusing on multi-class anomaly detection. This innovative model discriminates among various irregular texture products, simultaneously addressing both research questions with promising results compared to the state-of-the-art techniques.
In essence, this thesis not only explores the intricacies of irregular texture products but also contributes novel methodologies, addressing critical challenges and paving the way for the future application of computer vision in anomaly detection and localization of irregular texture products.
| Original language | English |
|---|---|
| Publisher | |
| Publication status | Published - Mar 2024 |