Robust Weld Defect Inspection Using Normalizing Flows with Gaussian Priors

다국어 초록 (Multilingual Abstract)

Automated defect detection, particularly in radiographic imaging used for industrial weld inspections, remains challenging due to low signal-to-noise ratios and limited examples of defects. This research presents an innovative framework that leverages normalizing flows for automated weld defect detection. We employed various state-of-the-art Normalizing flow architectures with different feature extractors to detect defects in weld radiographic images. We comprehensively compared the results with radiographic images of welded steel pipes collected from industrial sites. The results show that the combination of CFlow-AD with a wide residual network-50-2 outperforms the other methods, indicating its effectiveness in defect detection and highlighting its suitability and potential for real-world industrial applications. Furthermore, a mixture of Gaussian models is used in the Normalizing flow model's latent space. In real- world scenarios, even a single object class may exhibit diverse normal patterns, or across different courses, Normalizing flow models may map all inputs to very similar latent variables and then assign high likelihoods to both normal and defective features, thus failing to detect anomalies. The mixture of the Gaussian modeling approach can provide a stronger representation in the latent space, enabling better differentiation between defective and normal samples.

목차 (Table of Contents)

• Abstract i
• List of Figures iv
• List of Tables v
• CHAPTER 01: Introduction 1
• 1.1. Motivation 1
• 1.2. Problem Definition 1
• 1.3. Purpose of Study 2
• 1.4. Outline 3
• CHAPTER 2: Related Works 5
• 2.1. Supervised Methods 5
• 2.1.1 Support Vector Machine (SVM) 5
• 2.1.2. Radial Basis Function Neural Network (RBF-NN) 6
• 2.2. Self-Supervised Method 6
• 2.2.1. Reconstruction-Based Method 6
• 2.2.2. Feature Embedding-Based Method 9
• CHAPTER 3: Method 12
• 3.1 Model Description 12
• 3.1.1. Normalizing Flows 12
• 3.1.2. Coupling Blocks 13
• 3.1.3. Feature Extraction 13
• 3.2. Normalizing Flow Models 14
• 3.2.1. CFlow-AD 14
• 3.2.2. FastFlow 15
• 3.2.3. CSFlow 15
• 3.3. Anomaly Score 15
• Chapter 4: Experiments 16
• 4.1. Data Description 16
• 4.2. Experimental Setting 18
• 4.3. Performance Measures 18
• 4.4. Quantitative Comparison 18
• 4.4.1. Baseline Reconstruction Models 18
• 4.4.2. NF Models 18
• 4.5. Qualitative Comparison 20
• CHAPTER 5: Mixture of Gaussian and NF-Based Approaches for Defect Detection 25
• 5.1. Mixture of Gaussian Models in Defect Detection 25
• 5.1.1. Definition 25
• 5.2. Baseline Model 25
• 5.3. Model Interpretation 26
• 5.3.1. Separation Loss 27
• 5.3.2. Consistency Loss 27
• 5.3.3. Model Hyperparameters 27
• 5.3.4. Feature Extractor Model 28
• 5.3.5. Anomaly Score 28
• 5.4. Experimental Results 28
• 5.4.1. Applied Dataset 28
• 5.4.2. Quantitative Performance 29
• 5.4.3. FlowGoM vs HGAD: Geometry in Linear and Non-Linear Embedding Spaces 31
• 5.4.4. Qualitative Performance 31
• CHAPTER 6 34
• 6.1. Discussion 34
• 6.2. Conclusion 35
• REFRENCES 36