Enhancing Metal Surface Defect Recognition Through Image Patching and Synthetic Defect Generation

• Published in: IEEE access 2023, Vol.11, p.113339-113359
• Main Authors: Mustafaev, Bekhzod, Kim, Sungwon, Kim, Eungsoo
• Format: Article
• Language: English
• Subjects: Accuracy; Classification; Complexity; convolutional neural network; Convolutional neural networks; Data models; Datasets; Deep learning; Defective products; Effectiveness; Feature extraction; Generative adversarial networks; Housings; Human error; Image enhancement; image patching; Image quality; Image recognition; Industrial applications; Inspection; Material properties; Metal surfaces; Patching; Quality control; Servomotors; Surface contamination; surface defect recognition; Surface defects; Surface treatment; synthetic defect generation; Training
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3322734

Preventing surface defects of metal products during the production process is challenging due to manufacturing complexity, material properties and environmental factors. Relying on human inspectors for quality control can introduce human error, which increases the risk of delivering defective products to customers. To address these challenges, we propose an Inception-CNN model specifically designed for surface defect recognition in servo motor housings (SMHs). The model incorporates an inception module between convolutional layers to effectively capture multi-scale local information and extract complex and abstract features. Additionally, we introduce an image patching technique that enhances defect recognition for small defects by reducing image complexity while maintaining defect visibility. Moreover, we propose a surface defect generation GAN (SDG-GAN) method, a novel approach that addresses the data imbalance problem and improves the accuracy and robustness of the classification model through generating diverse and high-quality synthetic defect images. The original data was collected using a line scan camera installed in the SMHSI system. We ensured model generalization through 10-fold cross-validation using the SMHSD-P-GAN dataset. Evaluation results indicate that our classification model outperformed other CNN models and achieved strong generalization, with 99.40% accuracy in cross-validation and 99.23% on the original test data. This represents a substantial 32.31% improvement over a baseline CNN model trained on the SMHSD-O-TA dataset, underscoring the effectiveness of our proposed approaches in enhancing classification performance Our method efficiently processes 12 images per second, making it ideal for real-time defect inspection in SMHs. Its successful integration into the SMHSI system confirms its practicality and effectiveness in real-world industrial applications.