Detection and Segmentation of Manufacturing Defects with Convolutional Neural Networks and Transfer Learning (1808.02518v2)

Abstract: Quality control is a fundamental component of many manufacturing processes, especially those involving casting or welding. However, manual quality control procedures are often time-consuming and error-prone. In order to meet the growing demand for high-quality products, the use of intelligent visual inspection systems is becoming essential in production lines. Recently, Convolutional Neural Networks (CNNs) have shown outstanding performance in both image classification and localization tasks. In this article, a system is proposed for the identification of casting defects in X-ray images, based on the Mask Region-based CNN architecture. The proposed defect detection system simultaneously performs defect detection and segmentation on input images, making it suitable for a range of defect detection tasks. It is shown that training the network to simultaneously perform defect detection and defect instance segmentation, results in a higher defect detection accuracy than training on defect detection alone. Transfer learning is leveraged to reduce the training data demands and increase the prediction accuracy of the trained model. More specifically, the model is first trained with two large openly-available image datasets before finetuning on a relatively small metal casting X-ray dataset. The accuracy of the trained model exceeds state-of-the art performance on the GRIMA database of X-ray images (GDXray) Castings dataset and is fast enough to be used in a production setting. The system also performs well on the GDXray Welds dataset. A number of in-depth studies are conducted to explore how transfer learning, multi-task learning, and multi-class learning influence the performance of the trained system.

• The paper introduces a combined defect detection and segmentation framework using Mask R-CNN, achieving a mAP of 0.957 on the GDXray dataset.
• The paper leverages extensive pre-training followed by fine-tuning on metal casting X-rays to overcome limited training data challenges.
• The paper demonstrates that simple augmentation like flipping boosts performance, while techniques like Gaussian blur and noise impair accuracy.

Detection and Segmentation of Manufacturing Defects with Convolutional Neural Networks and Transfer Learning

The paper addresses the automation of quality control processes in manufacturing, particularly targeting the detection and segmentation of defects in casting and welding components using convolutional neural networks (CNNs) and transfer learning techniques. The methodology leverages the Mask Region-based CNN (Mask R-CNN) architecture to identify casting defects within X-ray images, with the capability to execute both defect detection and precise defect instance segmentation. This combined approach has demonstrated a superior accuracy in contrast to employing defect detection techniques in isolation.

A key feature of the proposed approach is its efficient use of transfer learning, which mitigates the challenges associated with limited training data. Initially, the model undergoes pre-training on extensive, publicly available image datasets before fine-tuning on a more compact metal casting X-ray dataset. The paper reports that the accuracy of the trained model outpaces existing state-of-the-art benchmarks on the GRIMA database of X-ray images (GDXray) Castings dataset, suitable for real-time deployment within production environments. Additionally, the model performs competitively on the GDXray Welds dataset.

The implementation results highlight the system's efficacy, citing a mean average precision (mAP) of 0.957 for bounding box predictions on the GDXray dataset. This superior performance is attributed to the multi-task learning framework which capitalizes on the simultaneous prediction of bounding boxes and segmentation masks, enhancing the model's accuracy compared to traditional methods. Furthermore, by incorporating a ResNet-101 feature extractor pre-trained on ImageNet, the paper effectively utilizes transfer learning, significantly boosting the model's generalization capabilities.

Experimentation also explored the implications of various data augmentation strategies, revealing that simple flipping notably augmented performance, whereas techniques like Gaussian blur and noise were detrimental. Training set augmentation, thorough hyperparameter tuning, and an insightful error analysis provided a comprehensive understanding of the model's robustness and potential limitations. Notably, zero-shot learning experiments exhibited the system's adeptness in identifying defects in novel datasets, highlighting its promise for broader defect detection applications beyond metal castings.

In summation, the research contributes a proficient framework for defect detection in manufacturing processes, combining the strengths of CNN-based architectures and transfer learning. Its ability to generalize beyond trained instances presents intriguing avenues for future exploration, including applications across diverse manufacturing materials and methods, such as additive manufacturing. Future work may center on refining the training workflows to facilitate the system's broad deployment in industrial scenarios. This paper underscores the potential of modern deep learning approaches in revolutionizing quality control in manufacturing, thereby promising significant advancements in efficiency and reliability.