Laboratory images acquired using the Marposs system (UV-lighting, 5 MP camera) and robotic pipeline images from Safe Metal provide primary high-quality data for automated MPI evaluation. For the current analysis, the dataset includes 30 systematically acquired laboratory images and 20 robotic pipeline images, with plans to expand the robotic dataset to 50-250 images. Secondary data, consisting of low- to mid-quality images from handheld cameras and commissioning phases, is also utilized to test the robustness of the models. Annotation plays a critical role in preparing the data for AI analysis. Using CVAT, cracks are labeled at the pixel level, enabling structured annotation for training, validation, and testing. This meticulous approach ensures the accuracy and reliability of the AI models in the evaluation phase. The deep learning framework employed for defect detection is based on a modified U-Net architecture with a ResNet encoder. By utilizing pre-trained models and transfer learning, the system achieves enhanced performance in semantic image segmentation. Training is conducted on laboratory images, with approximately 80% of the data allocated for training and 20% for validation. Robotic pipeline images serve as fresh test data to evaluate the model’s effectiveness. Initial results have demonstrated successful defect detection, with model predictions and probability maps showcasing the system’s capabilities. These advancements underscore the SeConRob project’s commitment to integrating advanced AI techniques into industrial applications, paving the way for more efficient and accurate automated defect detection and analysis.
