3 results
Now showing1 - 3 of 3
- Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Tuning domain randomization for 3D-rendered synthetic training data for defect detection in aircraft engine borescope inspectionVisual surface defect detection in complex industrial inspection scenarios remains predominantly manual. Deep learning–based vision systems promise to automate this task but require large quantities of representative, annotated defect images, which are often unavailable due to the rarity and diversity of relevant defects. We address this challenge with a modular synthetic data generation pipeline leveraging physically based 3D rendering, which enables systematic domain randomization throughout the entire rendering chain. The pipeline comprises dedicated modules for scene composition, geometry-based defect generation, procedural material modeling, rendering, and automatic pixel-wise annotation, allowing users to expose and tune key parameters such as defect size, position, and penetration depth, object materials, camera poses, and lighting. We first provide a structured analysis of all process steps that influence image appearance in defect synthesis and then instantiate the pipeline for a case study on aircraft engine borescope inspection, targeting missing-material defects on high-pressure compressor blades. Using a two-stage experimental design, we employ a Plackett–Burman screening over 22 candidate factors followed by a one-factor-at-a-time optimization of the most influential parameters. Our results show that the material model, defect positioning, defect size, and penetration depth randomization dominate detection performance. Training an object detector solely on synthetic images generated with the optimized configuration yields a maximum of 0.89 mAP@0.5IoU on a real-world borescope validation set, with performance saturating at ∼500 rendered training images. These findings demonstrate that carefully structured domain randomization in a physically based rendering pipeline can provide practically useful synthetic training data for industrial defect detection.Publicationtype: Journal Article9 13 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Introducing a tool for synthetic defect image data generation: enhancing industrial surface inspectionVisual surface defect detection in complex scenarios remains a manual endeavor in industrial inspection. Stateof- the-art machine learning approaches promise to automate this task through machine vision; however, they require substantial, suitable training data to be effectively trained. For industrial products, such data are often unavailable for several reasons, including the rarity of certain defects or their manifestations. This paper introduces a tool to generate synthetic data for training machine vision systems using rendering techniques. Our tool allows for targeted manipulation of process parameters throughout the rendering chain, including scene composition involving object, camera, and lighting, defect generation, placement, and material generation. This enables an exploration of the impacts along the rendering pipeline and facilitates selecting an appropriate data set for specific inspection tasks. We demonstrate our tool’s effectiveness by applying models trained on our synthetic data to a real-world inspection task.Publicationtype: Conference Paper77 75 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Adapting synthetic training data in deep learning-based visual surface inspection to improve transferability of simulations to real-world environments(The International Society for Optical Engineering, 2023); ; ; Learning models from synthetic image data rendered from 3D models and applying them to real-world applications can reduce costs and improve performance when using deep learning for image processing in automated visual inspection tasks. However, sufficient generalisation from synthetic to real-world data is challenging, because synthetic samples only approximate the inherent structure of real-world images and lack image properties present in real-world data, a phenomenon called domain gap. In this work, we propose to combine synthetic generation approaches with CycleGAN, a style transfer method based on Generative Adversarial Networks (GANs). CycleGAN learns the inherent structure from real-world samples and adapts the synthetic data accordingly. We investigate how synthetic data can be adapted for a use case of visual inspection of automotive cast iron parts, and show that supervised deep object detectors trained on the adapted data can successfully generalise to real-world data and outperform object detectors trained on synthetic data alone. This demonstrates that generative domain adaptation helps to leverage synthetic data in deep learning-assisted inspection systems for automated visual inspection tasks.Publicationtype: Conference Paper31 28