Reflection Detection in Inspection Images for Reactive Planning of Autonomous Inspections

Abstract

The adoption of automatic inspection systems is growing across various industries, such as manufacturing and energy, and is expected to expand significantly into other sectors, including aerospace. The potential for automation in this field is substantial, promising advancements in visual inspection to improve decision-making and performance. These systems often encounter challenges when inspecting highly reflective metallic surfaces, where varying light conditions can obscure critical surface details. Such limitations not only compromise inspection accuracy but also pose potential risks to safety. To address these issues, this thesis explores the implementation of various U-Net-based architectures for detecting specular light reflections in inspection images, facilitating reactive planning during autonomous inspections. A novel dataset comprising inspection images and corresponding masks of light reflections is introduced, serving as a foundation for training the U-Net models. Key findings reveal that CNN-based U-Nets significantly outperform their Transformerbased counterparts, with U-Net++ featuring a ResNet-50 encoder yielding the highest Intersection over Union (IoU) and Dice Similarity Coefficient (DSC) scores. In contrast, the proposed UNETR-Attention Fusion (UNETR-AF) struggles to detect larger reflections but performs comparably for medium and smaller reflections. This research contributes valuable insights into industrial inspection applications focused on reflective surfaces. While the findings predominantly pertain to 2D RGB images, future work may explore the adaptation of these techniques to RGB-D images to capture additional depth information, potentially improving the efficacy of reactive planning in autonomous inspections. Furthermore, the application of generative AI could facilitate the creation of expansive datasets, while few-shot learning methods may be employed to mitigate data scarcity challenges.