Reflection Detection in Inspection Images for Reactive Planning of Autonomous Inspections

Abstract

The adoption of automated visual inspection systems is growing across various industries, such as manufacturing and energy, and is expected to expand significantly into other sectors, including aerospace. However, these systems often encounter challenges when visually inspecting highly reflective metallic surfaces, as varying light conditions can obscure critical surface details, thus risking errors in defect detection. This paper addresses these challenges by developing methods to detect specular light reflections in order to automatically assess the quality of inspection images. This enables automated systems to take inspection images from different angles, to avoid undesired reflections. We show that U-Net based architectures trained on a novel dataset of inspection images and reflection masks lead to good detection under challenging conditions. The results demonstrate that Convolutional Neural Network (CNN)-based models, particularly U-Net++ with a ResNet-50 encoder, outperform Transformer-based approaches, achieving the highest accuracy in identifying reflective areas. While the proposed UNETR-Attention Fusion (UNETR-AF) model shows promise for smaller reflections, it struggles with larger ones. This research offers a practical solution for industries aiming to improve visual inspection reliability, particularly for safety-critical applications. By enabling automated systems to handle reflective surfaces effectively, it addresses a significant gap in current inspection technologies.