Research Article
A Study on Specular Highlight Removal for High-Reflective Surfaces Based on Two-Stage Network and Feature Alignment
@INPROCEEDINGS{10.1007/978-3-032-43712-9_10, author={Tsung-Hua Tsai and Ci-Yi Lai and Cheng-Hui Chen}, title={A Study on Specular Highlight Removal for High-Reflective Surfaces Based on Two-Stage Network and Feature Alignment}, proceedings={Smart Grid and Internet of Things. 10th EAI International Conference, SGIoT 2026, Taichung, Taiwan, November 21--22, 2026, Proceedings}, proceedings_a={SGIOT}, year={2026}, month={11}, keywords={Automated optical inspection specular highlight suppression highly reflective surface scratch inspection two-stage architecture spatial feature alignment}, doi={10.1007/978-3-032-43712-9_10} }- Tsung-Hua Tsai
Ci-Yi Lai
Cheng-Hui Chen
Year: 2026
A Study on Specular Highlight Removal for High-Reflective Surfaces Based on Two-Stage Network and Feature Alignment
SGIOT
Springer
DOI: 10.1007/978-3-032-43712-9_10
Abstract
Automated optical inspection (AOI) of highly reflective aluminum surfaces is often affected by specular highlights, which can obscure fine scratch features and introduce reflection-induced responses during subsequent image processing. This study investigates the suppression of specular-highlight interference for scratch inspection through controlled image acquisition and a two-stage processing architecture. A total of 48 grayscale images were acquired from the same scratched aluminum-surface region using six aperture settings and exposure times ranging from 5,000 μs to 40,000 μs. In the first stage, Sobel X edge extraction, image binarization, and connected component analysis are employed to enhance elongated scratch structures and suppress small or irregular reflection-induced responses. The spatial position and extent of the retained scratch candidates are then transferred through a candidate-guided spatial feature alignment strategy to define the search regions for the second stage, in which a sliding-window procedure analyzes continuous scratch responses and determines the final localization. Experimental results show that imaging conditions strongly influence reflection-induced foreground noise and subsequent candidate extraction. Among the tested combinations, f/16 with an exposure time of 35,000 μs produced the cleanest observed binary representation while preserving the target scratch response, resulting in fewer irrelevant candidates and more stable localization. These results demonstrate that controlling specular-highlight interference together with two-stage candidate extraction and localization provides an effective and interpretable approach for scratch inspection on highly reflective aluminum surfaces.