New Spectral and Textural Feature Combinations for Corrosion Detection in Hyperspectral Images of Special Nuclear Materials Packages

Abstract

This article presents a novel approach to corrosion detection on special nuclear material (SNM) packages using hyperspectral imaging (HSI). Laboratory samples of carbon steel are exposed to chloride salt solutions (NaCl and KCl) in concentrations ranging from 0.001 to 1.0 M. Images of these samples are captured using a hyperspectral sensor in the visible-near-infrared range [400–1000 nm]. Spectral and spatial features, namely principal components, windowed gradients (WGs), and local binary patterns (LBPs) are extracted from the hyperspectral images. The HSI feature vectors are then used to train a support vector machine (SVM) to detect corrosion. Literature in HSI for corrosion detection emphasizes the spectral features while neglecting the important information that can be gleaned from the spatial domain, for example, textural features. This work demonstrates that the combination of spectral and textural information in corrosion detection can outperform spectral or spatial information alone. The SVM trained on the laboratory samples is then applied to hyperspectral images of an SNM package. Here, the results show a consistency of the joint spectral and textural feature vector giving an excellent indication of where corrosion products have formed. This work introduces a novel nondestructive (ND) and noncontact method for assessing corrosion products on steel surfaces, significantly reducing the visual ambiguity in corrosion detection. Our proposed dual-feature HSI approach marks a significant advancement in the field, providing a more accurate and comprehensive means of detecting corrosion products when compared to existing approaches that focus on spectral or spatial features in isolation.