LIGHTWEIGHT MACHINE LEARNING FOR RAPID AND ACCURATE CORROSION DETECTION

Abstract

The deterioration of metallic structures due to corrosion presents substantial economic and safety concerns, contributing to nearly 4% of the global GDP. Despite the promise of machine learning in corrosion detection, progress is hindered by the scarcity of standardized datasets. In this research, the ARL-WPI corrosion dataset has been utilized to refine multi-label classification models through computationally efficient feature extraction and optimized machine learning techniques. However, Colour descriptors such as mean and standard deviation of Red, Green, and Blue (RGB) and Hue, Saturation, and Value (HSV) channels have been extracted for enhanced illumination invariance, along with texture-based attributes Histogram of Oriented Gradients (HOG), Local Binary Pattern (LBP), and Gray Level Co-occurrence Matrix (GLCM) to encapsulate structural patterns and wavelet-derived features (DWT) for multiresolution frequency analysis. Additionally, Principal Component Analysis (PCA) has been applied to preserve 99% of the variance while significantly reducing computational overhead. The optimization of machine learning models, particularly K-Nearest Neighbour(KNN) with 98.6% accuracy and 28.27s execution time and ensemble methods with 97.2% accuracy and 67.24s execution time has demonstrated superior performance over deep learning frameworks, which exhibit higher computational demands. The findings underscore the efficacy of lightweight classifiers with domain-specific feature representations in expediting corrosion assessment. This study fosters a robust synergy between AI-driven material diagnostics and industrial applications, offering a scalable, high-precision, and computationally viable approach for real-world corrosion monitoring.