Hyperspectral Imaging Applied for Pixel-Level Crack Detection with Background Interferences

Cracks in civil infrastructures are an important sign of structural degradation and may indicate the inception of catastrophic failure. Existing image-based crack detection techniques face challenges when it comes to the complex background scenes. These irrelevant background interferences are common in practice and may trigger false alarms in crack detection. To eliminate their influence, hyperspectral imaging is employed in this study, which captures hundreds of spectral reflectance values in a pixel in the visible and near-infrared region. Compared with the conventional greyscale/RGB images which are limited to one/three wide spectral bands (red, green, blue), hyperspectral imaging can therefore provide more rich spectral information for crack detection/distinguish cracks from other background interferences. Due to the high correlations in hyperspectral image data, this study proposed a hyperspectral crack detection method using the low rank representation-based algorithm. Moreover, a locality constraint together with the dictionary learning process is incorporated into the proposed method to train a multi-class classifier. The built classification model is tested based on a real-world hyperspectral imaging dataset, which contains eight different surface objects in total. The trained classifier achieves an overall accuracy of 92.1%. The results show that the proposed method can predict cracks and other materials under complex scenes.