Relevant channel selection in hyperspectral imaging to enhance crack segmentation in historic concrete buildings

Abstract

Recent years have been fruitful in the development of computer vision methods for a wide variety of applications. Despite the successful results achieved in the segmentation of cracks on concrete surfaces, poor results are still persisting during onsite application, mainly due to noise caused by biological colonization, which is present in most of historical heritage buildings. The authors have been working on this problematic and developed the SC-Crack method previously, however it still relies on the cumbersome task of acquiring sets of 17 channels images to compose an hyperspectral cube and still requires case-wise hyperparameter optimization. Consequently, it is important to define which spectral information mostly defines the success of the method, enabling to optimize both, the acquisition procedure and model processing. Following, a study aiming at the selection of the more informative channels was carried and the hyperparameter-free model is evaluated.

In this scope, images of concrete specimens were acquired sequentially to compose a 17 channel hyperspectral image cube. These were sere compute allowing to define the most informative channels sets that are processed using the SC-Crack+ method, presented in this work. The reduced image cubes of cracking on clean concrete surfaces and on surfaces with biological colonization were processed and analyzed. Relevant and improved results were achieved for crack segmentation, following this SC-Crack+ model. This enables the possibility of mounting cameras with sensors and lenses particularly adapted for prone acquisition targeting only the most relevant hyperspectral information for crack segmentation and still using traditional feature engineering image processing methods.