Anotace:
Ensuring the structural integrity of cooling towers is paramount for safety and efficient operation. This paper presents a novel approach for detecting damage on cooling tower shells, utilising textures derived from laser scanning and close-range photogrammetry. The proposed method delves beyond the limitations of solely relying on colour information by harnessing the rich details embedded in various textures, including diffuse, normal, displacement, and occlusion. The study demonstrates the efficacy of this approach for identifying significant concrete damage. A Convolutional Neural Network (CNN) trained on diffuse textures successfully detects high damage instances with minimal misdetection. However, accurately pinpointing low damage, often manifesting as subtle cracks, and mimicking other patterns like air pores, ribbing, and colour variations, presents a formidable challenge. To tackle this challenge, the authors introduce a novel "composed raster layer" that merges information from multiple textures. This pre-processed layer amplifies the visual cues associated with low damage, facilitating its differentiation from similar patterns. While the current implementation employing multi-resolution segmentation and rule-based classification exhibits promising results, further optimization is acknowledged to refine the accuracy of low damage detection. The successful application of textures commonly used in rendering techniques underscores their remarkable potential for enhancing damage detection in civil engineering applications. While acknowledging limitations such as the analysis of a single cooling tower and the reliance on specific software for damage detection, the study proposes future research directions. This research holds significant implications for the field of civil engineering by offering a promising approach for automated and efficient damage detection on cooling tower shells.