An automated crack detection method for underwater structures based on multilevel DWT and LPQ feature generation

• Published in: Multimedia tools and applications 2023-11, Vol.82 (27), p.42331-42352
• Main Authors: Yaman, Orhan, Tuncer, Turker, Ertam, Fatih
• Format: Article
• Language: English
• Subjects: Computer Communication Networks; Computer Science; Computer vision; Cracks; Data Structures and Information Theory; Datasets; Deep water; Discrete Wavelet Transform; Feature extraction; Flaw detection; Image processing; Image quality; Iterative methods; K-nearest neighbors algorithm; Multilevel; Multimedia Information Systems; Performance measurement; Pixels; Special Purpose and Application-Based Systems; Underwater construction; Underwater robots; Underwater structures; Wavelet transforms
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-15229-9

Underwater image processing is a very important research area to detect cracks in underwater constructions. In this study, a new light computer vision method is proposed to detect cracks in underwater structures. The proposed model uses Multilevel DWT (Discrete Wavelet Transform) and LPQ (Local Phase Quantization) for underwater crack classification. The walls of a pool with a size of 10 × 12 × 1.5 m have been monitored to create experiments. The entire surface of the pool has been scanned using the unmanned underwater robot. 3840 × 2160 × 3 pixels 472 cracked images, 229 robust images were obtained and 701 underwater image datasets were created in total. Collected images were obtained in a clean water environment. Image quality changes in turbid and deep waters. For this reason, three different scenarios were obtained by adding turbid water and deep water template on the collected images. Thus, a dataset of 701 clean water environments, 701 turbid water environments, and 701 deepwater environments was created. These images have been converted to 512 × 512 pixels using preprocessing. For the later feature extraction, a feature of 256 × 5 size has been obtained using Multilevel DWT and LPQ. By combining features, a feature of 1 × 1280 size has been created. After feature extraction, 1 × 368 features have been selected for each image using the ReliefF Iterative Neighborhood Component Analysis (RFINCA) feature selection algorithm. Selected features are classified using K Nearest Neighbor (KNN) algorithm. In the proposed method, 99.85%, 99.42%, and 99.14% accuracy have been obtained using clean water environment, turbid water environment, and deepwater environment images, respectively. According to the comparisons and the calculated performance metrics, our proposal is successful for crack detection of underwater structures.