Article Preview
TopIntroduction
Since global warming and extreme weather events have become increasingly intense, flood control has become one of the priorities among disaster protection systems. Dikes play a crucial role in flood control efforts, serving as critical elements for flood prevention and mitigation. However, as a result of prolonged operation and inadequate maintenance, dikes are vulnerable to seepage penetration and water scouring, which leads to varying degrees of danger and poses a serious threat to property and human safety in affected areas. Unfortunately, current dike collapse risk monitoring methods mostly rely on regular manual inspections, which not only consume significant human, financial, and material resources, but also are inefficient. Additionally, dikes in remote areas may not be readily accessible for manual inspections, leading to inspection loopholes.
In recent years, with the aid of neural networks and deep learning technology, visual defect detection, based on deep learning, has gained significant attention in the field of defect detection (Luo et al., 2021; Luo et al., 2022; Ofir et al., 2023; Tapamo et al., 2023). The deep learning-based computer vision defect detection is primarily trained by feeding a large number of defect samples into a deep network, which in turn learns the characteristics of the identified targets. In comparison with traditional feature extraction methods such as the Scale Invariant Feature Transform (SIFT) (Gao et al., 2015) and the Histogram of Oriented Gradients (HOG) (Chacon-Murguia et al., 2021), this kind of target detection method not only lowers false detection rate and leakage rate, but also has better feature expression capability and detection accuracy. There are two main categories of deep learning-based target detection algorithms. One category is two-stage target detection, which divides the candidate bounding box and classification tasks. The first stage extracts the potential location of the target through a specific neural network to form a target candidate region, and the second stage detects the target of candidate regions, which represent R-CNN (Agrawal et al., 2014) and Fast R-CNN (Girshick, 2015), in particular. The other category is one-stage target detection, which combines candidate region proposal and target detection into a single step. The YOLO (Bochkovskiy et al., 2020; Redmon et al., 2016; Redmon & Farhadi, 2017, 2018) series is a representative example of the other category.
However, traditional target detection algorithms such as YOLO are often utilized for defect detection by firstly identifying a large number of defective samples, labeling them, and then feeding them into the YOLO model for training. Although this approach is effective, it may be challenging to obtain a significant number of defective samples. In the context of dike collapse detection, the occurrence rate of collapsed and defective sections within an individual embankment is consistently observed to be minimal. Consequently, this scarcity poses a significant challenge in acquiring an ample defect sample data. Furthermore, obtaining certain specified data sets may not be feasible owing to the limited availability of openness. Accordingly, the detection of defects with zero defect samples remains a significant challenge that both industry and academia strive to overcome. In this paper, an UAV platform based zero defect sample defect detection method for dike collapse defect detection, which is also capable of providing heuristic method in further scenarios, is proposed.