AugMoCrack: Augmented morphological attention network for weakly supervised crack detection

Abstract

Most crack-detection methods adopt a pixel-level segmentation-based approach, which requires considerable time and complexity to detect the pixel area of the crack. Unlike a pixel-level segmentation-based approach, in this paper, the authors proposed an AugMoCrack network, a bounding box-level crack detection approach for weakly supervised crack detection achieved by augmenting the training data with Poisson blending, as well as high-frequency discrete cosine transform-based features. The proposed AugMoCrack detects the box position of the crack object from a morphological perspective, such as neighbour connectivity within the crack pixels and crack-area fitting. Based on this perspective, the authors also proposed new morphological attention loss functions for considering the neighbour connectivity and the box area border. Specifically, the authors trained the AugMoCrack network using two datasets and verified its performance based on 591 validation images in the concrete crack dataset and 672 validation images in the Crack500 dataset. Compared with the baseline architecture, with the authors’ proposed network, the authors achieved increases of approximately 4.5% points and 2.5% points in the mean average precision (mAP) in the concrete crack and Crack500 datasets, respectively. The proposed network also outperforms previous state-of-the-art crack-detection methods in a weakly supervised learning environment where the training data are insufficient.