The PFILSTM model: a crack recognition method based on pyramid features and memory mechanisms

Abstract

Crack detection is a crucial task for the structural health diagnosis of buildings. The current widely used manual inspection methods have inherent limitations and safety hazards, while traditional digital image processing methods require manual feature extraction and also have substantial limitations. In this paper, we propose a crack recognition method based on pyramid features and memory mechanisms that leverages a U-shaped network, long short-term memory mechanisms, and a pyramid feature design to address the recognition accuracy, robustness, and universality issues with deep learning-based crack detection methods in recent years. Experiments were conducted on four publicly available datasets and one private dataset. Compared with the commonly used FCN8s, SegNet, UNet, and DeepLabv3+ models and other related studies using the same evaluation criteria and datasets, our proposed model shows better overall performance in terms of all metrics evaluated.