Crack Detection via Hierarchical Multiscale Feature Learning and Densely Connected Conditional Random Field

Abstract

Crack analysis based on computer vision has become a common approach for crack detection and localization in civil infrastructure. In practice, many cracks show poor continuity, uneven gray levels, low contrast, complex topology, and background noise. These characteristics present significant difficulties for image-based crack detection. In this paper, we propose a novel framework that includes a deep fully convolutional network and densely connected conditional random field (dense CRF) to realize pixel-level crack detection in an end-to-end manner. The network learns and aggregates multilevel features at hierarchical convolutional stages. Specifically, the backbone of our network is novel self-attention modules with 1×1 convolution kernels for context information extraction across channels, and the network end with multiple parallel atrous convolution filters with different rate to capture objects and features at multiple scales. Finally, we combine the network output with a dense CRF to refine the final prediction results. The network in our study is trained and evaluated using three classical crack data sets. The experimental results clearly demonstrate that our method outperforms other approaches in terms of performance.