MCSNet:Multi-Scene Crack Segmentation Network Based on Few-Shot Learning

Crack segmentation networks heavily rely on large amounts of high-quality annotated images as training data. However, image labeling for pixel-wise is tedious and costly. Moreover, networks trained on crack images from some specific scenes struggle to generalize to other scenes. This paper proposes a few-shot segmentation approach for crack segmentation which performs highly accurate segmentation with only a few annotated images and few-shot learning enables network to segment multi-scene cracks. The network consists of an adjusted spatial convolution which extracts features of narrow target better, and a feature fusion module which enriches query features with multi-level and multi-source features. Additionally, an attention mechanism is incorporated to provide guidance for the query image segmentation. Extensive experiments on the simple public crack datasets such as Crack500, DeepCrack and a more complex self-collected dataset of concrete surface cracks demonstrate that the network outperforms other mainstream few-shot segmentation models.