Abstract: For the problem that there is scarce image annotation data and difficult feature extraction in the cracks of key parts such as chimneys, walls and metal equipment in thermal power plants under complex lighting conditions, a crack detection model named PowerCrackNet is proposed, which is based on a dual-branch prototype metric network with few-shot learning. The network integrates the reflection rate information irrelevant to lighting in the image using the Retinex theory, guiding the network to learn the invariance of lighting. During the feature extraction stage, a prior mask containing crack location and structural information is generated by measuring the similarity between query and support features in a high-dimensional space. Then, in the feature interaction stage, the convolution attention and atrous spatial convolution pooling pyramid mechanism are employed to integrate the feature information and prior information from different scales, thereby improving the spatial inconsistencies among samples in few-shot learning. Finally, in the prototype metric stage, a complementary attention mechanism is used to enhance the expressiveness of support prototype vectors for crack categories. Quantitative comparisons and visualization analysis on the public low-illumination crack dataset LCSD demonstrate that the proposed method achieves stronger crack segmentation performance under low-light conditions with limited samples(1-shot), outperforming the current other method. Further validation in real-world scenarios confirms that the proposed approach can provide efficient and reliable technical support for the early warning of cracks in key parts of thermal power plants, highlighting its significant practical value.

Key words: crack detection, few-shot learning, deep learning, thermal power plants