Enhanced Crack Detection with Limited Annotations: A Mixed Supervision Algorithm With Multi Attention Network

Crack detection is essential for maintaining the structural integrity and safety of infrastructures such as roads and buildings. However, traditional manual inspection methods are inefficient, while existing machine learning approaches rely on large-scale labeled datasets, which are both costly and time-intensive. To address this challenge, we propose an enhanced Mixed Supervised Learning (MSL) algorithm, which improves performance and reduces dependence on extensive labeled data by integrating multiple learning techniques. Additionally, we introduce an enhanced deep learning model, Efficient Multiscale Transformer v2 (EMTv2), which leverages multiscale feature extraction and advanced attention mechanisms to improve crack classification and segmentation. The enhanced MSL algorithm achieves competitive segmentation accuracy while utilizing only 25% of the labeled data required by conventional weakly supervised learning methods, surpassing state-of-the-art models across three benchmark datasets. Our approach significantly enhances both efficiency and segmentation quality, providing a scalable and practical solution for real-world crack detection applications.