A ground crack segmentation algorithm based on the fusion of attention aggregation mechanism and dense connection features

Abstract High-intensity ground operations in mining areas easily disrupt ecological balance and threaten ground safety, necessitating precise measurement and monitoring of surface deformation. Deep learning-based crack segmentation enables accurate localization and quantitative measurement of crack dimensions. Existing models mostly adopt large-scale improvement mechanisms or cumbersome modules to boost accuracy, but these introduce excessive parameters and computational costs. Conversely, lightweight models compromise segmentation precision, particularly for fine-scale crack measurement. Thus, this paper proposes the DP-MCUNet++ model for automatic and accurate extraction of fine ground cracks from unmanned aerial vehicle images, achieving an optimal balance between accuracy and efficiency. Building upon UNet++, it implements three key innovations: (1) removes the deep supervision pruning mechanism, reducing parameters by 15% while maintaining accuracy; (2) replaces standard convolutional modules with a dual-path encoder that separately captures semantic context and spatial details through parallel processing, enabling lightweight operation (10.25 M parameters) with diverse feature extraction; (3) introduces convolutional block attention module to integrate pixel spatial and feature channel dependencies for suppressing background interference, and designs multi-scale attention aggregation mechanism. For adaptive multi-scale feature aggregation to handle crack scale variations. The model is validated on three diverse datasets including our self-collected Huipodi mining area dataset, CrackTree200, and Crack-Flickr Dataset (CFD), demonstrating strong generalization capability across different crack types and environmental conditions. Experiments demonstrate that the proposed DP-MCUNet++ achieves 81.05% Recall, 83.17% F1-score, and 81.02% mean Intersection over Union on the Huipodi dataset, with an inference speed of 5.98 FPS. It outperforms traditional image processing methods and state-of-the-art deep learning networks while maintaining computational efficiency suitable for practical deployment. Cross-validation on two public benchmarks, CrackTree200 and CFD, confirms robust generalization capability. The method provides quantitative crack measurement accuracy of ±2.58 cm for width and ±3.87 cm for length at a ground sampling distance of 1.29 cm, meeting measurement standards for ground safety assessment and ecological protection.