A low-cost regularization strategy for the 3D reconstruction of wind turbines based on remote sensing images

Unmanned aerial vehicle remote sensing images are easy to acquire and cost-effective. Their application provides a new approach for the maintenance of offshore wind farms. To obtain an accurate model of wind turbines, thereby improving detection efficiency and mitigating the risks associated with high-altitude workings, a low-cost method for wind turbine three-dimensional (3D) reconstruction is proposed. The main advantages of deep learning-based Multi-View Stereo Network (MVSNet) include high accuracy, fast processing speed, and ease of extension to multiple viewpoints. Most networks utilize multi-scale 3D Convolutional Neural Network (CNN) for regularization. However, when dealing with images that have high depth and resolution, CNN regularization strategy leads to a significant increase in memory usage. In this paper, the two-dimensional cost maps are regularized along the depth dimension sequentially using Stacked Convolutional Gated Recurrent Unit (Stacked Conv-GRU), instead of reducing the 3D cost volume all at once. Localized convolution operations replace undifferentiated fully connected operations, substantially reducing memory consumption and enabling high-resolution reconstruction. We conduct comparisons and model evaluations on the "Nordtank" wind turbine remote sensing image dataset. Experimental results demonstrate that, compared to other 3D reconstruction methods, Stacked Conv-GRU regularized MVSNet reduces run time and hardware requirements while maintaining similar accuracy.