Enhancing point clouds for multi-scale object monitoring in large railway infrastructure scenes

Point clouds are widely used in rail infrastructure. However, generating complete point clouds remains challenging due to device limitations and external factors such as occlusions, shadows, and surface reflectivity. These issues often result in sparse, noisy, and incomplete data, leading to the loss of essential geometric and semantic details needed for effective railway maintenance and inspection. Consequently, developing detailed, high-quality point cloud models for large-scale railway networks spanning hundreds of miles remains a challenge. We introduce the Edge-Aware Spatial Enhancement Network (EASE-Net), a hierarchical encoder-decoder framework that integrates multi-scale processing with spatial and color information to preserve both fine details and overall structural integrity. Additionally, to address limitations of existing loss functions, which treat all points equally, we propose the Context-Aware Point Cloud (CAPC) loss function. Experiments on real-world railway datasets demonstrate that our method achieves the highest reconstruction performance, with Chamfer distances of 36.35–85.11 mm, Earth Mover’s distances of 26.50–62.07 mm, F1-scores above 0.979, and coverage rates between 99.38 % and 99.54 %. Most importantly, EASE-Net maintains structural consistency, with volume ratios close to one and Jensen-Shannon Divergence scores below 0.0007. The results show that the reconstructed point clouds preserve the original point distribution and avoid distortions, even under noisy conditions where traditional methods have difficulty recovering fine details. Overall, the proposed method can generate detailed 3D models from only a few images, enabling project managers to potentially monitor large-scale railway networks without requiring supercomputers.