Runze Wu

The quality and feature representation of crack images are critical for crack detection. Traditional RGB-based methods are intuitive but highly susceptible to environmental interferences. Multispectral imaging enhances detection robustness through spectral stability and band complementarity. Although fusing RGB with infrared or other multispectral bands has been applied to crack detection, existing methods often select fixed bands, lacking an adaptive band selection strategy that generalizes across environments. Additionally, most fusion approaches focus on mitigating environmental interferences under specific crack conditions, leading to poor model transferability and difficulty handling compound interferences such as illumination, temperature, occlusions, and complex backgrounds. Moreover, the superposition of environmental interferences poses greater challenges for feature extraction, boundary enhancement, image stabilization, and resolution improvement in sensitive area fusion. This paper proposes a multispectral fusion approach based on adaptive band selection and multi-level decomposition (MFAM). First, a multi-criteria adaptive band selection algorithm is proposed to extract the most discriminative spectral bands using environment-aware weighting. Second, the dual-scale transform-domain dual-nested decomposition strategy mitigates feature loss and fusion instability caused by poor environmental transferability, establishing hierarchical representation spaces via multi-level frequency band decomposition to precisely integrate multi-scale information. Finally, to mitigate boundary blurring and small-scale feature loss due to low spatial resolution, a spectral-guided enhancement re-fusion method is proposed. By re-fusing the preliminary fused image with the optimal band, edge details are refined, and spatial resolution and clarity enhanced. Experiments on some multispectral crack datasets demonstrate the superior environmental adaptability and image detail enhancement of the proposed MFAM.

This study presents a multimodal data fusion framework for road damage detection and prediction, integrating RGB images, LiDAR point clouds, and GPR (Ground Penetrating Radar) data to enable high-precision detection of surface cracks, potholes, and underground voids, as well as dynamic trend forecasting. By leveraging Deep Mapping 2.0 and the RAFT algorithm, the alignment accuracy between RGB and LiDAR data was significantly improved, reducing registration error to 2.3 mm. Concurrently, the spatial mapping accuracy of GPR data was enhanced to 4.8 mm, ensuring precise multimodal data fusion. A Cross-Attention Transformer combined with a Feature Pyramid Network (FPN) was used for dynamic feature weighting, achieving a crack detection IoU of 97.3% and an AP@0.5 of 93.7% for underground void detection, thereby substantially enhancing the model's performance in detecting complex road damage. Moreover, a trend prediction model integrating ConvLSTM and a spatiotemporal attention mechanism achieved an MAE of 8.7% in a six-month damage trend prediction experiment, reducing prediction error by 34% compared to existing methods, underscoring the model's effectiveness in forecasting damage progression.The experimental results demonstrate that the proposed framework exhibits strong adaptability and stability across diverse road damage detection tasks, particularly excelling in the joint detection of cracks and underground voids with high accuracy. Furthermore, the framework is readily extendable to infrastructure health monitoring applications, such as bridges and tunnels, providing robust technological support for intelligent road maintenance and offering data-driven insights for the long-term optimization and sustainability of urban transportation infrastructure.

In multispectral image fusion, edge distortion, spectral loss, and geometric mismatch seriously affect the fusion accuracy, especially in complex road scenes with shadows or occlusions. Multispectral image fusion aims to preserve surface details and spectral data. In order to solve the problems of edge distortion and spectral loss in image fusion, this paper proposes a multispectral and hyperspectral fusion method based on edge chromatic aberration. Swin Transformer is used for multi-scale feature extraction, and the GRDB module is added to preserve edge details, which improves the clarity and accuracy of diseased edges in road scene fusion images. In addition, saliency weight mapping can identify and highlight key disease areas, ensuring that they are prominent in the fused image.Experiments show that the multispectral image fusion method based on edge color difference significantly improves the performance of road damage analysis on the self-built BUCEA-MS-Road-Damage dataset: the edge IoU in the detection task is increased to 80.1% (+1.3%), and the target detection accuracy is 92.3% (+3.6%); the accuracy and recall of the classification task are increased to 91.3% (+3.0%) and 89.8% (+3.0%) respectively; the Dice coefficient of the segmentation task is 83.3% (+3.0%). In the cross-sensor test, the fusion result is still robust (SSIM=0.93, SAM=2.7°), and the edge color difference index (ECD-Index=6.3) is 25.9% lower than the baseline. This method effectively solves the problems of multi-scale feature extraction and texture distortion of cracks through adaptive color difference correction and spectral consistency constraints, providing high-precision data support for intelligent road maintenance.

To address the issues of uneven feature point distribution, environmental interference, and insufficient real-time performance in UAV image stitching in complex urban environments, this paper proposes an improved ORB algorithm based on Gaussian scale-space optimization and dynamic grid division, combined with a global geometric consistency optimization strategy. First, local adaptive noise filtering and bilateral filtering are applied to enhance image quality. Then, multi-scale feature detection is achieved using a Gaussian scale-space pyramid, and dynamic grid division is employed to balance feature point distribution. Finally, a global energy function, including reprojection error and smoothness constraints, is constructed to iteratively optimize the homography matrix and suppress stitching distortions. Experimental results show that the proposed method achieves high processing speed on low-performance hardware platforms, improves feature point distribution uniformity to 0.89, and achieves stitching accuracy (RMSE) of 3.5 pixels, significantly outperforming ORB and SIFT algorithms, while remaining robust in dynamic occlusion and lighting variation scenarios. This method provides a lightweight and efficient solution for UAV image stitching in urban environments, supporting applications such as urban planning and disaster assessment. Future work will explore lightweight deep learning integration and edge computing acceleration to further improve dynamic scene adaptability.