Integrated 3DGS with Enhanced YOLO Network for Spatiotemporal Defect Monitoring on Aircraft Skin

With the transition of civil aviation maintenance towards predictive maintenance, traditional General Visual Inspection (GVI) suffers from the disadvantages of low efficiency and high subjectivity, as the existing 3D inspection techniques rely on expensive equipment, and 2D detection methods lack spatiotemporal analysis capabilities. This paper proposes a method which integrates 3D Gaussian Splatting (3DGS) with an enhanced YOLO network for spatiotemporal defect monitoring on aircraft skin. By utilizing the videos recorded during GVI inspections, 3D reconstruction of aircraft surfaces is achieved through adaptive density control and differentiable rendering in 3DGS. The lightweight YOLO11n network is improved by introducing a dynamic feature fusion module to enhance multi-scale feature representation, where a multi-scale edge enhancement module is employed to improve contour recognition and an adaptive threshold focal loss is incorporated to optimize learning performance on difficult samples. Tested on a self-constructed dataset, the 3D reconstruction achieved an average Peak Signal-to-Noise Ratio (PSNR) of 29.1 dB, Structural Similarity Index Measure (SSIM) of 0.86, and a rendering speed of 12.9 frame per second (FPS); the enhanced YOLO network achieved an mAP0.5 (mean Average Precision evaluated at an intersection over union threshold of 0.5) of 85.0% and an FPS of 131.6. Testing on two public datasets yielded mAP0.5 scores of 77.8% and 83.5%, respectively, validating its generalization capability. This method effectively integrates the efficiency of 2D detection with the advantages of 3D information, reduces reliance on specialized equipment, supports subsequent spatiotemporal evolution analysis of defect, and provides technical support for full lifecycle health management of aircraft.