Research on Defect Detection in Digital Radiographic Images of Fuel Rods Based on Gradient-Domain-Guided Multi-Scale Retinex and Improved YOLOv12

Abstract: Weld inspection during nuclear fuel rod manufacturing requires reliable detection of internal defects that vary widely in size and may have low contrast in digital radiographs. We present a detection workflow that combines gradient-domain guided multiscale Retinex (GDGF-MSR) with an improved YOLOv12 detector. The preprocessing stage uses gradient-domain guided filtering to estimate the background at multiple scales while preserving structural edges. This improves the visibility of faint Porosity and elongated regions of lack of penetration(LOP). The detector incorporates coordinate attention to encode direction and position, a high-resolution P2 branch to retain spatial detail for small defects, and a bounding-box regression loss that combines complete intersection over union with normalized Wasserstein distance. The workflow was evaluated on a production-line digital radiography dataset containing LOP, tungsten inclusions(Tungsten), and Porosity. In the model comparison, the proposed workflow achieved 96.5% mean average precision at an intersection-over-union threshold of 0.50 and 98.5% recall, with a measured processing rate of 5 frames per second(FPS). These results support the use of edge-preserving preprocessing and morphology-aware detection for automated radiographic inspection of nuclear fuel rod welds.