Abstract: Nuclear power plant construction imposes stringent quality requirements on welds, underscoring the significance of utilizing intelligent methods to replace manual radiographic inspection. However, the image quality of X-ray photographs taken at nuclear power sites is generally low, leading to reduced accuracy in intelligent detection of weld defects.In addressing cold cathode X-ray images, this study maps CT values to grayscale values, then employs grayscale inversion and the CLAHE algorithm to enhance image contrast, thus processing X-ray data into clearer images. For weld defect detection,the study modifies the YOLOv8n model by adding a layer for detecting small objects, replacing the boundary loss function with Wise-IoU, and incorporating the GAM attention mechanism to boost the network’s feature extraction capability.Experimental results indicate that, through two-phase optimization of both image and model, the improved defect detection model raised the mAP@0.5 ∶0.95 by 29.8% relative to the original model. With accuracy nearing that of the more complex YOLOv5x model, the parameter count and floating-point operations are only 4% and 6% of the latter, respectively, and the model size is under 7 MB, facilitating deployment on portable, low-performance devices in industrial settings.