Abstract: A YOLOv8n-based system for the localization and detection of electron beam weldments was proposed to address the challenges of centroid localization and feature extraction in circular weldments during electron beam welding as well as the inefficiency of traditional visual inspection methods. The system was developed using Python as the primary programming language, combined with PyQt5 for designing the human-computer interaction interface. Deep learning and multi-threading technology were used to enable parallel processing of video streams and user interface update, thereby avoiding interface lag. Furthermore, the accuracy of labeling results was improved by enhancing the YOLOv8n pre-training model and employing an iterative labeling method. To tackle the issue of difficult feature extraction caused by high similarity between the features to be extracted and the background of circular weldment images, redundancy in model parameters was reduced through spatial and channel reconstruction units in the backbone. Normalization technique was used to adjust the mean and variance of each feature channel, a new low-order feature fusion module SCFF was finally designed, which retains key circular feature information while reducing model parameters and computational cost. The enhanced model achieves automatic circle center identification, weld seam positioning, and visualization of circle center coordinates. The experimental results demonstrate that the improved model increases average prediction accuracy by 1.8% and reduces parameter count by 18.7% compared to the original model, making it more suitable for deployment and embedded use.