Abstract: Surface defect recognition of fuel rod end plug welds is critical for nuclear fuel manufacturing quality control, as defects such as poor forming and oxide color directly affect the safety and service life of nuclear fuel assemblies. Mainstream defect recognition methods rely on deep learning models that require large manually annotated datasets. However, manual annotation of weld defects is time-consuming, labor-intensive, and prone to subjective errors due to subtle defect differences and complex backgrounds, limiting the efficiency of automatic recognition. To address these issues, achieve automatic and accurate defect recognition, and reduce reliance on manual annotation, this study proposed an annotation-free method integrating traditional image processing and deep learning-based semantic segmentation. In this study, a multi-step approach was adopted. Firstly, the Sobel edge operator extracted edge features of weld images, followed by connected component analysis to locate weld regions. Morphological operations eliminated noise and optimized edge contours to improve localization accuracy. Secondly, images with accurate weld localization were selected to train a semantic segmentation model for detailed edge extraction, and predefined rules determined poor forming defects based on contour features. For oxide color defects hard to identify via edge-based methods, a region growing algorithm was used for coarse localization. Weld images were converted from RGB to HSV color space to enhance distinguishability of oxide color. Pre-experiments summarized the HSV range of oxide color to set seed points for the algorithm. Finally, images with accurate coarse localization were selected as the training set to train a deep learning semantic segmentation model for final oxide color defect determination. Data augmentation enhanced model generalization. Experimental results show that the proposed method significantly reduces manual annotation costs. Tests on a self-built dataset covering various defects under different conditions demonstrate good overall recognition accuracy. These results indicate that the method can reduce pixel-level annotation effort while maintaining acceptable segmentation performance on the collected dataset. This study indicates that integrating traditional image processing with deep learning provides a feasible annotation-efficient solution for surface defect recognition of fuel rod end-plug welds. The method saves annotation costs and exhibits strong robustness to complex backgrounds, meeting quality control requirements of nuclear fuel manufacturing. Its technical idea can be extended to other industrial weld defect detection fields, with important theoretical and practical value. Future research will optimize model structure to improve recognition speed and expand defect types.