Abstract: Nuclear fuel rods operate under harsh operational conditions in nuclear reactors. Their surfaces act as the first safety barrier of the reactor. Accurately detecting surface defects on these nuclear fuel rods is crucial for the safe operation of nuclear reactors. However, this task is confronted with challenges due to the unstructured nature and variable scales of defect features. This study aims to develop an intelligent defect detection algorithm for the surface appearance inspection of nuclear fuel rods to enhance the accuracy and robustness of the detection model, specifically addressing the aforementioned challenges. To address these issues, a novel hierarchical adaptive feature interaction network (HAFI-Net) for defect detection was proposed. Specifically, the network incorporated a novel shape-adaptive feature extraction module (SAEM) within its backbone. This module effectively enhances the model’s adaptive extraction capacity for unstructured defect features by integrating a multi-scale deformable convolution (DCN) structure with a stacked residual block structure, laying the foundation for efficient feature fusion in subsequent stages. In the Neck network, a hierarchical multi-level feature global-local perception enhancement network (HGLPN) was designed. This network first hierarchically fused the multi-level feature maps generated by the backbone, using a hierarchical multi-level path aggregation network (HML-PAN), to achieve sufficient cross-level information interaction. Subsequently, a global-local representation synergy module (GLSM) was employed to fully exploit the fused multi-level features. This process significantly enhances the model’s global context awareness while precisely capturing the positional information of local details, thereby remarkably improving the model’s detection accuracy for multi-scale defects. To validate the effectiveness of the proposed network, a dataset of nuclear fuel rod surface defects (including scratch, dent, and foreign material) was constructed based on 1 392 collected images. Extensive comparative and ablation experiments were conducted. Comparative results demonstrate that HAFI-Net achieves superior performance compared to existing state-of-the-art methods, with 91.2% in precision, 88.5% in F1-score, 92.0% in mAP50, 65.0% in mAP50∶95, 35.8% in mAP for small defects (mAPs), and 70.0% in mAP for large defects (mAPl). The model also meets real-time detection requirements and exhibits excellent performance in detecting various defect types, accurately identifying scratches, dents, and foreign materials. Ablation experiment confirms the synergistic effectiveness of the proposed modules. SAEM and HGLPN significantly improve the detection performance for unstructured and multi-scale defects, respectively. The overall model outperforms the baseline network across all mAP series metrics and substantially reduces localization and false negative errors. In conclusion, the proposed HAFI-Net exhibits significant advantages for the task of surface defect detection on nuclear fuel rods, achieving a good balance between detection accuracy and computational efficiency. This work holds important technical value for advancing the intelligent quality inspection of nuclear fuel rods and ensuring the safe operation of nuclear reactors.