<p>Given that garment-surface defects commonly exhibit small-scale appearance, low contrast, weak texture, and susceptibility to interference from complex backgrounds, existing general-purpose object detectors tend to suffer from missed detections and unstable localization in practical production-line scenarios. To address these issues, this paper proposes RDNT-YOLO, a parameter-efficient and high-accuracy method for small object defect detection on garment surfaces. Taking YOLOv8s as the baseline, RDNT-YOLO introduces improvements from three perspectives: detection hierarchy, small object representation, and regression robustness. First, a dynamic small object detection layer, DSODL, is designed, and DySample is incorporated to optimize feature upsampling and multi-scale fusion, thereby enhancing the detectability of fine-grained defects. Second, a C2f-RFAG module is constructed in the downsampling stage, where dynamic receptive-field aggregation and attention gating strengthen feature representation for complex textures and weakly salient defects. Third, an NWD–TDIoU combined regression loss is proposed to jointly account for scale sensitivity and localization accuracy of small objects, improving the stability of bounding-box regression. Experiments on a self-built garment-surface small object defect dataset demonstrate that RDNT-YOLO achieves 84.3% mAP@0.5 and 38.1% mAP@0.5:0.95, while maintaining 7.43&#xa0;M parameters and delivering 40.63 FPS real-time inference. Furthermore, RDNT-YOLO is deployed online on a garment small object defect inspection device, and the results show that RDNT-YOLO can generate defect detection and alarm outputs under practical production conditions, preliminarily verifying the feasibility of integrating the proposed method into garment surface defect inspection equipment. This online validation provides a prototype basis for subsequent production-line application and engineering-level testing.</p>