Intelligent recognition of embroidered purse patterns: comparing YOLO series and RT-DETR

Digital preservation of traditional embroidered purse patterns faces challenges due to complex textures and dense distributions. This study constructs a specialized dataset of eight motif categories and evaluates the YOLO series against a transformer-based RT-DETR model. To address the unique characteristics of embroidery stitches and textures, the RT-DETR model was profoundly optimized by integrating the modern pure convolutional architecture ConvNeXt-Large as the backbone and incorporating a Focal Loss strategy.Experimental results demonstrate that the optimized RT-DETR achieves a mAP@0.5 of 0.5433, marking a 32.8% improvement over the YOLOv5m baseline. Specifically, for complex “Figures and Stories” patterns, the model reached an AP of 0.833, effectively reducing missed and false detections in intricate textile scenarios. Statistical analysis confirms the significance of these gains (95% CI: [0.538, 0.548]). This research establishes a benchmark for detecting dense heritage patterns and provides robust technical support for the intelligent archiving and digital transmission of intangible cultural heritage.