基于结构感知花瓣拓扑图的细粒度牡丹识别

Fine-Grained Peony Recognition Based on a Structure-Aware Petal Topology Graph

摘要: 目的：针对牡丹图像中重瓣花冠遮挡、花瓣边界粘连和局部纹理相近等问题，构建基于花瓣拓扑关系的识别模型PTRGNet。方法：训练阶段采用SAM候选掩码与人工校正获得实例级花瓣标注；测试阶段由训练后的实例分割器自动输出花瓣候选。模型以花瓣实例为节点，融合空间形态特征和ROI语义特征，并将距离、方向角、重叠率、外接框IoU和面积差作为边属性输入边感知GAT，再与视觉分支门控融合。结果：在Peony-8上，PTRGNet取得67.8% Top-1准确率和67.1% F1分数；在Oxford Flowers-102上，弱伪拓扑设置取得89.8% Top-1和98.6% Top-5。结论：花瓣拓扑可为牡丹识别提供颜色、纹理之外的结构证据，但模型性能仍受实例分割质量影响。

Abstract: Objective: Peony recognition is often disturbed by dense petals, weak boundaries and similar local textures. This work builds a Petal Topology Relation Graph Network (PTRGNet) to introduce structural cues into fine-grained peony classification. Methods: SAM is first used as an annotation aid to produce candidate petal masks, which are then corrected manually for instance-level supervision. At test time, petal masks are generated by the trained instance segmenter without manual input. PTRGNet represents each petal as a graph node, combines spatial morphology with ROI semantic features, and describes pairwise relations using distance, direction, mask overlap, bounding-box IoU and relative area difference. These edge attributes are fed into an edge-aware graph attention module and fused with the visual branch through a gated head. Results: On Peony-8, PTRGNet reached 67.8% Top-1 accuracy, 66.9% precision, 67.3% recall and 67.1% F1 score under predicted masks, exceeding EfficientNet-B0 by 8.1 percentage points. Human-refined masks gave 70.2% accuracy, while zero-shot SAM masks gave 63.5%, showing the effect of mask noise. On Oxford Flowers-102, the weak pseudo-topology setting achieved 89.8% Top-1 and 98.6% Top-5 accuracy. Conclusion: Petal topology supplies useful structural evidence for peony recognition, although stronger validation still depends on reliable automatic masks and larger instance-level flower datasets.