摘要
场景图生成技术存在谓词预测长尾分布和不合理预测等问题,为此提出一种基于场景辅助的场景图生成(Scene-Assisted Scene Graph Generation, SA)模型。在Faster R-CNN目标检测框架下,引入场景先验知识,结合图像视觉特征,通过双分支结构进行场景类别推理和谓词预测。实验结果表明,SA模型在谓词分类和场景图检测任务中性能良好,相比传统模型,谓词分类准确率提高了4个百分点,场景图检测提升了0.8个百分点,消融实验验证了双分支模块在提升模型性能方面的有效性。
Scene Graph Generation technology still has a predicate prediction long-tailed distribution and unreasonable prediction problems,therefore,a scene-assisted scene graph generation(SA)model was proposed.Under the Faster R-CNN framework,scene prior knowledge was introduced and combined with image visual features.Scene category inference and predicate prediction were performed through a dual-branch structure.Experimental results demonstrate that the SA model performs effectively in predicate classification and scene graph detection tasks.Compared with traditional models,predicate classification accuracy improves by 4 percentage points,and scene graph detection accuracy increases by 0.8 percentage points.Ablation experiments confirm that the dual-branch module effectively enhances model performance.
作者
顾非凡
周萌萌
宋世淼
葛家尚
杨杰
GU Fei-fan;ZHOU Meng-meng;SONG Shi-miao;GE Jia-shang;YANG Jie(College of Mechanical and Electrical Engineering,Qingdao University,Qingdao 266071,China;Qingdao QCIT Technology Co.Ltd.,Qingdao 266100,China)
出处
《青岛大学学报(自然科学版)》
2025年第4期39-45,100,共8页
Journal of Qingdao University(Natural Science Edition)
基金
山东省自然科学基金(批准号:ZR2021MF025)资助。
关键词
场景图生成
先验知识
目标检测
深度学习
scene graph generation
prior knowledge
object detection
deep learning
