A novel material attribute annotation model based on gradient boosting decision tree (GBDT) and hybrid transfer learning is proposed.A new material attribute dataset named MattrSet is created firstly.Several features i.e.LBP, Gist, SIFT are extracted to characterize the images.GBDT algorithm is applied to complete annotation by optimizing the annotation model based on the log-likelihood loss function.Furthermore, a hybrid transfer learning strategy is designed to resolve the problem of missing data or unbalanced data.Experimental results show that:1) GBDT algorithm improves the annotation performance about 2.78％ compared to the best competitor before transfer learning.2) Basic transfer learning strategy further improves the annotation performance about 11.02％ compared to the best competitor.3) A reasonable combination of the models after using the hybrid transfer learning strategy further improves the annotation performance about 22.5％ compared to the best basic transfer learning, and promotes the accuracy about 16.80％compared to the best competitor.%提出了基于梯度提升决策树(gradient boosting decision tree,GBDT)与混合型迁移学习策略的材质属性标注模型,创建全新的材质属性数据集MattrSet,提取图像LBP、Gist、SIFT特征;引入GBDT算法,基于对数似然损失函数优化标注模型,实现图像的材质属性标注;设计混合型迁移学习策略,弥补样本缺失或不平衡问题,并进一步改善标注性能.实验结果表明:迁移学习前,梯度提升决策树算法的标注性能较最强基线提升2.78％;执行基本迁移学习策略后,标注性能比迁移学习前提升11.02％;合理地组合模型并执行混合型迁移学习策略,标注性能比基础迁移学习提升22.5％,较最强基线提升16.80％.
展开▼
