摘要
提出为商品图像标注句子,以便更准确地刻画图像内容。首先,执行图像特征学习,选出标注性能最优的梯度核特征完成图像分类和图像检索,该特征能客观描绘商品图像中形状和纹理这两类关键视觉特性。然后,基于语义相关度计算结果从训练图像的文本描述中摘取关键单词,并采用N-gram模型把单词组装为蕴涵丰富语义信息且满足句法模式兼容性的修饰性短语,基于句子模板和修饰性短语生成句子。最后,构建Boosting模型,从若干标注结果中选取BLEU-3评分最优的句子标注商品图像。结果表明,Boosting模型的标注性能优于各基线。
Product image sentence annotation was presented because sentence describes online products more accurately than single words.Firstly,image feature learning was executed.Gradient kernel feature that achieves the best annotation performance was chosen because the feature describes the key visual characteristics of product image such as shape and texture better than other features.Therefore,the gradient kernel feature was selected to complete image classification and image retrieval.Secondly,several key words were summarized from training images' captions based on semantic correlation computing.Thirdly,a modified sequence that not only contains rich semantic information but also satisfies syntactic mode compatibility was created based on these key words by N-gram model.Sentence was generated according to predefined sentence template and the modified sequence.Finally,a Boosting model was designed to choose those sentences that obtain the best BLEU-3 scores to annotate product images.Experiments show sentences generated by the boosting model achieve the state of art annotation performances.
出处
《计算机科学》
CSCD
北大核心
2016年第5期269-273,287,共6页
Computer Science
基金
国家自然科学基金重点项目(61133012)
国家社科重大招标项目(11&ZD189)
教育部人文社会科学研究青年项目(12YJCZH274)
江西省科技厅科技攻关项目(20142BBG70011
20121BBG70050)
江西省高校人文社科基金项目(XW1502
TQ1503)资助