基于特征关联的服装图像特征描述研究

Research on Garment Image Features Description Based on Feature Association

为了自动生成更全面的服装商品图像的特征描述,突破图像分类技术获取信息的局限性,建立一种基于特征关联的自然语言描述模型,以优化用户在网上购物的体验。根据服装描述语句的文本特点,对大量文本语料潜在的丰富信息进行分析,设计出改进的统计加权技术,在此基础上结合条件概率的方法最大化地获得高精度的商品文本特征,设计并构建关联规则与图像识别特征构建出更高级的语义概念,训练长短期记忆的自然语言描述模型,最后生成描述服装特征信息的语句。通过实验评估表明在多种模型上利用特征关联的方法对服装信息的特征描述在评测分值上都有显著的提高。

In order to automatically create a more comprehensive description of clothing commodity image features,and to break through the limitation of acquiring information with image classification technology,a natural language description model is established.This model is based on the feature association and can improve user’s online shopping experience.According to the text characteristics of clothing description sentences,the paper analyzes the potential rich information from a large amount of text data and designs a improved statistical weighting technique,based on which the text features are obtained combining with the conditional probability method.Then association rules associates text and image recognition features that construct a more advanced semantic concept to train the Long Short-Term Memory(LSTM)natural language description model.Finally,the trained LSTM model generated description statements of clothing features information.The experimental evaluations show that using the method of feature correlation to describe the features of clothing information improve evaluation score significantly.