基于互补特征和类描述的商品图像自动分类

Auto Classification of Product Images Based on Complementary Features and Class Descriptor

实现电子商务中的在线商品自动分类是电子商务智能化的迫切要求。该文研究如何运用图像特征和分类算法对商品图像的一些具体信息进行自动分类,如长袖衬衫和短袖衬衫、圆领T恤与V领T恤等。图像特征采用了具有互补特性的塔式梯度方向直方图(PHOG)和塔式关键字直方图(PHOW)相结合;在分类器设计方面提出了基于图像类描述的改进最近邻分类算法。实验结果证明本算法能使2类和3类商品图像分类正确率达到70%-99%,且能够实现快速实时分类,相对于现有方法有了明显提升。

Auto-classification of online goods is a great need for intelligent e-commerce.Some valuable information of product,such as long sleeve T-shirt vs short one,round collar vs V-neck collar,can be tagged based on the image features and classification algorithms.As two complementary features,PHOG（Pyramid Histogram of Orientated Gradients） and PHOW（Pyramid Histogram Of Words） are adopted to extract and describe the features of product-images.An improved nearest-neighbor classifier based on the class descriptor of images is proposed.Experimental results show that the accuracies have been achieved between 70% to 99% on 2 or 3 real-time classification task,which is markedly improved compare to the existent result.