基于多特征融合和深度学习的商品图像分类

Classification of commodity image based on multi-feature fusion and depth learning

针对现有单一特征描述及浅层结构分类算法分类精度不高等问题,提出一种基于图像内容特征的深度置信网络(DBN)商品图像分类算法。对于从图像中提取到的颜色、纹理和形状等特征进行融合,构建5层DBN分类器对所得的特征数据进行学习训练和分类。采用京东商城提供的商品图像库,通过训练权重进行测试,测试结果表明,该算法在时效性和精确度方面优于使用单一特征的分类算法以及其它主流分类算法。

Taking the single feature and the major classification algorithms into consideration,an image classification algorithm based on fusion of multi-feature for deep belief network（DBN）was proposed to classify the product images.The multi-feature algorithm was used to fuse the features of image color,texture and shape.The samples were trained and classified using the constructed DBN with five levels.The proposed method was evaluated on the product image data set by train weight provided by Jingdong Mall.The results show that the proposed algorithm is better than single feature algorithm and other mainstream algorithms in terms of accuracy and timeliness.