【作者】 高健；

【导师】 杨小平；

【作者基本信息】 江西财经大学 ， 工程硕士（计算机技术领域）（专业学位）， 2016， 硕士

【摘要】 随着互联网技术的快速发展,在线购物已经成为一种潮流和趋势,越来越多人选择在京东、淘宝等电子商务平台上进行商品的选择和购买。各大电子商务平台为了能够更好的满足用户的需要,推出了越来越多的新颖的服务性功能,商品相关性推送就是其中一个。为了能够更好的对平台上的商品进行管理和分类,同时能够根据用户的浏览记录对相关商品进行推送,商品图像分类技术被提出,同时也成为了目前的一大研究热点。在早期的一些研究中,学者们都是采用人工手动分类或者是基于文本内容的图像分类方法,这些方法不仅需要耗费大量的人力和物力,同时在分类精确度方面仍未达到良好的效果。鉴于这样的问题,后来研究者们提出了基于图像内容的分类方法,基于图像内容分类方法是根据图像内容所具有的相关特征进行特征提取,然后根据这些特征进行图像的分类。目前,常用于提取图像内容特征的方法主要是基于图像的底层视觉特征,如图像的颜色、纹理、形状和相关空间关系等。这一方法在对图像进行分类时,较之先前的分类方法有了明显的改进。因此,对图像内容进行特征提取和分类仍是主流的研究领域。“基于多特征融合和深度学习算法的商品图像分类”也就是利用图像的内容进行有关特征提取,以期获得对图像内容更本质的描述。多特征融合算法实际上就是先对图像的颜色,纹理和形状这三种基本特征信息进行特征提取,然后将这些特征根据多特征融合算法进行融合;将这一特征融合信息作为深信度网络模型的输入层数据,以便其根据这些特征数据进行有关的样本训练和分类。在利用该深度学习模型进行学习分类过程中,利用BP算法对该学习模型的预训练结果与训练样本标签数据进行差异比较,以此来作为深信度网络模型的权值更新依据,提高该算法的分类精度。最后,利用京东商品图像库对“基于多特征融合和深度学习算法的商品图像分类”这一算法进行实验分析。将实验结果与利用单一特征算法的实验结果进行比较,得出利用单一特征算法对测试样本库进行分类,所得的分类精确度在60%上下,而本文所提出的利用多特征融合的分类算法能够将分类结果提高到82.3%,同时保证了运行时效性;为了能够更加全面的对本文算法的有效性进行验证,将本文算法的实验结果与其他主流的几大分类算法实验结果做对比分析,结果表明本文算法较之其他主流的算法都有明显的分类精确度优势,而且它的平均每张图像的处理时间也不会较之其他算法有明显的差距。综上所述,本文所提出的算法对测试样本库进行分类,在时效性和精确度方面都有较好的结果。更多还原

【Abstract】 With the rapid development of Internet technology,online shopping has become a trend and trends,more and more people choose to select and buy goods in Jingdong,Taobao and other e-commerce platform.Major e-commerce platform in order to meet the needs of users,more and more new service has be launched,the function of product recommendation is one of them.In order to manage and classification of goods better,and according to users browsing history to push related products,product image classification techniques have been proposed,but also become a major research focus at present.In some of the early studies,scholars were used the methods of manual classification or image classification method based on text content,which not only requires a lot of manpower and material resources,and the classification accuracy has yet to achieve good results.In view of such problems,the researchers have proposed classification method based on image content,Content-based image classification method is based on the relevant characteristics of the image content having feature extraction,and then accordance with these characteristics to classify the image.Currently,commonly used in image content feature extraction methods are mainly based on low-level visual features of images,such as color,texture,shape and Related spatial relationship,etc.This method has been significantly improved over the previous classification,Accordingly,the image classification method based on image content is still the mainstream of research field."Product Image Classification Based on Multi-feature Fusion and Deep Learning Algorithm",Using the content of the image to extract features,in order to obtain a more nature description of the image content.Multi-feature fusion algorithm is use the basic visual information on the image to extract features,including image color features,texture features and shape features,then use multi-feature fusion algorithm to fuse these features.This information of the multi-feature fusion will be the data as the input layer of the Deep Belief Network,the model of Deep Belief Network will use this data for training and classification.With the process of the learning and classification based on the model of deep learning,using BP algorithm to update the weights of the model,to improve the classification accuracy of the algorithm.Finally,we use the Jingdong product image library to test and analysis "Product Image Classification Based on Multi-feature Fusion and Deep Learning Algorithm",The results of this algorithm comparing with the experimental results of using single feature algorithm,it is concluded that by using single feature algorithm to classify the test sample library,the classification accuracy obtained around 60%,and this article proposed the use of multiple features fusion classification algorithm to the classification results to 82.3%,at the same time to ensure the operation timely;In order to be more comprehensive to validate the effectiveness of the algorithm,The results of this algorithm comparing with several other major classification algorithm,the results show that the algorithm is compared with other mainstream algorithm has obvious advantages of classification accuracy,and its average processing time per image compared with other algorithms do not have obvious difference.In summary,the proposed algorithm for the test sample library classification,in terms of accuracy and timeliness have better results.更多还原