基于卷积神经网络的电商图片物体识别

发布时间：2018-04-16 08:40

本文选题：物体识别 + 卷积神经网络　；参考：《电子科技大学》2016年硕士论文

【摘要】：电子商务自1999年在中国出现以来,飞速发展。中国电子商务市场到2013年末,交易规模达到了10.2万亿元,网络零售市场在这个时间,交易规模达到了18851亿元。现在人们可以随时随地地使用移动终端设备,在电商平台中选购自己满意的商品。如何让用户更快更方便地找寻到自己的想要的商品方式,在互联网信息的电商平台中显得尤其重要。现有的也是较为传统的商品搜索方式是利用关键字进行查询,查询方便且快捷。相对于文字提供的信息而言,图片有着其天然的优势,它可以将信息高质量的保存,移动终端的智能设备提供了强大的截屏和拍照功能,使得人们可以方便地获取商品图片。可以预见的是,以图搜物的方式将逐渐兴起。在使用图像来搜索电商商品时,首先通过物体识别提取出物体中的商品信息,再从商品信息中抽取出图像特征,最后使用图像特征来搜索相似商品图片。在搜索过程中使用物体识别技术能够带来大量好处,通过物体识别可以去除环境的干扰,让系统只关注于用户感兴趣的部分。在用电商图片搜索相似商品时,用户关心的是输入图片中的商品,而不是输入图片中的背景。在电商图片的物体识别任务中,遇到的物体类别通常是鞋子、裤子、上衣等,这些物体通常由多个部分组成,比如鞋子通常由左右两只鞋子组成,而传统的物体识别算法并不能很好的处理这种情况。因此本文针对这种情况,设计了一个基于组合备选物体的物体识别算法。首先通过旋转图片,减小物体实例偏转对物体识别的影响。然后将同一张输入图片旋转之后的多张图片,通过传统的物体识别算法识别出其中的备选物体。每个备选物体可能对应一个完整物体,也可能对应一个完整物体的一个个组成部分。最后通过图像特征聚类备选物体,找出图片中有多少物体,每个备选物体属于哪个物体,再通过能量函数确定每个物体的位置。从实验结果中可以看出本文模型在物体类别识别的精确度稍稍高于Faster R-CNN。而在物体位置的预测上,本文模型的表现要明显优于Faster R-CNN。考虑到图像搜索对于物体位置的要求更高,综合来说,本文模型要比Faster R-CNN更好。
[Abstract]:Since the emergence of electronic commerce in China in 1999, the rapid development.China's e-commerce market reached 10.2 trillion yuan by the end of 2013, compared with 1.8851 trillion yuan in the online retail market.Now people can use mobile terminal devices anytime, anywhere in e-commerce platform to choose their own satisfied goods.How to make users find their desired goods more quickly and conveniently is particularly important in the e-commerce platform of Internet information.The traditional commodity search method is to query by keyword, which is convenient and fast.Compared with the information provided by the text, the picture has its natural advantage, it can save the information with high quality, and the intelligent device of the mobile terminal provides the powerful function of screen capture and photo taking, which makes it convenient for people to obtain the commodity picture.Predictably, there will be a gradual rise in search methods.When using images to search for e-commerce commodities, first of all, the commodity information is extracted by object recognition, then image features are extracted from commodity information. Finally, image features are used to search similar commodity images.The use of object recognition technology in the search process can bring a lot of benefits, through object recognition can remove the interference of the environment, so that the system only focuses on the part of interest to the user.When consumers search for similar items, the user is concerned about the items in the input picture, not the background in the image.In the object recognition task of ecommerce images, the objects encountered are usually shoes, trousers, jackets, etc. These objects are usually made up of several parts, such as shoes, which are usually made up of two shoes left and right.But the traditional object recognition algorithm can not deal with this situation very well.In this paper, an object recognition algorithm based on combinatorial alternatives is proposed.Firstly, the effect of object deflection on object recognition is reduced by rotating images.Then, several images of the same input image are rotated, and the candidate objects are identified by the traditional object recognition algorithm.Each alternate object may correspond to a complete object, or it may correspond to a single component of a complete object.Finally, by clustering the candidate objects with image features, we find out how many objects are in the picture, which object each candidate belongs to, and then determine the position of each object by the energy function.It can be seen from the experimental results that the accuracy of this model in object classification recognition is slightly higher than that in Faster R-CNN.In the prediction of object position, the performance of this model is obviously better than that of Faster R-CNN.Considering that image search requires higher position of objects, the model is better than Faster R-CNN.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183

【相似文献】