基于多特征索引的服饰检索和后验证过程

发布时间：2019-03-26 13:27

【摘要】：随着电子商务的快速发展,新兴的网上购物模式已经慢慢的的取代传统的实体店购物模式。尤其体现在选购服装的方面上,网上购物比起去实体店来回奔走的选购更加的体现出其优势来。我们只需通过点击鼠标,就可以实现足不出户就能够在网络上浏览各式各样的服装同时在海量的服饰中挑选出自己喜欢的服饰来。但是现如今的各大购物网站在服饰搜索上大部分都是基于产品的标签的形式来检索的,即通过事先对服饰加入标签,检索的时候通过关键字对服饰进行检索。由于在语言描述方面上存在着很大的主观差异性,同时对于我们在大街上看到别人穿的漂亮的服饰或者电影电视上或者网络上看到我们喜欢的服饰时,我们想要在网上找到相类似的服饰时,我们不可能描绘的那么准确,进而搜索出来与之相同款式的服装。所以设计一个通过目标图像来检索相似图像的系统就显得格外的重要,人们只需要把想要查询的服饰图像输入就可以很精确的检索出来相类似的服饰。本文提出了基于多特征索引的服饰检索和后验证过程,通过实验结果可以看到在大量的图像库中该方法体现出来了一个很好的准确度。文章主要是由如下四部分组成:(1)第一部分工作是对图像进行预处理阶段通过SMQT特征和SNOW分类器实现人脸检测。当输入一幅查询图像的时候,通过实现人脸检测的方法,图像中人脸的位置会被检测出来。然后知道了人脸的位置后就能大致的确定服饰的位置。当我们判断出输入图像中服饰的位置后由Grab Cut算法对图像进行分割得到图像中服饰的区域。图像经过预处理之后只保留了原图像中的服饰区域。(2)在传统的BOW检索框架,我们加入了颜色描述符(CN),增强了颜色特征的匹配。同时将sift特征和颜色特征加入到二重多维索引的框架中进行特征的匹配。(3)针对BOW模型量化后的视觉单词降低了局部区域的辨别能力同时也不具备特征间的几何关系,所以我们提出了基于特征尺度的几何后验证方法来验证错误的匹配特征。这种后验证方法是根据特征之间的几个关系来去除误匹配点从而对于候选图像进行重新打分的过程。(4)通过实验来验证我们提出的方法并且横向的比较在检索框架中每进行的改进对实验精确度的影响。实验结果可以得出在大量的图像的数据库我们的方法取得了很好的效果。
[Abstract]:With the rapid development of e-commerce, the new online shopping mode has slowly replaced the traditional physical shop shopping mode. Especially reflected in the selection and purchase of clothing, online shopping compared to physical stores to go back and forth shopping more reflected its advantages. By clicking on the mouse, we can make it possible to browse all kinds of clothing on the Internet and pick out the clothes we like from the vast amount of clothing. But nowadays, most of the shopping websites are based on the form of product labels, that is, by adding labels to the clothing in advance, and searching the clothing by keywords. Because there are a lot of subjective differences in language description, and when we see other people's beautiful clothes on the street, or when we see our favorite clothes on movies, television, or the Internet, When we want to find similar clothing on the Internet, we can't describe it as accurately as we can, and then search for the same style of clothing. So it is very important to design a system to retrieve similar images by target images. People only need to input the clothing images they want to query and then they can retrieve the similar clothing accurately. This paper presents a multi-feature index-based clothing retrieval and post-verification process. The experimental results show that this method has a good accuracy in a large number of image databases. This paper mainly consists of the following four parts: (1) the first part of the work is to pre-process the image through SMQT features and SNOW classifier to achieve face detection. When a query image is inputted, the position of the face in the image will be detected by the method of face detection. Then know the location of the face will be able to roughly determine the location of the dress. When we determine the position of the dress in the input image, the Grab Cut algorithm is used to segment the image to get the region of the dress in the image. After pre-processing, only the dress regions in the original image are preserved. (2) in the traditional BOW retrieval framework, we add the color descriptor (CN), to enhance the matching of the color features. At the same time, the sift feature and color feature are added to the frame of double multi-dimensional index to match the features. (3) aiming at the quantized visual words of BOW model, the discrimination ability of local region is reduced and the geometric relationship between features is not possessed. Therefore, we propose a geometric post-verification method based on feature scale to verify the wrong matching features. This post-validation method is a process of removing mismatched points and re-scoring candidate images according to several relations between features. (4) the proposed method is verified by experiments and the horizontal comparison between the proposed method and the cross-sectional comparison in the detection process is carried out. The effect of each improvement in the cable frame on the accuracy of the experiment. Experimental results can be obtained in a large number of images in the database of our method has achieved good results.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】