毕业生知识库

  随着互联网的发展，网络上的图片信息成爆炸式增长，人们对于图像检索的需求越来越大，其中，对于商标图像的检索，无论在科研还是商业上都有极大的研究与应用价值。
本文围绕基于内容的商标图像检索展开了一系列的研究，提出了一种大规模商标图像数据集下的高效检索算法，并在公开数据集和实验室收集的千万量级的大规模数据集上进行了实验，并实现了一个商标检索系统。具体地，本文的主要内容如下：
  (1)   针对商标图像的局部匹配问题，本文采用一系列预处理方法，将图像边缘图分割成几个部件，再提取RIDE SIFT特征，并使用FV算法进行聚合描述，在公开数据集METU v2上的实验结果显示，该方法为当前性能最好的方法，甚至胜过一些深度学习方法。
  (2)   针对商标图像的语义相似问题以及纯文字型商标图像检索问题，本文采用深度学习方法提取CNN特征。主流的CNN网络和传统SIFT特征无法很好地表征纯文字型商标图，本文采用一种用于手写汉字识别的HCCR CNN，提升了网络的文字表征能力。
  (3)   针对数据量大时检索难、检索慢问题，本文采用了IMI索引加上OPQ特征编码的方法，OPQ编码能对特征进行降维并量化至汉明空间，IMI索引将传统KNN检索转变为ANN检索，保证精度的同时大大提升了检索速度。
  (4)   在Windows平台上使用C++实现了一个商标检索系统，并且本文提出的检索架构已经投入用于商标侵权检测的商业应用（详见标掌柜网址http://www.biaozhanggui.com/ ），目前上线半年，运行良好。

  With the development of Internet, the image information on the web has grown exponentially, and people’s demand for image retrieval is rapidly increasing. Especially, the search for trademark images has great research and application value both in scientific research and in commerce.
In this paper, a series of researches have been carried out on content-based trademark retrieval, and an efficient search algorithm for large-scale trademark image data sets has been proposed. We have conducted the experiments both on public dataset and our lab’s dataset with more than 10 million trademark pictures. What’s more, we implemented a trademark retrieval system in C++ on Windows platform. Specifically, the main content of this paper is as follows:

1. For the problem of partial matching of trademark images, this paper uses a series of preprocessing methods to segment the image edge into several components, extract proposals, extract RIDE SIFT features on proposals, and use FV algorithm to get aggregation. The experimental results on the public dataset METU v2 show that this method is state-of-the-art and even performs better than some deep learning methods.
2. For the semantic similarity problem and the problem of matching text trademark images, this paper uses deep learning method to extract CNN features. Mainstream CNN networks and traditional SIFT features do not characterize pure text-based trademark maps. This article adopts a HCCR CNN for handwritten Chinese character recognition, which enhances the network's textual representation capability.
3. For the problem of the difficulty low efficiency when searching the large-scale datasets, this paper uses the method of IMI index with OPQ feature coding. OPQ coding can reduce the dimension of features and quantize them into Hamming space. IMI index converts traditional KNN search into ANN search, which has greatly improved the retrieval speed while assuring the accuracy.
4. A trademark retrieval system was implemented using C++ on the Windows platform, and the search architecture proposed in this paper has been put into commercial applications for trademark infringement detection (see http://www.biaozhanggui.com/), for the time being, it has launched for half a year, and it is working well.