應用在限制類別搜尋整合與視覺化服飾檢索之研究

Effective Approaches for Category-Constrained Search Integration and Visual Clothing Search

林淇翔(Chi-Hsiang Lin)

指導教授 ： 蔡志忠

國立中正大學/工學院/資訊工程學系/博士(2013年)

摘要

對於資訊來源而言，使用限制類別查詢在已經分類好的目錄底下，是一種非常普遍的方法。它可以節省搜尋範圍而已可以增加效能。已經有越來越多的線上網站採用此方法，此方法困難在需要建立一個meta-search的系統，可以提供使用者只需要下一次的查詢就能把所有的網站資訊結果找回來並整合。

一個基本的問題是去建立一個像是meta-search的系統，從類別提供的資訊來源裡去對應所選擇的類別在一個網頁目錄底下做類別對應的動作。值得注意的是類別選擇查詢可以減少搜尋範圍且可以改善搜尋時間及結果。對於一個整合購物的搜尋系統網站來說，能夠快速且有效的回應是一個非常重要的事。

在另一部分中，衣服產品中含有許多的特徵，這是很難描述關鍵字，如質地，形狀，或關係對服飾的描述。基於這個問題，我們開發了一個基於圖像的視覺化服飾檢索系統，對服飾的影像提取和使用的特徵來找很難通過文字來形容，沒有文字解釋的衣服。

在這篇論文中，我們研究了分類網頁自動對應的問題，並對服飾檢索系統提供四種視覺化的搜索條件。我們把對應的問題轉換為一個搜尋的問題，且提供了一個有效的演算法來找到最佳的對應。我們的方法不僅可以使用在單層的網頁架構，也可以應用在多層的網頁目錄裡。為了評估我們的方法，我們從網站上搜集的很多的網頁。實驗結果顯示我們的方法是有效的，而已可以應用在整合限制類別的搜尋。在服飾檢索的實驗，我們的實驗圖像有1043張圖像，結果顯示我們方法是更利於挑選服飾產品。

並列摘要

Category-constrained search in which a category is selected to restrict the query to match instances in that category is a very popular mechanism
provided by information sources. It reduces the search space, and improves the response time and the quality of query results. As more and more online
sources are available, it is challenging to build a meta-search system which provides a unified directory and a meta-search capability to search and
access all sources from different websites in one query submission.

One of the fundamental problems in building such a meta-search
system is category mapping which maps the selected category in the
unified directory to categories provided by the information sources.
Note that category selection reduces the space to search, and
improves the query time and quality of the query result. Quick and
high quality response from each shopping site is very crucial to the
performance of an integrated shopping search system.

In the other part, clothes products contain many characteristics which are difficult to describe in keywords, such as texture, shape, or the relationship
between object and space. Based on this issue, we develop an image-based visual clothing retrieval system, which extracts and uses the features of
clothing images to find objects that are difficult to describe by text, or without text annotations.

In this thesis, we study the problem of automatic web taxonomy mapping and propose a part-based retrieval system framework, which provides four visual
search conditions. We formulate the mapping problem as a search problem, and present an efficient algorithm that aims to find the optimal mapping. Our
method can be used in flat mapping, but also use in hierarchical mapping. To evaluate our approach, we carry out experiment on data sets collected from
many Web sites. Experiment shows that our approach is very convincing, and can be used to implement automatic web taxonomy mapping for integration of
category-constrained search. In clothing experiment, we experiment on an image database with about 1043 general-purpose images, results showing that our
approach is more helpful for picking clothes product.

並列關鍵字

Information Retrieval ； directory mapping ； meta-search ； category-constrained search ； text categorization ； ontology mapping

參考文獻

[1] Bin He, Kevin Chen-Chuan Chang. Statistical Schema Matching across

(SIGMOD 2003), San Diego, California, June 2003.

Matchings across Web Query Interfaces: A Correlation Mining Approach.

Washington, August 2004.

[7] Choi, N., Song, I-Y. and Han, H. (2006) ’A Survey on Ontology Mapping’,