- 學位論文
透過Google圖片搜尋引擎作個人資料網頁搜尋
Search for Personal Web Pages through Google Image Search Engine
摘要
現今一般網頁搜尋引擎（例如Google或Yahoo）對於個人資料（包括個人臉部相片、服務單位、學歷、經歷、地址、電話號碼等）的搜尋，並沒有提供完善的搜尋功能，假如使用者想搜尋個人資料而輸入人名至搜尋引擎中，搜尋引擎將找出非常多不相關的資料，使用者往往需要花費不少時間來作資料的過濾篩選。在本研究中，我們提出利用圖片搜尋引擎來搜尋個人資料，當使用者輸入人名後，我們即呼叫Google圖片搜尋功能，取得Google所回傳的相關圖片，並提出了三種不同的圖片排序方式來重新排序Google所回傳的圖片，使使用者能快速從較前面圖片找到個人資料。在第一種排序方式中，我們分析Google回傳圖片所伴隨的文字描述，得到一個排序分數，然後以此分數當作排序的依據，將分數大的圖片排序在較前面的位置。在第二種排序方式中，我們利用圖片中所偵測出來的臉孔數目當作重新排序的依據，含有一個正面臉孔的圖片將被排序在最前面，然後依序是二個臉孔到多個臉孔，最後則是無臉孔之圖片。在第三種排序方式中，我們連結至Google所回傳圖片所在之原始網頁，若網頁內的文字含有較多的個人資料關鍵字，該圖片將被排序在較前面的位置。使用者可以依據個人對回應時間之需求選擇所要搜尋的方式，實驗結果也說明了本研究之可行性、有效性。
並列摘要
Current web-based search engines (such as Google or Yahoo) do not provide adequately for the search of personal data (including personal facial image, address, telephone number, etc.). If a user inputs a person’s name to a search engine, he/she may need to spend much time in finding the desired data from those returned by the search engine. In this research, we propose to find personal data based on an image search engine. A person’s name is used in the study as a query. After a user inputs a query, we call Google image search engine and obtain related images from Google. Three approaches are also proposed to rank images returned by Google such that the user may find the desired personal data from the fronter images. In the first proposed approach, we analyze for each image the string of words which is also returned by Google and then obtain a ranking score. The images are ranked based on the scores. Images with large scores are ranked in front. In the second method, we re-rank images based on the number of frontal faces in each image. Images with only one face are ranked at the front, and then images with two, three, and more faces are ranked. In the third method, we link to each web page on which an image returned by Google is put, and analyze the text on the web page. If the text has more key words about personal data, the image is ranked at fronter locations. Users may select one of the three approaches accordiny to their requirement in response time. Experimental results show the feasibility of the proposed approach.