| DC 欄位 |
值 |
語言 |
| DC.contributor | 資訊管理學系 | zh_TW |
| DC.creator | 李佳容 | zh_TW |
| DC.creator | Chia-Jung Lee | en_US |
| dc.date.accessioned | 2013-7-29T07:39:07Z | |
| dc.date.available | 2013-7-29T07:39:07Z | |
| dc.date.issued | 2013 | |
| dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=100423027 | |
| dc.contributor.department | 資訊管理學系 | zh_TW |
| DC.description | 國立中央大學 | zh_TW |
| DC.description | National Central University | en_US |
| dc.description.abstract | 隨著近年來網際網路的蓬勃發展，現今網路已成為現代人重要的資訊來源之一，因此如何快速且正確的搜索出所需要的資訊，在近年來便成為一門重要的課題，而使許多研究紛紛提出各種圖片檢索技術。其中一種常用的圖片檢索技術為以圖像內容作檢索的一種方法，稱為CBIR (Content-Based Image Retrieval)，但其具有低階圖片特徵無法完全表達高階語意的缺點 (semantic gap)，許多研究中皆提到相關式回饋技術(Relevance Feedback)為改善的方法之一。RF為CBIR系統在得到檢索結果後，讓使用者對Query圖片和檢索結果進行比較，並提供回饋以使下一次檢索結果較佳的方法，而虛擬相關式回饋技術(Pseudo Relevance Feedback)被提出，以電腦取代使用者以避免誤判情形發生。
基於PRF的假設，本研究提出一個演算法可以改良Feedback結果中，回饋資訊的影響力沒有隨著重要性而提升的缺點，稱為Block-Based Pseudo Relevance Feedback (BBPRF)。此演算法的概念為將Feedback結果中，每k個結果畫分為同一區間，並將結果乘以該區間的權重值，以所得到的新的結果做下一個回合(iteration)的回饋；該權重值隨著第一次檢索結果的排行遞減，以達到「排名越前面的結果，相對重要性越大」的目的。
本研究以常見的NUS-WIDE和Caltech256兩種資料集進行實驗，使用RF中常見的演算法之一Rocchio演算法進行比較。實驗一為探討本研究所出的改良演算法檢索結果是否有比傳統演算法的結果來的好，實驗二為加入使用者回饋進行檢索結果的探討。
實驗結果中，NUS-WIDE和Caltech256兩種資料集中，改良的演算法所得到的效果皆比傳統的虛擬相關式回饋演算法好，且過程所費時間並無過大差距。以此可看出本研究所提出之改良式的虛擬相關式回饋演算法具有更好的效能。 | zh_TW |
| dc.description.abstract | Nowadays the network has become one of the important ways to obtain information. Therefore, it is important to effectively search for information. For image search, CBIR (Content-Based Image Retrieval) is major technique. However, the semantic gap problem limits the performance of CBIR systems. In literature, RF (Relevance Feedback) can be used to improve the retrieval performance of CBIR systems. It is usually based on asking users to give feedbacks, and the retrieval results are re-ranked. One major limitation of RF is the need of the user in the loop process. To this end, PRF (Pseudo Relevance Feedback) was proposed that considers top-k images as the pseudo feedbacks to re-rank the retrieval results.
This thesis proposes an algorithm called Block-Based Pseudo Relevance Feedback (BBPRF) to improve the traditional PRF approach. The idea of this algorithm is to assign higher weights to higher ranked images. Particularly, top-k images as the feedbacks are divided into two to k blocks and each block has a specific weight, so the weighted feedbacks will benefit the next feedback iteration.
The experiments are based on the NUS-WIDE and Caltech256 datasets and the Rocchio algorithm is used as the traditional feedback algorithm. The first experimental results show that our proposed BBPRF performs better than the traditional PRF approach in terms of precision at 10, 20, and 50. In particularly, using top 30 images with 30 blocks perform the best. The second study further integrates the user’s feedbacks and BBPRF, and the retrieval performance is even better than using BBPRF alone. | en_US |
| DC.subject | Pseudo Relevance Feedback | zh_TW |
| DC.subject | block | zh_TW |
| DC.subject | 影像檢索 | zh_TW |
| DC.title | 一個基於區塊式虛擬相關回饋演算法： 以影像檢索為例 | zh_TW |
| dc.language.iso | zh-TW | zh-TW |
| DC.title | A Block-based Pseudo Relevance Feedback Algorithm for Image Retrieval | en_US |
| DC.type | 博碩士論文 | zh_TW |
| DC.type | thesis | en_US |
| DC.publisher | National Central University | en_US |