應用支持向量機器提升模板匹配法於圖像檢索的準確性

Abstract

由於資訊科技的快速發展，大量的圖像資料以數位化的方式儲存，如何以一張查詢圖片即可從中檢索出與其同類別的圖像是我們關注的目標。模板匹配法是一種計算圖像相似度的方法，將圖像資料庫的圖像對於查詢圖片計算互相之間的相似度，以檢索出相似度較高的圖像。由於在計算相似度時並不需用知道圖像的類別，因此可以節省對於圖像資料庫標注類別的人力成本。然而我們發現相同類別的圖像之間仍可能有些許差異，被稱為組內變異，當兩張圖像的組內變異過大，會導致計算出的相似度較低，使得模板匹配法無法檢索出我們所想要的圖片，因此，本論文提出在模板匹配法的基礎上輔以支持向量機器，以求降低組內變異的影響，並結合多個支持向量機器的檢索結果，使得圖像檢索的準確性會比使用單一支持向量機器來得良好。本論文選用Caltech-101的圖像資料庫，比較模板匹配法在使用支持向量機器前後的檢索結果，根據實驗結果證明，支持向量機器的確可以提升模板匹配法檢索結果的準確性。

In this thesis, we propose a content-based image retrieval method to enhance the accuracy of acquiring images from large image databases. The term “content” refers to the information that can be extracted from the image itself. The proposed method combines template matching and support vector machine. With the content of images, the similarities among them can be measured according the chi-square distance. The advantage of template matching is that prior knowledge of the databases is unnecessary. However, the variance of images for the same object, called within-class variance, could affect the similarities among images and lead to wrong query results. To solve this problem, support vector machine is introduced in the second stage. Support vector machine can classify images into two categories. We select the training images of support vector machine based on the similarities computed in the first stage. Support vector machine is able to generate non-linear decision boundaries that can reduce the effect of within-class variance. In this thesis, we utilized Caltech-101 databases for experiments. The experimental results show that the proposed content-based image retrieval method has better accuracy than template matching.