根據主成分分析及邊緣資訊的先進顯著區域擷取技術

Abstract

在這十年內，因應大量影像處理對於預先定義出影像中顯著區域的需求，像是影像切割、影像壓縮和影響大小調整等等。定義顯著區域為自動判斷出影像中較重要區域，並且產生出以顯著程度為權重的影像， 這個研究已經成為影像處理中不可或缺的前處理方法。 在這過去十年內，顯著區域偵測已經發展出主要的四個方向，像是基於區域法、基於方塊法、中心與周圍差異法和物體分類法。而不同於中心與周圍差異法，基於區域法在這幾年發展出較為可靠的演算法，這是因為此法可以利用不同的影像切割方式在不同的影像中互相擬補各個方法的缺點。 根據這個理論，我們將精力聚焦於基於區域法和基於方塊法來做改良的基石。基於區域法是採用不同的切割方式做為定義顯著區域的前處理，此法可以結合不同影像切割的優點來達到較好的結果。再另一方面，基於方塊法計算出每一個八乘八方塊的差值，並且使較少出現的方塊給予較高的顯著權重，即為越少在影像中出線的方塊即為顯著區域。在本篇論文當中，我們提出了兩個新定義顯著區域方法。第一，區域對筆法為現在領域中最佳的方法，我們結合此法並且結合我們提出的背景定義概念的不同基於區域法而成為一個完整的系統。第二，我們提出一個新方法為結合基於區域法以及基於方塊法，此法巧妙的運用兩個方法的優點達到比現在最好的方法更好的結果。 依據實驗結果顯示出我們提出的方法有效的並且成功的超越現在顯著區域研究領域中大部分的方法，此項結果可以支持我們的研究結果是卓越的並且有理論根據的。

Research of saliency map detection has fast development in the last ten years because of the great demand of image processing application, for instance image segmentation, image compression and image resizing etc. Saliency map, defining salient region of image automatically, has been indispensible part of preprocessing method in several image processes. In past decade, saliency map has developed four primary concepts in the field such as the following: region-based, block-based, center-surround and object classification methods. Region-based method shows the robust performance during the recent year which is contrary to center-surround method. This is due to different image segmentation method, it can complement each methods. According to the theorem, we focus on the region-based and block-based method for saliency detection. Region-based method is adopted image segmentation for preprocessing determining the salient regions, which can combine different segmentation methods with several concepts for enhancing diverse image. On the other hand, block-based method which calculates difference of each 8x8 block in image determining seldom appearing patches in high score to be a salient patch. In this thesis, we provide two novel methods below: First, a mixed region-based method which improves the state-of-the-art RC method into a region-based system with background determination concept. Second, the combination of region-based and block-based methods uses both advantages of two leading to a novel region-based method in our system. The simulation results show that our two novel methods have better performance compared to different methods in saliency detection field.