Abstract
影像中的對焦物通常是取像者所感興趣且要分析的東西。影像分析常首先 取出對焦物。在此篇論文中探討當背景失焦時，對焦物的選取方法。此類 背景失焦的照片很容易從競賽、人像、近攝、或是生態攝影中取得。它的 用處可在影像編輯、合成、分析、分類、了解、壓縮、....，等等。此篇 論文第二章先探討相機成像的光學幾何模型、光學散射模型、和成像的數 學模型。第三、四、五章分別以灰階區分法、梯度法、及線性預測法這三 種方法來做前置處理，接著再加上邊緣點連結、中位濾波器、或去除不必 處理區及亮度對比調整的中間處理，最後一步驟是背景連通區搜尋法及最 大面積法，如此可以成功的將對焦物取出。第六章則針對背景有高亮度區 域而設計出斜面濾波器來消去背景，留下對焦物。然後加上中置及後置處 理法，可成功的將對焦物取出。此方法最主要目的是在背景有高亮度區時 ，若使用梯度法或線性預測法，並不能處理的很好而設計出來的方法。前 面方法皆利用灰階影像作對焦物的選取。第七章將討論彩色影像作選取對 焦物時可利用的方法。其方法最主要考慮彩色影像可以轉換色彩空間而使 得對焦物容易被選取出來。第八章為結論及討論未來可以繼續做的改進及 研究。 Focused objects are often the objects of interest in an image. The way to analyze an image is often to extract the focused objects at first. The thesis discusses the extraction of the focused objects from an image with defocused background. This kind of image can be easily found in race, person in a scene, closed objects, or ecological photograph. The extracted object can be used in image editing, synthesis, analysis, classification, understanding, compression, ..., etc. Chapter 2 introduces defocus models from optical models and image formation process. In Chapter 3,4,5,6,7, existing methods and new approaches are investigated. Different images will be processed with slightly different steps in order to fit their properties. Five different first-stage methods using gray- level thresholding, gradient-based method, linear prediction method, ramp filter method, and color image processing method have been implemented to detect the boundary of the focused objects. The mid-stage processing methods using median filtering, edge linking, removal of local regions without large gradients, prediction errors, or ramp errors, and local contrast adjustment have been implemented to enclose the focused objects. The final-stage methods using connected background search method and finding the largest non-background region method have been implemented to extract the focused objects. The advantages and limitations of these methods will also be explored. Simulation examples are provided to justify their performance. In Chapter 8, conclusions and directions for future research will be given.