Measuring Shape Similarity in Cluttered Images

• Other Authors: Sinn, Lok Tsun (author.)
• Format: Others
• Language: English; Chinese
• Published: 2016
• Online Access: http://repository.lib.cuhk.edu.hk/en/item/cuhk-1292279

在本論文中，我們提出一個用於雜亂圖片，並能容忍尺度差異及局部摭擋的形狀相似度計算。由於在雜亂圖片中含有大量閉合及打開的形狀，雜亂圖片中的形狀相似度計算目標為找出圖片的部分以使當中的形狀與查詢形狀為最相似。我們提出使用我們先前提出的弧長金字塔描述符(PAD)，並配合變換聚類分析以進行在雜亂圖片中的形狀相似度計算。PAD是一個基於形狀上某點的形狀描述符，並能以局部尺度不變地描述該點附近的形狀。PAD可以描述閉合及開放形狀上的局部形狀。比較PAD可以得知某點在尺度不變性下的局部形狀相似度。我們可以使用PAD以對雜亂圖片中的形狀及查詢形狀進行描述，並比較雜亂圖片及查詢形狀的任意兩點，即一匹配對的局部形狀相似度。然而，PAD只能對局部區域的形狀進行比 === 較。為了考慮整體形狀，我們提出一個對於變換的聚類分析。每一個匹配對的PAD比較均與一個能使該匹配對附近的局部形狀相似度最大化的變換關聯。我們對於這些變換進行聚類分析，以找出一個由多對匹配對並有的變換。而此變換亦代表能使更多地方的局部形狀相似度最大化，以達到更多部分的形狀相似。相較於先前提出，基於兩形狀的接近度去決定是否部分形狀是否匹配的PAD形狀匹配，我們確保了匹配部分形狀的形態亦相似。這能避免當形狀包含多條邊緣，或者在處理雜亂圖片中的「海量形狀」現像時出現的假陽性。我們進行了五個實驗，以評估我們提出的形狀相似度計算對於局部摭擋、邊緣衰變及雜亂形狀下的容忍度。當中亦包括與先前提出的PAD形狀匹配及其他現有的形狀匹配的比較。我們亦討論了我們提出的形狀相似度計算在其他電腦視覺中的應用。 === In this thesis, we propose a shape similarity measurement for cluttered image with tolerance to scale changes and partial occlusion. While there are numerous closed and open shapes lying in the cluttered image, the shape similarity measurement aims at finding a part of the image that is most similar to a given query shape. Our shape similarity measurement uses a transformation clustering approach, together with our previously proposed pyramid of arclength descriptor (PAD) in achieving this goal. PAD is a point-based shape descriptor which describes local shape near a point in a locally scale-invariant manner, and it is capable in describing local shape on both open curves and closed shape. Comparing PAD can tell the similarity of local shapes scale-invariantly. With the aid of PAD, we can well describe the shapes in the cluttered image and the query shape, as well as measuring the local shape similarity of any two points, or a matching pair, between cluttered image and the query shape. However, PAD can only determine the shape similarity within a limited range and in a local sense. To measure the shape similarity in a global sense, we further propose clustering on transformations. For every PAD comparison of a matching pair, a transformation is associated in maximizing the local shape similarity for the matching pair. By performing clustering on these transformations, we can find out the transformation shared among multiple matching pairs. This transformation can maximize the local shape similarity at multiple positions, leading to longer shape portion sharing similar shape. In contrast to previous PAD matching, in which two shape portions are regarded as matched when they are close together after transformation, our transformation clustering ensures the matched portions are having similar shape. This can avoid false positive when dealing with shapes with multiple boundaries, as well as the sea of shapes" phenomenon in the edge map of cluttered image. Five experiments are conducted in evaluating the capability of our shape similarity measurement for cluttered images under occlusion, boundary decay and cluttering shapes. Comparison with our previous PAD matching approach as well as other existing shape matching methods are also included. We further discuss the potential applications of our shape similarity measurement in other computer vision problem." === Sinn, Lok Tsun. === Thesis M.Phil. Chinese University of Hong Kong 2016. === Includes bibliographical references (leaves ). === Abstracts also in Chinese. === Title from PDF title page (viewed on …). === Detailed summary in vernacular field only. === Detailed summary in vernacular field only.