靜態影像中樣式的發現與計數

Abstract

物件的計數一直是生活中很重要且常見的問題，現行也已有許多對於靜態影像中物件計數的相關研究，然而大部分的方法都是基於已知目標物的部分特徵，設定了相當程度的限制條件而進行偵測，適用的資料範圍及彈性不甚理想。本研究以重複及明顯為樣式的概念，發展一般化物件偵測與計數系統，不預先設定目標型態，而是以重複出現的明顯視覺性特徵，找出影像中的樣式，進行特徵描述並計算數量，進而達到物件計數的目的。本研究對於樣式( Pattern )的定義，即同時在形狀、顏色、尺寸三方面有相同特徵且與影像中相對明顯的封閉邊緣輪廓。本研究對靜態影像中的封閉邊緣進行特徵擷取並分群，保留相對明顯的，即為樣式。主要的內容議題分為兩個部分，第一部分為輪廓的分群，以邊緣資訊找出影像中的輪廓，再根據形狀、顏色以及尺寸將相似的輪廓分在同一群。第二部分則是篩選出相對明顯的輪廓分群，進行標示、外型特徵的描述以及數量計算。最後的結果即為已標示出樣式輪廓的影像及相應樣式的特徵描述。本研究以兩個不同面向的實驗進行成果探討，其一為系統找出的物件是否能符合人的期望，另一項則是不同物件的分群成效。實驗結果證明，基於本研究提出的特徵定義而實作的樣式偵測系統，確實能有效的找出達成靜態影像中的一般化物件偵測與計數。

Object counting is common and important in life. Currently there are many research about counting objects in an image, but most of them need to adjust parameters base on target profiles. It leads to some limits on data type which make these approaches inflexible. In this paper, we take repeat and obvious contours as pattern, describing their profiles and counting numbers without setting data conditions in advance. In this paper, the definition of pattern is the contours with similar features on shape, color and size simultaneously.We extracts contour features from image, divides contours into many groups and keeps obvious ones as patterns. The proposed approach include two parts. The first part is contour clustering. We get contours from edge information in image, then divide them into many groups referring to similar shape, color, and size. In second part, we make a filter to keep those comparatively obvious groups, then mark contours, describe their profiles and count. Final results are images with marked contours and pattern feature descriptions.This paper evaluates the performance from two aspect. One is the ability to find objects that matches human thoughts. The other one is the effect of clustering. The results show that the proposed approach can find most of object profiles and count numbers in general images.