Title: 以形態學運算及模糊運算於影像特徵抽取之研究
A Study on Feature Extraction by Morphological Operations and Fuzzy Reasoning
Authors: 林瑞盛
Rey-Sern Lin
Yuang-Cheh Hsueh
Keywords: 數學形態學;模糊推理;雜訊去除;角點偵測;邊緣偵測;多維訊號排序;特徵抽取;梯度權重濾波器;Mathematic Morphology;Fuzzy Reasoning;Noise Removal;Corner Detect;Edge Detectionion;Multivariate Data Ordering;Feature Extraction;Gradient Weighted Filters
Issue Date: 1999
Abstract: 近來,為了提升影像處理系統的效能,雜訊過濾、影像切割、角點偵測、邊界偵測這些影像處理的問題, 受到了相當多的注目。 形態學運算是跟據物體的形狀變化理論定義而來的,因此在影像處理及分析上,已經成功的發展於濾波器的設計、邊的偵測、角點的偵測、物體的骨架化等等應用。由於它具有簡單的整數運算,可平行處理,可讓影像處理上的及時問題可解的特性,因此十分吸引人。另一方面,模糊系統可以解決不精確和不確定性的輸入及輸出資訊, 因此在影像處理上,逐漸成為十分受歡迎且有用的工具。 特徵抽取在影像分析及辨識上是一個基本且重要的步驟。特徵可用統計的方式表現或特徵化,也可以直接以物體幾合上的形狀如線、邊、角點呈現物體外形的線索,這種幾合特徵對低階的影像處理特別重要。本論文中,主要集中於抽取幾合特徵,我們將特徵抽取分成兩個步驟:前處理的雜訊去除和實質上的特徵抽取,在前處理步驟中,空間上的雜訊過濾和盡可能保留邊是主要的目的, 而實際的特徵抽取則著重於改善效能和簡化處理過程。 在這論文中,首先,我們提出一個一般化的可調適性的梯度權重濾波器,它可以用三個參數: 區域參數、區域權重函數、和全域權重函數特徵表示,當中我們分別對這三個參數作改善及實驗探討,實驗顯示這三個參數對可調適性的梯度權重濾波器的影響。更進一步,我們將它擴充至彩色影像雜訊的去除。因為形態運算天性上繼承了幾合特徵,因此特別適用於特徵抽取,一個修改式的角點偵測法於是被提出,這方法的優點是只須簡單的整數運算及較準確的位置性。數位影像從自然景像取像時,在灰階值上和物體邊界總是存在一些模糊性,因此我們結合模糊推理和形態學運算的特性,將其應用於影像邊界的偵測。實驗結果顯示較好的正確位置、較細的邊、且使用的參數也較CANNY Operator少。
Recently, considerable attentions have paid to improve many image problems such as noise removal, filtering, segmentation, corner detection, and edge detection. Morphological operators are defined in terms of geometric shapes and patterns in an image and have been developed for many aspects of image processing and analysis, such as noise suppression, filtering, edge detection, and skeletonization. What makes morphological operators so attractive is the fact that they involve simple logical operations and can be implemented in parallel, making real-time applications possible. On the other hand, fuzzy systems can process imprecise or uncertain input and output information defined by fuzzy sets and expressed by linguistic terms. They have developed gradually as useful and well-known tools in image processing system. In image analysis and pattern recognition, feature extraction is the most fundamental and important processing phase. Features in images can be characterized by statistic extraction and comprised by itself geometric shape, color and texture. Many geometrical features such as lines, edges, and corners represent some cues about object’s profile. They are especially important for low-level image analysis. In this dissertation, we focus on the extraction of geometrical features and model it by pre-processing phase and physical processing phase. The pre-processing, spatial filtering, will aim to eliminate the interference factors such as contamination and preserves sharpness as many as possible. The physical processing, features extracting, will aim to enhance performance, simplification, and modification by morphological operations and fuzzy reasoning. In this dissertation, an adaptive gradient weighted filters characterized by three factors, called the local constants, the local weighted functions, and the global functions are proposed in the pre-processing phase. Moreover, the multichannel filters by gradient information are extended. Since morphological operations are inherently geometric in nature, they are ideally suited for the geometrical feature extraction. A modified morphological corner detector with simple integer computation is developed firstly in physical processing phase. Digital images, being mappings of natural scenes, are always accompanied by a reasonable amount of fuzziness due to imprecision of gray values of the pixels and vagueness in the region boundaries. Then, an edge detector by multi-scale fuzzy reasoning and morphological operations is introduced. The experimental results disclose that our approaches obtain thinner edge magnitude, more correct localization, and need less input parameters than Canny operator.
Appears in Collections:Thesis