Fusion of Tensor Voting and Hough Transform for Image Analysis and Its Application to Video Stablization

瀏覽統計 Email 通知 RSS Feed

簡歷

基本資料

Project title

Code/計畫編號

NSC99-2221-E019-037

Translated Name/計畫中文名

基於張量投票及霍福轉換之影像分析技術及其在視訊穩定系統之應用

Project Coordinator/計畫主持人

Shyi-Chyi Cheng

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Computer Science and Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=2118600

Year

2010

Start date/計畫起

01-08-2010

Expected Completion/計畫迄

01-07-2011

Bugetid/研究經費

552千元

ResearchField/研究領域

資訊科學--軟體

"快速發展的網際網路引進包括影像、視訊資料在內之大量多媒體資訊在網路上傳送，前瞻的影像分析技術有助於建立符合語意、適切的物件結構，然而現存自動的影像處理技術往往無法精確地剖析語意物件的結構，尤其是當影像背景雜亂時，這個問題顯得更加明顯。高階的視覺系統與低階的影像分析技術的差異限制了網際網路多媒體系統提供有效的影像、視訊資訊共享服務，因此研發有助於縮短高階語意與低階特徵的影像、視訊分析技術是一項具有挑戰性且從根做起的研究課題。本計畫提出一基於投票機制的視覺樣本群組技術整合區域性的低階特徵為廣域性的高階語意特徵。本計畫所提出的視覺樣本群組技術是由感官群組技術演化而來，利用感官群組技術我們可偵測在具有雜訊或資訊不完全的情況下，有效地從影像中偵測出有意義的結構或樣本，這些結構的準確性已被證明會影響高階功能如物件偵測、辨識、檢索的可靠性。事實上，大多數藉由影像處理方法得到的低階、區域性特徵如邊緣點、線段、或影像區域都隱含不確定性，因此借用感官群組技術整合這些帶雜訊的區域特徵得到的整體性高階特徵也因而不太可靠，為了解決這個問題，我們使用舉量保持技術取得一個影像小區塊的線邊資訊，結合張量投票的概念，區塊的線邊資訊提供穩定的區塊與區塊結合方向，進而得到抗雜訊的高階特徵擷取方法。為了驗證本計畫所提出的方法的準確性，我們應用所提出的視覺樣本群組技術於偵測視訊內容多重物件、多重運動的問題，並以數位穩定應用當例子，設計了一個可抗雜訊的視訊穩定演算法，初步實驗結果驗證本計畫提出的方法的可行性。" "As the rapid advance of the Internet, mass multimedia information including image or video data is transmitted through the network. The development of advanced image analysis facilitates the segmentation of semantic objects or patterns from images. Unfortunately, most of the image processing operators either fail in detecting accurate structures of semantic object from cluttered background or suffer from high computational complexity. The gap between high-level visual system and low-level image analysis limits the functionality of multimedia sharing in the Internet. It is a challenge to shorten this gap through the development of advanced image analyses. In this project, we propose a perceptual grouping scheme to integrate local features into global features describing high-level image objects or patterns using tensor voting. In this project, we describe a visual pattern grouping scheme which is basically a variant of perceptual grouping proposed in the literature. Perceptual grouping is a useful tool to detect organized structures or patterns even when the information is in-complete. Accurate extraction of meaningful objects or patterns provides reliable preprocessing information to higher level functions, such as object detection, recognition, and retrieval. In fact, most of the low-level functions, such as edges, curves, and surfaces detection generate noisy output which leads to un-reliable global object description using perceptual grouping. To tackle the problem, we divide an image into multiple non-overlapping blocks whose edge equations are derived by the moment-preserving edge detector proposed in our past work. These blocks, based on the detected edge parameters and tensor voting, are then grouped to form robust higher-level objects. The proposed visual pattern grouping algorithm is also verified by applying it to detect multiple motions of multiple objects in video sequence. A video stabilization system is finally constructed to demonstrate the performance of the method. Preliminary experimental results show the feasibility of the proposed method."

Keyword(s)

電腦視覺
影像分析
視覺樣本群組
張量投票
霍福轉換
矩量保持技術
Computer vision
image analysis
perceptual grouping
tensor voting
Hough transform
moment-preserving technique

DSpace CRIS

Fusion of Tensor Voting and Hough Transform for Image Analysis and Its Application to Video Stablization

基本資料

Description