使用膚色與陰影機率高斯混合模型之移動膚色區域偵測; Gaussian Mixture Models with Skin and Shadow Probabilities for Moving Skin Region Detection

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/48450

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/48450

Title:	使用膚色與陰影機率高斯混合模型之移動膚色區域偵測;Gaussian Mixture Models with Skin and Shadow Probabilities for Moving Skin Region Detection
Authors:	吳宗憲;Tsung-Hsien Wu
Contributors:	資訊工程研究所
Keywords:	膚色偵測;陰影偵測;高斯混合模型;detection;shadow;skin detection;GMSM;GMM
Date:	2011-07-26
Issue Date:	2012-01-05 14:55:07 (UTC+8)
Abstract:	膚色偵測在影像應用中扮演一個非常重要的角色，它的應用相當廣。單純的膚色偵測會先設定閥值，當色彩區域落在閥值之內才會被判斷為膚色，但若背景有著接近膚色的物體或牆壁則會造成誤判，此外，光線昏暗或變化強烈的地方，如果單純根據閥值判斷，膚色很容易受到光線影響造成判斷錯誤；同樣地，陰影偵測也有類似的情形，因此我們利用膚色及陰影特徵分別計算出膚色與陰影機率，然後結合高斯混合模型的學習機制，來解決此問題。本篇論文提出一種利用膚色與陰影機率結合高斯混合模型來偵測移動膚色區域的方法，稱之為膚色與陰影機率高斯混合模型。本系統分為兩大部分，首先利用膚色機率結合高斯混合模型偵測移動的膚色區域，接著再使用陰影機率高斯混合模型去除手或其他移動物體在背景上造成的陰影，藉以得到所要的移動膚色區域。其中的陰影機率高斯混合模型是以高斯陰影混合模型為基礎改良而來，由於高斯陰影混合模型需先判斷像素是否為陰影來決定要不要建入模型中，但由於陰影範圍難以界定，因此我們提出將單純的陰影判斷改為計算出陰影機率，然後結合高斯混合模型來建立陰影模型，並利用此模型來偵測陰影區域。我們在實驗中分別測詴膚色偵測及陰影偵測結果，而經實驗證明膚色高斯混合模型可有效偵測移動膚色區域，並且不會偵測到接近於膚色的背景及移動的非膚色物體，而高斯陰影機率混合模型雖不能提升陰影偵測率，但可降低陰影偵測的誤判率。 Skin detection plays an important role in a wide range of image processing applications. Common skin detection methods need to set skin color cluster decision boundaries in different color space components. When the pixel values fall within these decision boundaries, they would be defined as skin pixels. However, this method may not work well in the scenes with complex and time-varying illumination. Common property-based shadow detection methods have the same problem. To overcome these shortcomings, a composite improved approach to detect moving skin regions is presented in this thesis. This thesis proposes Gaussian mixture models with skin and shadow probabilities (GMM-SS) to detect moving skin regions. The system is separated into two parts. The first part of GMM-SS uses Gaussian mixture models with skin probability to detect moving skin regions. The learning rate in GMM-SS is lower for pixels with higher skin robabilities. The moving skin regions would be constructed into background Gaussian more slowly by the lower learning rate, and vice versa. The second part of GMM-SS uses Gaussian mixture models with shadow probability to remove casting shadows to get pure moving skin region. The shadow Model is used to identify distributions of pixel values that could represent shadowed surfaces. In this model, it modifies shadow distribution learning rate with each pixel’s shadow probability and makes it more flexible for shadow detection. The experiment results show that our system is more efficient and robust for moving skin region detection and removing shadow regions. We compare it with common skin detection and Gaussian mixture models. It keeps steady skin detection rates and low false alarm rates in most situations than common skin and shadow detection methods.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	622	View/Open

社群 sharing

Loading...