Please use this identifier to cite or link to this item:
標題: 基於直方圖統計方式的類神經模糊網路之方法做彩色影像分割
Color Image Segmentation By Histogram-Based Neural Fuzzy Network
作者: 彭懷生
Peng, Huai-Sheng
關鍵字: Match score;相似計分;PCT;FCM;BPNN;SONFIN;主值成分轉換;模糊C-平均聚類;直方圖統計;倒傳遞類神經模糊網路;自我架構類神經模糊推論網路
出版社: 電機工程學系
摘 要
本論文提出基於直方圖統計方式的自我架構類神經模糊推論網路(SONFIN)之方法來做彩色影像分割。我們使用色調(H)和濃度(S)二維的彩色空間來表示每一彩色像素,因為它具有排除光線影響色彩的能力,而且只需用二維空間提高了運算速度。為增加以直方圖統計表示色彩的精確度,我們分別使用四種不同的HS空間切割方式。訓練SONFIN所用的直方圖統計資料乃是由不同環境所拍的影像來獲得,以便增加方法的強健性。為測試所提方法的好壞,我們將其應用在以膚色為基礎的人手切割上。我們並比較了其他切割方法,其中包含直方圖統計之相似計分(match score)法、主值成分轉換(PCT)、模糊C-平均聚類(FCM)和基於直方圖統計方式的倒傳遞(BP)類神經模糊網路四種方法。在這些分割的方法裡,我們發現使用SONFIN方法是最好的而且可得到不錯的分割結果。此外,我們所提出的不同的HS空間切割方式也對於實驗結果有不錯的改善。

Color image segmentation by histogram-based self-constructing neural fuzzy inference network (SONFIN) is proposed in this thesis. Each color pixel is represented by a HS space for brightness or lightness component is irrelevant to chromatic information of images in the space. In addition, the usage of only two-dimensional space increases computation speed. To represent a color by histogram as accurately as possible, four types of HS space division approaches are studied. Histogram information from images under different environments is used to train SONFIN to make the method as robust as possible. To verify performance of the proposed method, experiments on human-hand segmentation based on skin color are performed. For comparison, other four segmentation methods, including color image segmentation by match score of histogram, principal component transformation (PCT), fuzzy c-means (FCM), and histogram-based back propagation (BP) neural network, are applied to the same problem. From comparisons, we find that SONFIN achieves the best performance, and the segmentation result is good. In addition, the HS division approaches proposed in the thesis also help to improve the segmentation performance.
Appears in Collections:電機工程學系所

Show full item record

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.