Please use this identifier to cite or link to this item: `http://hdl.handle.net/11455/7598`
 標題: 基於直方圖統計與輪廓模糊類神經網路 做花朵的分割與辨識Flower Image Recognition by Histogram-Based and Contour-Based Neural Fuzzy Network 作者: 黃怡碩Huang, Edward 關鍵字: Array精確度 直方圖 飽和度 建構類 分類器;SONFIN 出版社: 電機工程學系 摘要: 摘 要本論文首先提出基於直方圖統計的自我建構類神經模糊推理網路(SONFIN)之法執行花朵的切割。之後，利用切割的花朵之輪廓與顏色兩種特徵值及SONFIN辨識器執行花朵的辨識。這邊以色調（Hue）與飽和度(Saturation-)二維彩色空間來表示每一像素的特徵。除了有簡化計算的功能，還有就是我們希望能夠排除亮度對於像素的影響。為增加以直方圖統計表示色彩的精確度，我們使用非均勻的HS空間切割方式。所得到的直方圖再經過SONFIN分類器來完成花朵的切割。透過與混合高斯分類器(Gaussian Mixture Model Classifier)及直方圖分類器(Histogram-based Classifier)可以比較出基於直方圖的SONFIN分類器是較為優異的切割法。經由SONFIN切割出來的花朵，再以灰階化、膨脹與蝕刻的運算來得到更完整的結果。利用邊界抽取與快速離散傅立葉轉換(FFT)，可得到輪廓特徵值。顏色特徵值則是取切刻花朵得HS平均值。所得到的輪廓與顏色特徵值則用來作SONFIN辨識器的輸入。經由十種花朵的辨識可驗證所提方法具有不錯的辨識率。AbstractThis thesis proposes flower segmentation by histogram-based Self-cOnstructing Neural Fuzzy Inference Network (SONFIN). Then, recognition of the segmented followers by SONFIN using both contour and color features is proposed. Color information of each pixel is represented by two-dimensional colored space of H (Hue) and S (Saturation). Besides the simplification benefits, it also enables us to rule out effects on the elements from brightness. To represent a color by histogram as accurately as possible, non-uniform partition of HS space is used. The histograms are fed as inputs to SONFIN classifier for flower segmentation. It is shown from comparisons with Histogram-based Classifier and Gaussian Mixture Model Classifier that SONFIN is the comparatively optimized method for segmentation. Segmented flowers by SONFIN will undertake gray-scale, erosion and dilation operations to achieve a better result. Contour feature is extracted via boundary extraction and Fast Fourier Transform (FFT). As to color feature, it is extracted by taking average HS values of the segmented flowers. Both the extracted contour and color features are fed as inputs to SONFIN recognizer. Experimentations on recognition of ten kinds of flowers show that a good performance is achieved by the proposed method. URI: http://hdl.handle.net/11455/7598 Appears in Collections: 電機工程學系所