本論文提出一個新的白血球分類系統，此系統包含兩個階段，第一階段執行白血球影像的切割；第二階段則執行白血球辨識工作。為了有效切割白血球，我們提出了一種新的切割演算法，首先利用主成份分析 (principal component analysis) 在 HSI 色彩空間內定義出一個橢圓形區間，在橢圓形區間內之像素點會被視為白血球細胞核和細胞質之顆粒，再經過形態處理(morphological process) 後，才完成血球切割工作。接著，從切割後之白血球中萃取出色彩、形狀和區域方向圖樣 (local directional pattern) 等特徵，然後，將這些特徵輸入類神經網路予以辨識出是何種白血球。本論文採用三種不同之類神經網路：多層感知機 (multi-layer perceptron)、支持向量機 (support vector machine) 和模糊化多維矩形複合式神經網路 (fuzzy hyper-rectangular composite neutral networks) 來辨識白血球種類。 為了測試本論文所提之白血球分類系統之有效性，共使用了450張白血球影像，其影像大小為360×360的 JPEG 格式。整體來說，每張影像從一開始的切割到辨識的處理時間約為1秒。在整體資料的正確率比較上，多層感知機可達到最佳之效果，其正確率可達99%的正確性；支持向量機則有97%的整體正確率，雖然模糊化多維矩形複合式神經網路對於訓練資料有著高達百分百的辨識結果，但其測試集之辨識結果不佳。 This thesis presents a new white blood cell classification system. The system involves in two steps. While the first step is the segmentation of a white blood cell from an image, the second step focuses on the recognition of the types of white blood cells. We propose a new segmentation algorithm for the segmentation of white blood cells. First of all, we use the principal component analysis (PCA) to define an elliptical region in HSI color space. Pixels with color in the elliptical region will be regarded as the nucleus and granule of cytoplasm of white blood cell. Through a morphological process, we can segment the white blood cell from the image. Then, several features (e.g., color, shape, local directional pattern)are extracted from the cell. These features are fed into a neural network to recognize the types of the white blood cells. In this thesis, three different neural networks (i.e., multi-layer perceptron (MLP), support vector machines (SVM) and hyper-rectangular composite neutral networks (HRCNN)) are used to recognize white blood cell type. To test the effectiveness of the proposed white blood cell classification system, a total of 450 white blood cells images were used. The image size is 360 × 360 with the JPEG format. Overall, the processing time of each image from the start of segmentation to the end of recognition is about one second. The MLP could achieve the best results with the recognition rate 99% of accuracy. As for the SVM, it could achieve 97% of accuracy in overall data. Although the HRCNN could achieve 100% correct ratio for the training data set, the accuracy for the testing data set was not as high as the other two types of neural networks.
