以Kernel為基礎之模糊分群演算法硬體架構實現

Abstract

本論文根據文獻[12]以及文獻[17]，以此兩則文獻中提到的FCM-SC分群演算法的硬體架構和KFCM演算法的硬體架構為基礎，實作以非線性高斯核函式為核距離計算之KFCM[12] 再加上空間資訊[17] 後的分群演算法硬體電路，具有管線化以及可以同時計算所有分群之權重係數的能力。此架構改良了以往KFCM分群演算法對於有雜訊的資料做分群的問題，並且配合KFCM本身可以對非線性資料分群效果較好的能力，所以能夠廣泛地使用在許多的分群資料上，並且都有良好的辨識率。本論文使用FPGA實現我們提出的硬體架構，並使用人工雜訊圖片作為實驗測試資料。實驗結果顯示本架構對於有雜訊的非線性資料分群效果確實較KFCM佳，且架構簡單提供了日後高度的延伸性。

Based on the FCM-SC (Fuzzy C-Mean with spatial constraint) architecture in reference [12] and the KFCM (Kernel-Based Fuzzy C-Means) architecture in reference [17], KFCM-SC (Kernel-Based Fuzzy C-Means with spatial constraint) hardware architecture is proposed here with non-linear Gaussian kernel function and spatial constraint. Moreover, the KFCM-SC architecture also takes the advantage of the pipeline and it can compute all of the membership coefficients and centers concurrently. Compared to KFCM architecture, KFCM-SC architecture improves the segmentation ability for noisy data by computing the spatial information. With these advantages, it can deal with the non-linear data due to the kernel function, KFCM-SC architecture can be applied to wide of data and it can achieve better segmentation results. KFCM-SC architecture is implemented on FPGA and tested with noisy picture data. The segmentation result shows that KFCM-SC architecture definitely has a better ability with non-linear noisy data compared to KFCM. Because of the simple architecture of the KFCM-SC, it can be extended easily.