本文提出一個新的邊緣偵測演算法。啟發自CFA 插補(interpolation)法,本文提出可以在拜耳圖樣(Bayer pattern)上執行類高斯平滑(Gaussian smoothing)及類拉氏(Laplacian)邊緣偵測之演算法。經由適當延伸,本演算法亦可偵測彩色及灰階影像之邊緣。使用本演算法之優點包括插補運算及色彩轉換運算之節省,並能節省記憶體使用量。在邊緣偵測演算法部分,相較於5×5 高斯-拉氏遮罩(Laplace of Gaussian mask),本文提出之方法在彩色邊緣偵測上可節省約5/6 之運算量,相較於零點偵測法(zero-crossing detection) 可節省2/3 之運算量,於灰階邊緣偵測亦可節省約1/3 之高斯-拉氏遮罩運算。實驗顯示本文提出之演算法有不錯的偵測結果,並可藉由調整標準差及門檻值兩參數亦可增加此演算法之適應性。
A new edge detection algorithm is proposed in this Thesis. Inspired by the Color Filter Array (CFA) interpolation kernels, we design two other kernels for the algorithm to perform Gaussian-like smoothing and Laplacian-like edge detection directly on a Bayer-patterned image. Also, the proposed algorithm can be easily extended to existing color and grayscale images. That is, it is capable of detecting edges in a Bayer-patterned, a color, or a grayscale image. Benefits of performing edge detection on a Bayer-patterned image include the computation saving of the interpolation and/or color space transform to a full color or grayscale image, and lower memory usage. With the proposed 5×5 kernels, the extension to color edge detection theoretically presents approximately 5/6 of computation saving from the existing color Laplace of Gaussian (LOG) operations, and 2/3 saving from the three-channel zero-crossing detection, while for grayscale edge detection presents approximately 1/3 of computation saving from the existing grayscale LOG operation. Experimental results show that the proposed algorithm has great localization and flexibility by tuning its standard deviation σ and threshold parameter th.