Edge Detection by Scale Multiplication Vahid Rezaei Professor William Hoff Course Project for Computer Vision Department of Electrical Engineering and Computer Science Spring 2012 Outline • Review of Edge Detection methods • Scale Multiplication approach 1D signals The discrimination of singularity and noise. The scale multiplication. The thresholding. 2D images • Simulations 2 Outline • Review of Edge Detection methods • Scale Multiplication approach 1D signals The discrimination of singularity and noise. The scale multiplication. The thresholding. 2D images • Simulations 3 Review – Why Edge Detection? Edges are used to show considerable change in some physical aspect of the image such as intensity changes in a neighborhood. Applications are in object recognition, image registration, image segmentation, data compression, and image reconstruction . 4 Review – Classical Approaches For example: The Laplacian operators compute some quantity related to the Laplacian of the underlying image gray tone intensity surface. The zero-crossing operators determine whether there is a zero-crossing within the pixel or not. A threshold value is used for gradient to find the edges. 5 Review – Classical Approaches The major drawbacks of such operators are: A. It is difficult to find the actual location of the edge. B. The choice of threshold value (and other parameters) is based on trial and error which may leads in meaningless edges. C. dealing with noisy images is another challenge for classical approaches. 6 Review – Canny Edge Detector Canny’s work on edge detection is based on three criteria:), and good n(x)~N(0, G(x) isdetection, a step edge withgood of A. localization, magnitude low spurious response. (Optimal edge detector) Consider the 1D-signal W(x)=G(x)+n(x) and assume is a FIR filter on [-T,T] (edgedetector filter). The response of f toW : . 7 Review – Canny Edge Detector Canny assumed the edges are the local maxima of and introduced the following criteria: 1. Good detection: It simply means to keep the 2. Good Localization: It means the detected edge SNR at x=0 as high as possible. should be as close as possible to x=0. 3. Low Spurious Response: Since input is single step, it is not allowed to have multiple maxima. 8 Outline • Review of Edge Detection methods • Scale Multiplication approach 1D signals The discrimination of noise. The scale multiplication. The thresholding. 2D images • Simulations 9 The idea of Scale-Multiplication An important issue is the scale of detection filter where: small-scaled filter is sensetive to noise and large-scaled filter filters some useful details. 10 Assumptions ∗ 2 2 2 ∗θ 1.4 1.2 1 0.8 0.6 0.4 0.2 0 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 11 Edge vs Noise Local regularity can be measured by Lipschitz exponent. A function f(x) is Lipschitz at iff there exists a constant such that in the neighborhood of =0 for Step Edge & =-0.5 for White Noise. 12 Scale Multiplication 13 Thresholding . for c . Where 1 2 . 14 2D Image , , , , . . 15 2D Image , , , , , . , , . , arctan 0.8 16 Outline • Review of Edge Detection methods • Scale Multiplication approach 1D signals The discrimination of singularity and noise. The scale multiplication. The thresholding. 2D images • Performance evaluation 17 Performance evaluation Figure of Merrit (The greater the F, better detection ) MS of Distance 18 Performance evaluation Noisy cameraman with SNR=16.5278 dB [Canny vs SMED] 19 Performance evaluation LoG vs SMED 20 Performance evaluation Canny vs SMED 21 Conclusion While the classic methods are sensitive to “noise”, the proposed method is able to handle the effect of the noise on detected edges. Tuning of different parameters is a concern in other methods while the proposed method is almost 100% automatic. 22 References 1. 2. Zhang, Bao; “Edge Detection by Scale Multiplication in Wavelet Domain”; Pattern Recognition Letters; pp. 1771-1784; Vol 23; 2002. Bao, Zhang, Wu; “Canny Edge Detection Enhancement by Scale Multiplication”; IEEE Trans. On Pattern Analysis and Machine Intelligence; pp. 1485-1490; Vol 27 No 9; 2005 23 24