Journal of Visual Communication and Image Representation
Scale multiplication in odd Gabor transform domain for edge detection
Introduction
Edge detection, which has attracted the attentions of many researchers, is one of the most important areas in lower level computer vision because the success of higher level processing such as object detection, object recognition, and scene interpretation relies heavily on good edge detection. Generally, edges are referred to those pixels around which there is a large gray variation. Most existing edge detection methods are gradient-based, such as some simple edge detection operators Roberts and Sobel as well as some sophisticated operators Canny [1] and Susan et al. [2]. An overview of edge detection techniques was given in [3]. Some new developed edge detection algorithms can be found in [4], [5]. In this paper, we focus on the multi-scale analysis of odd Gabor transform to develop a new edge detection technique.
Gabor transform, or named by short time Fourier transform (STFT), can obtain the lower limit of joint time-frequency resolution under Heisenberg Uncertainty Principle. Daugman [6] has taken a deeply investigation on it from the view of visual nerve perceptibility and found that the receptive profile of simple cells in mammalian visual systems can be closely fitted by two-dimension Gabor function. Now two-dimension Gabor transform has obtained comprehensive applications in computer vision field such as image segmentation, object recognition, and so on.
Under Canny’s three optimal criterions for edge detection, R. mehrotral [7] pointed out that odd Gabor filter has good edge detection performance. In real applications, as suggested in [7] a simple edge operator should be used to estimate the gradient orientation. Thus the estimation error based on the above simple edge operator, which generally is high especially in the noise case, will largely influence the accurate construction of the odd Gabor filter templates. In addition, we cannot implement the convolution between odd Gabor filter and the input image by using FFT since each pixel corresponds to different odd Gabor filter template.
Multi-scale analysis is a popularly adopted technique in computer vision. A systematic analysis for multi -scale space theory has been given by Tony Lindeberg [8]. The applications of multi scale in edge detection can be dated back to Marr and Hildreth’s work [9]. Based on the framework of Canny [1], some techniques for edge detection based on multi scale wavelet transform have been developed [10], [11], [12]. In addition, Park et al. [13] have proposed a region based multi scale edge detection method. Recently Zhang [14] extended the ideas of Xu [15] and Sadler [16] and proposed a scale multiplication based edge detection scheme, in which adjacent scales are multiplied in wavelet domain to sharpen the edge responses. In our work, the similar scale multiplication idea was applied to the odd Gabor transform domain, but a more favorable edge response output energy function was presented.
Voorhees et al. [17] pointed that the derivative based edge response can be fitted by Rayleigh distribution, which is also shown in our work to be applicable for odd Gabor filter based edge response. Thus the pdf of sharpened edge response output can be approximately modeled by an exponential distribution since there exist strong correlation between two edge response outputs with two adjacent scale factors. Moreover, an adaptive threshold determining scheme is proposed based on the analysis of the pdf of sharpened edge response output.
This paper is organized as follows: Section 2 gives an introduction to odd Gabor based edge detection and proposes an improvement on it. To sharpen the edge response, a scale multiplication in odd Gabor transform domain is proposed and a more favorable edge response output energy function is given in Section 3. In Section 4, we theoretically analyze the sharpened edge response output energy distribution property and propose an adaptive threshold selection scheme. Some experimental results of synthetic and natural images are presented in Section 5. Finally we give our conclusions in Section 6.
Section snippets
Odd Gabor filter based edge detection
Gabor filters, which have been shown to fit well the receptive fields of the majority of simple cell in the primary visual cortex [6], are modulation products of Gaussian and complex sinusoidal signals. A 2D Gabor filter oriented at angle θ is given by:where is a Gaussian function, σg is the standard deviation of the circle Gaussian along x and y, and ω denotes the spatial frequency.
Fig. 1 shows the even real and odd
Scale multiplication in odd Gabor transform domain
Multi-scale method is a commonly adopted technique in computer vision. Here, we form the multi-scale odd Gabor filter aswhere k ∈ Z is the scale factor and i ∈ [1,2] denotes orientation index.
Following the idea of Zhang [14], Xu [15] and Sadler [16], an adjacent scale multiplication in odd Gabor transform domain is performed to obtain a sharpened edge response output , and we have:where Ek(x,y) is the
Rayleigh distribution of edge response based on odd Gabor filter
In [17] Voorhees et al., have pointed that the derivative based edge response can be fitted by Rayleigh distribution. For the first derivative property of odd Gabor filter, the odd Gabor filter based edge response E (x,y) can also be fitted well by Rayleigh distribution. To evaluate the degree of fitness between the real pdf p(x) and the estimated distribution q(x) of odd Gabor based edge response, the Bhattacharya distance is adopted as similarity measure, and we have:
Experimental results and analysis
Because of the subjectivity of edge detection, it is difficult to compare the performance of two edge detectors on most real-world natural images [19]. Here, we use some synthetic [20] and real world images to test the effectiveness of ESMG.
Conclusion
We proposed an adaptive edge detection technique based on the multi scale analysis in odd Gabor transform domain (ESMG). To overcome the shortcomings of original odd Gabor-based edge response obtaining method, an improved scheme is suggested. With adjacent scales multiplication in odd Gabor transform domain, a sharpened edge response output, which can more effectively resist the inverse influence from noise contamination on the performance of edge detector, can be obtained. By analysis of edge
Acknowledgments
This work was funded in part by National Natural Science Foundation of China (Nos. 60373028, 60475010 and 90604032), Program for New Century Excellent Talents in University, Specialized Research Fund for the Doctoral Program of Higher Education (No. 20030004016), Specialized Research Foundation of BJTU (No. 2005SZ005), Research Foundation of Beijing Jiaotong University (No. 2005SM013).
References (22)
- et al.
Odd Gabor filter-based edge detection
Pattern Recognit.
(1992) - et al.
Multiscale edge detection and feature binding: an integrated approach
Pattern Recognit.
(1998) - et al.
Multiresolution edge-detection techniques
Pattern Recognit.
(1995) - et al.
Edge detection by scale multiplication in wavelet domain
Pattern Recognit. Lett.
(2002) - et al.
Comparison of edge detector: a methodology and initial study
Comput. Vis. Image Understand.
(1998) A computational approach to edge detection
IEEE. Trans. Pattern Anal. Mach. Intell.
(1986)- et al.
SUSAN—A new approach to low level image processing
Int. J. Comput. Vis.
(1997) - et al.
Edge detection techniques-an overview
Int. J. Pattern Recognit. Image Anal.
(1998) - et al.
Edge detection with embedded confidence
IEEE Trans. PAMI
(2001) Image field categorization and edge/corner detection from gradient covariance
IEEE Trans. PAMI
(2000)
Complete discrete 2-D Gabor transforms by neural networks for image analysis and compression
IEEE Trans. Acoust. Speech Signal Process.
Cited by (29)
Gabor filter-based edge detection: A note
2014, OptikCitation Excerpt :Meanwhile Mehrotra also noted that this conclusion was applicable to 2-D step edge. Since then, Gabor filter is widely used in edge detection [2–6] and other fields. But there are two shortcomings about Mehrotra's conclusion: Firstly, it lacked strict mathematical explanation for σw≈1 in the 2-D case [4] though it is reasonable; secondly, the filter response was maximized only at the midpoint of step edge and it cannot represent the entire edge.
Towards finger-vein image restoration and enhancement for finger-vein recognition
2014, Information SciencesCitation Excerpt :Using Euler formula, a Gabor wavelet can be decomposed into two parts: real and imaginary. The real part, usually called even Gabor wavelet, is suitable for ridge detection [73], while the imaginary part, usually called odd Gabor wavelet, is suitable for edge detection [80]. Since veins present ridges, even Gabor wavelets are desirable for the purpose of venous region enhancement.
Automated defect detection in uniform and structured fabrics using Gabor filters and PCA
2013, Journal of Visual Communication and Image RepresentationCitation Excerpt :Although optimal filters have some obvious advantages over other methods, the selection of their parameters is crucial and difficult. Another approach for selecting Gabor filter parameters is reported in [27], where an odd-symmetric filter is designed to match with the texture features of defect-free fabric image; single-scale odd Gabor filters are also used in [28] for purposes of edge detection, but the feature identification and dimension reduction strategies are based on mean and variance threshold comparison of the edge pixels statistical distribution. In this paper, we extend the work presented in [29] by exploiting the performance of Gabor filters considering an algorithm devoted to detect texture defects, the dimension and orientation of which vary randomly in uniform and structured fabrics.
Image bilevel thresholding based on stable transition region set
2013, Digital Signal Processing: A Review JournalFinger-vein ROI localization and vein ridge enhancement
2012, Pattern Recognition LettersImage bilevel thresholding based on multiscale gradient multiplication
2012, Computers and Electrical EngineeringCitation Excerpt :The FDoG filters with small scale are sensitive to the edge signals but also prone to the noise, whereas the filters with large scale are robust to the noise but could filter out the fine details. To overcome this dilemma, an alternative scheme is the multiscale multiplication, namely, multiplying the responses of the filters at several different scales [15]. This idea is introduced into this paper to compute the multiscale gradient information PGM.