EDGE DETECTION Presentation by Sumit Tandon Department of Electrical Engineering University of Texas at Arlington Course # EE6358 Computer Vision
September 13, 2005EE Computer Vision2 CONTENTS Introduction Types of Edges Steps in Edge Detection Methods of Edge Detection First Order Derivative Methods First Order Derivative Methods - Summary Second Order Derivative Methods Second Order Derivative Methods - Summary Optimal Edge Detectors Canny Edge Detection Edge Detector Performance Line Detection Convolution based technique Hough transform Application areas
September 13, 2005EE Computer Vision3 INTRODUCTION Edge - Area of significant change in the image intensity / contrast Edge Detection – Locating areas with strong intensity contrasts Use of Edge Detection – Extracting information about the image. E.g. location of objects present in the image, their shape, size, image sharpening and enhancement
September 13, 2005EE Computer Vision4 TYPES OF EDGES Variation of Intensity / Gray Level Step Edge Ramp Edge Line Edge Roof Edge
September 13, 2005EE Computer Vision5 Steps in Edge Detection Filtering – Filter image to improve performance of the Edge Detector wrt noise Enhancement – Emphasize pixels having significant change in local intensity Detection – Identify edges - thresholding Localization – Locate the edge accurately, estimate edge orientation
September 13, 2005EE Computer Vision6 Noisy Image Example of Noisy Image
September 13, 2005EE Computer Vision7 METHODS OF EDGE DETECTION First Order Derivative / Gradient Methods Roberts Operator Sobel Operator Prewitt Operator Second Order Derivative Laplacian Laplacian of Gaussian Difference of Gaussian Optimal Edge Detection Canny Edge Detection
September 13, 2005EE Computer Vision8 First Derivative At the point of greatest slope, the first derivative has maximum value E.g. For a Continuous 1- dimensional function f(t)
September 13, 2005EE Computer Vision9 Gradient For a continuous two dimensional function Gradient is defined as
September 13, 2005EE Computer Vision10 Gradient Approximation of Gradient for a discrete two dimensional function Convolution Mask Gx = Gy = Differences are computed at the interpolated points [i, j+1/2] and [i+1/2, j] Demo
September 13, 2005EE Computer Vision11 Gradient Methods – Roberts Operator Provides an approximation to the gradient Convolution Mask Gx= Gy = Differences are computed at the interpolated points [i+1/2, j+1/2] and not [i, j] Demo
September 13, 2005EE Computer Vision12 Roberts Operator - Example The output image has been scaled by a factor of 5 Spurious dots indicate that the operator is susceptible to noise
September 13, 2005EE Computer Vision13 Gradient Methods – Sobel Operator The 3X3 convolution mask smoothes the image by some amount, hence it is less susceptible to noise. But it produces thicker edges. So edge localization is poor Convolution Mask Gx = Gy= The differences are calculated at the center pixel of the mask. Demo
September 13, 2005EE Computer Vision14 Sobel Operator - Example Compare the output of the Sobel Operator with that of the Roberts Operator: The spurious edges are still present but they are relatively less intense compared to genuine lines Roberts operator has missed a few edges Sobel operator detects thicker edges Will become more clear with the final demo Outputs of Sobel (top) and Roberts operator
September 13, 2005EE Computer Vision15 Gradient Methods – Prewitt Operator It is similar to the Sobel operator but uses slightly different masks Convolution Mask Px = Py = Demo
September 13, 2005EE Computer Vision16 First Order Derivative Methods - Summary Noise – simple edge detectors are affected by noise – filters can be used to reduce noise Edge Thickness – Edge is several pixels wide for Sobel operator– edge is not localized properly Roberts operator is very sensitive to noise Sobel operator goes for averaging and emphasizes on the pixel closer to the center of the mask. It is less affected by noise and is one of the most popular Edge Detectors.
September 13, 2005EE Computer Vision17 Second Order Derivative Methods Zero crossing of the second derivative of a function indicates the presence of a maxima
September 13, 2005EE Computer Vision18 Second Order Derivative Methods - Laplacian Defined as Mask Demo Very susceptible to noise, filtering required, use Laplacian of Gaussian
September 13, 2005EE Computer Vision19 Second Order Derivative Methods - Laplacian of Gaussian Also called Marr-Hildreth Edge Detector Steps Smooth the image using Gaussian filter Enhance the edges using Laplacian operator Zero crossings denote the edge location Use linear interpolation to determine the sub-pixel location of the edge
September 13, 2005EE Computer Vision20 Laplacian of Gaussian – contd. Defined as Greater the value of , broader is the Gaussian filter, more is the smoothing Too much smoothing may make the detection of edges difficult
September 13, 2005EE Computer Vision21 Laplacian of Gaussian - contd. Also called the Mexican Hat operator
September 13, 2005EE Computer Vision22 Laplacian of Gaussian – contd. Mask Demo Discrete approximation to LoG function with Gaussian = 1.4
September 13, 2005EE Computer Vision23 Second Order Derivative Methods - Difference of Gaussian - DoG LoG requires large computation time for a large edge detector mask To reduce computational requirements, approximate the LoG by the difference of two LoG – the DoG
September 13, 2005EE Computer Vision24 Difference of Gaussian – contd. Advantage of DoG Close approximation of LoG Less computation effort Width of edge can be adjusted by changing 1 and 2
September 13, 2005EE Computer Vision25 Second Order Derivative Methods - Summary Second Order Derivative methods especially Laplacian, are very sensitive to noise Probability of false and missing edges remain Localization is better than Gradient Operators
September 13, 2005EE Computer Vision26 Optimal Edge Detector Optimal edge detector depending on Low error rate – edges should not be missed and there must not be spurious responses Localization – distance between points marked by the detector and the actual center of the edge should be minimum Response – Only one response to a single edge One dimensional formulation Assume that 2D images have constant cross section in some direction
September 13, 2005EE Computer Vision27 Mathematical Formulation of Performance Criteria Notation Edge : G(x) Noise : n(x) Impulse response of Filter : f(x) Assumptions The edge is located at x=0 Filter has finite response bounded by [-W, W] Noise is of Gaussian nature n 0 : Root mean squared noise amplitude per unit length
September 13, 2005EE Computer Vision28 First Criterion: Edge Detection Response of filter to the edge: RMS response of filter to noise : First criterion: output Signal to Noise Ratio
September 13, 2005EE Computer Vision29 Second Criterion: Edge Localization A measure that increases as the localization increases is needed Reciprocal of RMS distance of the marked edge from center of true edge is taken as the measure of localization Localization is defined as:
September 13, 2005EE Computer Vision30 Maximizing the product The design problem is reduced to maximization of the product:
September 13, 2005EE Computer Vision31 Third Criterion – Elimination of multiple responses In presence of noise several maxima are detected – it is difficult to separate noise from edge We try to obtain an expression for the distance between adjacent noise peaks The mean distance between the adjacent maxima in the output is twice the distance between the adjacent zero crossings in the derivative of output operator
September 13, 2005EE Computer Vision32 Elimination of Multiple Responses – contd. Expected number of noise maxima in the region of width 2W is given by Fixing k will restrict the number of noise maxima that could lead to false response – this is a constraint imposed on the design
September 13, 2005EE Computer Vision33 Finding Optimal Detectors by Numerical Optimization It is not possible to find a function f which will maximize the SNR, Localization product in the presence of the multiple response constraint If the function f is discrete then the computation is simplified to four inner products Additional constraints can be included by the use of Penalty Function – it has non-zero value when the constraint is violated f is then found out for maximizing the following expression
September 13, 2005EE Computer Vision34 Detector for Step Edges G(x) = Au -1 (x)
September 13, 2005EE Computer Vision35 Spatial scaling We can scale the filter width Detection improves with width, localization degrades with increase in width The uncertainty principle is
September 13, 2005EE Computer Vision36 Efficient approximation Depending on the above principles, several optimal edge detectors are calculated Best approximation to the above detectors is the First Derivative of Gaussian It is chosen because of the ease of computation in 2 dimensions
September 13, 2005EE Computer Vision37 Noise estimation Important to estimate the amount of noise in the image to set thresholds Noise component can be efficiently isolated using Weiner Filtering – requires the knowledge of the autocorrelation of individual components and their cross- correlation Noise strength is estimated by Global Histogram Estimation
September 13, 2005EE Computer Vision38 Thresholding Broken edges due to fluctuation of operator output above and below the threshold – results in Streaking Use double thresholding to eliminate streaking
September 13, 2005EE Computer Vision39 Two Dimensional Edge Detection In two dimensions edge has both position and direction A 2-D mask is created by convolving a linear edge detection function aligned normal to the edge direction with a projection function parallel the edge direction Projection function is Gaussian with same deviation as the detection function The image is convolved with a symmetric 2-D Gaussian and then differentiated normal to the edge direction
September 13, 2005EE Computer Vision40 Need for multiple widths and Feature Synthesis Width is chosen to give best tradeoff between detection and localization For different edges, width has to be different since SNR is different Feature Synthesis Smaller operator responses are used to predict the larger operator response New edge points are marked if the actual large operator output varies significantly from the predicted output
September 13, 2005EE Computer Vision41 Implementation of Canny Edge Detector Step 1 Noise is filtered out – usually a Gaussian filter is used Width is chosen carefully Step 2 Edge strength is found out by taking the gradient of the image A Roberts mask or a Sobel mask can be used
September 13, 2005EE Computer Vision42 Implementation of Canny Edge Detector – contd. Step 3 Find the edge direction Step 4 Resolve edge direction
September 13, 2005EE Computer Vision43 Canny Edge Detector – contd. Step 5 Non-maxima suppression – trace along the edge direction and suppress any pixel value not considered to be an edge. Gives a thin line for edge Step 6 Use double / hysterisis thresholding to eliminate streaking
September 13, 2005EE Computer Vision44 Canny Edge Detector – contd. Compare the results of Sobel and Canny
September 13, 2005EE Computer Vision45 Edge Localization – Subpixel Location Estimation Take Gaussian edge detector output without non-maxima suppression Error displacement is computed by taking the weighted mean of distance, along the edge gradient on either side of the edge
September 13, 2005EE Computer Vision46 Edge Detector Performance Criteria Probability of false edges Probability of missing edges Error in estimation of edge angle Mean square distance of edge estimate from true edge Tolerance to distorted edges and other features such as corners and junctions
September 13, 2005EE Computer Vision47 Figure of Merit Basic errors in a Edge Detector Missing Edges Error in localizing Classification of noise as Edge I A : detected edges I I : ideal edges d : distance between actual and ideal edges : penalty factor for displaced edges
September 13, 2005EE Computer Vision48 Line Detection Lines in an image present another form of discontinuity It is worthwhile to consider them separately since they are very frequent Methods Use of Convolution Mask Hough Transform
September 13, 2005EE Computer Vision49 Line Detection using Convolution Mask We need to decide the orientation and width of line to be detected Following masks are for single pixel wide lines in four directions
September 13, 2005EE Computer Vision50 Line Detection using Convolution Mask – contd. The mask for horizontal orientation is applied Lines are 5 pixels wide, mask is tuned for detecting 1 pixel wide line Line detector acts as edge detector Demo
September 13, 2005EE Computer Vision51 Line detection – Hough Transform Method to isolate the shapes from an image Performed after edge detection Not affected by noise or gaps in the edges Technique Thresholding is used to isolate pixels with strong edge gradient Parametric equation of straight line is used to map the edge points to the Hough parameter space Points of intersection in the Hough parameter space gives the equation of line on actual image
September 13, 2005EE Computer Vision52 Hough Transform - contd Parametric Equation of Straight Line
September 13, 2005EE Computer Vision53 Hough Transform – contd. Edge Detected Image Mapping of Edge Points in Hough Parameter Space
September 13, 2005EE Computer Vision54 Applications Enhancement of noisy images – satellite images, x-rays, medical images like cat scans Text detection Mapping of roads Video surveillance, etc.
September 13, 2005EE Computer Vision55 Applications Canny Edge Detector for Remote Sensing Images Reasons to go for Canny Edge Detector – Remote sensed images are inherently noisy Other edge detectors are very sensitive to noise
September 13, 2005EE Computer Vision56 Edge map of remote sensed image using Canny
September 13, 2005EE Computer Vision57 Image enhancement and sharpening using canny
September 13, 2005EE Computer Vision58 MATLAB Demo Get an image from the internet Apply following operators on the image and display the results Roberts Sobel Prewitt Laplacian of Gaussian Canny
September 13, 2005EE Computer Vision59 References Machine Vision – Ramesh Jain, Rangachar Kasturi, Brian G Schunck, McGraw-Hill, 1995 INTRODUCTION TO COMPUTER VISION AND IMAGE PROCESSING - by Luong Chi Mai Department of Pattern Recognition and Knowledge Engineering Institute of Information Technology, Hanoi, Vietnam The Hypermedia Image Processing Reference A Survey and Evaluation of Edge Detection Operators Application to Medical Images – Hanene Trichili, Mohamed-Salim Bouhlel, Nabil Derbel, Lotfi Kamoun, IEEE, 2002 Using The Canny Edge Detector for Feature Extraction and Enhancement of Remote Sensing Images - Mohamed Ali David Clausi, Systems Design Engineering, University of Waterloo, IEEE 2001 A Computational Approach to Edge Detection – John Canny, IEEE, 1986