190 likes | 515 Views
Color Edge Detection in Multiple Color Spaces. Jae Mo Park. March 17, 2008. I. Introduction. Edge Detector Overview Color Space Overview. Laplacian of Gaussian. Sobel. Roberts. Prewitt. Canny. Zero-crossing. Edge Detector Overview. Gray/Monochromatic Edge. I. Introduction.
E N D
Color Edge Detectionin Multiple Color Spaces Jae Mo Park March 17, 2008
I. Introduction • Edge Detector Overview • Color Space Overview
Laplacian of Gaussian Sobel Roberts Prewitt Canny Zero-crossing Edge Detector Overview Gray/Monochromatic Edge I. Introduction Edge Detection
Edge Detector Overview Color Edge I. Introduction
Edge Detector Overview Color Edge (1) Monochromatic-based techniques I. Introduction Monochromatic Edge Detector Red Color image Output Fusion Monochromatic Edge Detector Green Monochromatic Edge Detector Blue Ignores the relationship between the vector components ! (2) Vector-valued techniques • Treat the color information as color vectors in a vector space • without decoupling • Color Variants of the Canny Operator, Cumani Operator, • Vector Order Statistics, etc • Grayscale transform with Principal Component Analysis (PCA)
Color Space Overview RGB Hue + Saturation CIE Luminance + Chrominance CMYK • h1h2h3 • h1=R-G • h2=G-B • h3=B-R • : Showed best performance • in Sobel, Roberts, Vector • gradient, Difference vector • comparison • RGB • sRGB • Adobe RGB • Adobe Wide Gamut RGB I. Introduction • HSV(HSB) • HSL(HSI) • CIE XYZ • CIE LAB • CIE LUV • CIE UVW • YIQ(NTSC) , YUV, YDbDr • YCbCr • xvYCC
II. Methods • Color Variants of Canny Operator • Monochromatic Canny Operator Using Principal Component Analysis • Experiment Block Diagram
Input Image Linear Filtering Intensity Gradient Non-maximum Suppression (Thinning) Hysteresis Thresholding Output Edge Map Determine whether the gradient magnitude is a local maximum in the gradient direction Smoothing by Gaussian (σ) Then take x- and y- derivatives Refine edges using T-high, T-low, and direction Linear Filtering Intensity Gradient Non-maximum Suppression Input Image Output Edge Map Linear Filtering Intensity Gradient Non-maximum Suppression Hysteresis Thresholding Linear Filtering Intensity Gradient Non-maximum Suppression Choose II/L∞ I II III Combination Operator: L1, L2, L∞ norm Color Variants of Canny Operator Monochromatic Canny Operator II. Method Color Variants of Canny Operator
Calculate , C : ML estimate of mean C: ML estimate of covariance matrix Grayscale Canny Operator Using PCA Monochromatic Canny Operator II. Method Input Image Solve for λi, vi (i=1, 2, 3) Decorrelate Color Components (g1,g2,g3) Monochromatic Canny Operator Output Edge Map λ: eigenvalues of C v: eigenvectors of C λ1>λ2>λ3 gi [m,n]=viT f [m,n] Take g1, the first principal component
Experiment Block Diagram II/L∞ Color Canny Operator II. Method Color Input Image RGB to other color spaces Normalize each color component Compare edge maps PCA, Monochromatic Canny Operator Implement and use rgb2xyz, xyz2lab, rgb2ycbcr,rgb2h1h2h3 rgb2hsv, and rgb2ntsc Compare based on accuracy Implement CannyColor.m, CannyGray.m, and PCA.m
III. Result • Test images • Color Canny Operator in multiple color spaces • Monochromatic Canny with PCA in multiple color spaces
Performance Indicator • Can you read the words? • Can you count the number of Lemons/oranges? • Can you recognize the shape of banana/grape/pineapple? Test images III. Result
Results of Color Canny Operator RGB YCbCr h1h2h3 CIE XYZ HSV CIE L*a*b* YIQ
Results of Monochromatic Canny with PCA RGB YCbCr CIE XYZ h1h2h3 HSV CIE L*a*b* YIQ
Gray with equal weight Gray with stnd weight RGB YIQ Comparison CIE XYZ RGB III. Result Gray=0.30R+0.59G+0.11B
Conclusion • Best result: • Color Canny-RGB, CIEXYZ, YCbCr • Mono Canny with PCA-RGB 2. The results obtained in RGB color space outperformed results in other color spaces. IV. Conclusion 3. The results using Color Canny operator generally better than results obtained using Gray Canny operator with PCA. 4. Both detectors did not show good performance in h1h2h3 color space. 5. Not easy to decide which colorspace is the best.