Image Blending W ES L E Y N A S LU N D CO LO R A D O S C H O O L O F M I N ES D EC E M B E R 2 , 2 0 1 3 How to put bears in your hot tub (safely) What’s my motivation? PHOTO SPHERE IMAGE STITCHING Registration ◦ Match features Calibration ◦ Correct lens distortion, exposure differences, vignetting, etc. Blending ◦ Seamline ◦ Fusion Simplify the problem SATELLITE LANDSAT IMAGES GLOBAL ORTHO AERIAL IMAGES Assumptions IMAGE STITCHING Registration ◦ Match features Calibration ◦ Correct lens distortion, exposure differences, vignetting, etc. Blending ◦ Seamline ◦ Fusion Seamline Blending Simple Solution: Feathering , 1 , , ∗ 1 , 2 , , 1 ∗ 2 , Seamline Blending Better: Multiresolution spline (Pyramid Blending) • Avoids blurring, doubling • Computationally efficient Seamline Blending Color images • “Exposure” differences still an issue • Resolved later • Could use radiance maps to balance Seamline Blending Alternate method: Optimal Seamline Cut • Apply Dijkstra’s Algorithm • Cost map from cross correlation • Find least cost path (“avoid obstacles”) 1 → 0 → 1 Image Fusion • Methods • Average • Details obscured • Laplacian Pyramid • Spectral detail loss • Wavelet Analogy: Logo is Landsat image Apple is Ortho Aerial image Image Fusion Wavelet Families: Properties Mother Father Daughter Image Fusion Why do these properties matter? ◦ Discrete Wavelet Transform (DWT) ◦ Wavelet family provides orthonormal basis Image Fusion Wavelet Fusion ◦ Spatial detail contained in high frequencies ◦ Decomposition extracts detail ◦ Several combination methods ◦ Substitutive ◦ Low resolution detail replace with high resolution detail ◦ Additive ◦ Add detail coefficients ◦ Weighted ◦ Based on mean/standard deviation or other criteria Image Fusion MATLAB’s wavemenu GUI for wfusimg() ◦ Look at source code: edit wfusimg Comparing methods ◦ Wavelets (Haar) ◦ Fusion methods Image Fusion Low Detail (1) High Detail (2) Max, Max Image Fusion Decomposition Comparison 1 Level 3 Levels 5 Levels Image Fusion Want image 1 consistency and image 2 detail Max, Max Max, Mean Max, img2 Image Fusion Quantitative Comparisons ◦ Entropy ◦ Higher entropy, higher detail ◦ Correlation coefficient ◦ How close to original the fused image is ◦ Warping ◦ Measures spectral distortion Image Fusion Approximation Detail max max max mean max img2 Level 2 2 2 Entropy 5.002327 4.825146 4.711708 Image Entropy Logo 3.61948 Apple 5.150525 Correlation 0.725695 0.721694 0.70458 Warping 0.629166 0.398735 0.888424 Image Fusion Satellite Landsat Image Global Ortho Aerial Image Image Fusion SATELLITE LANDSAT IMAGES GLOBAL ORTHO AERIAL IMAGES Image Fusion Max, Max Max, Img2 Improvements/Future Work Fusion methods ◦ Mathematical formulation is limiting for weighted methods ◦ Need inseparable filters for best results ◦ Quantitative comparisons are somewhat misleading ◦ Methods are not standardized ◦ Difficult to compare directly ◦ Compare results of different wave families ◦ Haar wavelets are approximations of Daubechies wavelets Questions?