Computer Vision, Part 1 Topics for Vision Lectures 1. Content-Based Image Retrieval (CBIR) 2. Object recognition and scene “understanding” Content-Based Image Retrieval Example: Google “Search by Image” Basic technique Extract Features (Primitives) Matched Results Similarity Measure Query Image Image Database Features Database Relevance Feedback Algorithm From http://www.amrita.edu/cde/downloads/ACBIR.ppt Each image in database is represented by a feature vector: x1, x2, ...xN, where xi = (xi1, xi2, …, xim) Query is represented in terms of same features: Q =(Q1, Q2, …, Qm) Goal: Find stored image with vector xi most similar to query vector Q • Distance measure: x i Q x1iQ1 x2iQ2 ... xm iQm • Possible distance measures d(Q, xi): – Inner (dot) product – Histogram distance (for histogram features) – Graph matching (for shape features) . . . Some issues in designing a CBIR system • Query format, ease of querying • Speed • Crawling, preprocessing • Interactivity, user relevance feedback • Visual features — which to use? How to combine? • Curse of dimensionality • Indexing • Evaluation of performance Types of Features Typically Used • Intensities • Color • Texture • Shape • Layout http://www.clear.rice.edu/elec301/Projects02/ artSpy/intensity.html Intensity histograms Color Features Hue, saturation, value Color Histograms (8 colors) http://www.owlnet.rice.edu/~elec301/Projects02/artSpy/patmac/mcolhist.gif http://www.owlnet.rice.edu/~elec301/Projects02/artSpy/patmac/mcolhist.gif Color auto-correlogram • Pick any pixel p1 of color Ci in the image I. • At distance k away from p1 pick another pixel p2. • What is the probability that p2 is also of color Ci? Red ? k p2 p1 Image: I From: http://www.cse.ucsc.edu/classes/ee264/Winter02/xgfeng.ppt • The auto-correlogram of image I for color Ci , distance k: C(k ) (I ) Pr[| p1 p2 | k, p2 IC | p1 IC ] i i i • Integrates both color information and space information. From: http://www.cse.ucsc.edu/classes/ee264/Winter02/xgfeng.ppt Two images with their autocorrelograms. Note that the change in spatial layout would be ignored by color histograms, but causes a significant difference in the autocorrelograms. From http://www.cs.cornell.edu/rdz/Papers/ecdl2/spatial.htm From: http://www.cs.cornell.ed u/rdz/Papers/ecdl2/spati al.htm Color histogram rank: 411; Auto-correlogram rank: 1 Color histogram rank: 310; Auto-correlogram rank: 5 Color histogram rank: 367; Auto-correlogram rank: 1 Texture representations • Gray-level co-occurrence • Entropy • Contrast • Fourier and wavelet transforms • Gabor filters Texture Representations Each image has the same intensity distribution, but different textures Can use auto-correlogram based on intensity (“gray-level co-occurrence”) Texture from entropy Images filtered by entropy: Each output pixel contains entropy value of 9x9 neighborhood around original pixel From: http://www.siim2011.org/abstracts/advanced_visualization_tools_ss_pao.html Texture from fractal dimension From: http://www.cs.washington.edu/homes/rahul/data/ic cv07.pdf Texture from Contrast • Example: http://www.clear.rice.edu/elec301/Projects02/artSpy/grainin ess.html Texture from Wavelets http://www.clear.rice.edu/elec301/Projects02/ artSpy/dwt.html http://www.clear.rice.edu/elec301/Projects02/ artSpy/dwt.html Shape representations Some of these need segmentation (another whole story!) • Area, eccentricity, major axis orientation • Skeletons, shock graphs • Fourier transformation of boundary • Histograms of edge orientations From: http://www.lems.brown.edu/vision/researchAreas/ShockMatching/shock-ed-match-results1.gif Histogram of edge orientations From: http://www.cs.ucl.ac.uk/staff/k.jacobs/teaching/prmv/Edge_histogramming.jpg Visual abilities largely missing from current CBIR systems • Object recognition • Perceptual organization • Similarity between semantic concepts Examples of “semantic” similarity Image 1 Image 2 Examples of “semantic” similarity Image 1 Image 2 Examples of “semantic” similarity Image 1 Image 2 “In general, current systems have not yet had significant impact on society due to an inability to bridge the semantic gap between computers and humans.” Image Understanding and AnalogyMaking Bongard problems as an idealized domain for exploring the “semantic gap”