A sketch-based approach to Image Retrieval

I. Thesis Demo Applets

II. Thesis Document

(Download Full Document)


Chapter 1: Introduction

Imagina System Overview: The user starts with an abstract mental image of the object they would like to retrieve. This image is made contrete by drawing a sketch. Features are extracted from the sketch, constructing multiple representations of the query. Specific similarity functions compare each representation of the query to representations previously extracted from the database images. The different similarity metrics are combined and the best matches overall are returned to the user along with visual information on how each match occured. The user can then use this information to formulate a more precise query.

Chapter 2: Query Systems

Chapter 3: Theoretical Foundations

Chapter 4: Color


A comparison of the similarity measures of the RGB and HSV color spaces. The colored points shown are all of the points considered similar for a given similarity cutoff value. The cutoff value has to be different in the two spaces in order to result in approximately equal areas of coverage since the two measures are unrelated. The actual cutoffs used are 0.88 for HSV and 0.74 for RGB. The colored points were selected from the HSV display in Figure 4-3, in comparison with a point roughly at the center of the colored points selected in by either similarity measure.

An actual color histogram match. Although the two faces are very different in lightness, both of them have a similarly shaped distribution of color in the pink hue color range, resulting in a high match value, of 0.97. This is in contrast to the figure following.

Chapter 5: Filters

The result of the application of a spot filter over two images. For some images the results are very good, while for others the results are very noisy

Chapter 6: Image Segmentation

The algorithm applied to the same image previously segmented with the Grow algorithms. The segmentations shown use cutoffs of 0.94, for the one next to the original image, 0.90, for the one below, and 0.80 for the third segmentation. The segmentation using a cutoff of 0.90 is the segmentation which the Imagina system defaults to upon loading this image. The color labels displayed are simply color markers used to visually denote the twenty most populous regions which are segmented by the algorithm, and have no intrinsic meanings.

Chapter 7: Boundary Edge Extraction

The Edges Boundary Extraction Algorithm finds an approximation to an edge within a window of computation. It calculates the center of mass C_in of the points belonging in the region and the center of mass C_out of the poitns belonging to the outside of the region within the window of computation. It uses the number of points belonging to the inside and the outside to calculate a weighted average of C_in and C_out, that will serve as an approximation to the tangent point to the region's boundary. A tangent direction is calculated to be parallel to the perpendicular bisector of C_inC_out. Finally, a new window center is calculated by moving along the tangent.