Motivation
The cognitive importance of parts of visual form in human perception has been theoretically and experimentally verified. However, the identification of shapes given their parts is still an unsolved problem in shape similarity. The main facets of this problem are:
1) A fish tail contour
2) A contour of a partially occluded fish
3) A contour of a partially occluded fish with a distortion.
1. the length problem,
2. the scale problem, and
3. the distortion problem.
These problems are best illustrated by the figures at right. The first figure is that of a fish tail contour. The second is that of a partially occluded fish contour. The third is another partially occluded fish contour, though with a distortion in the form of a large spike. Given a significant part of visual form as a query, such as the fish tail shown in the first image, our goal is to find similar shapes containing the query part; this means that we need to find the tail as part of some other fish contours, such as the contours in the second and third images. Length and scale differences are readily apparent between the first image and the others. The second and third contours cover a larger area of the fish body, but are therefore zoomed out relative to the first image. By current methods, normalizing the length would not deal with the scale problem, and would even worsen it in this case. Normalizing the scale by current methods would require an equal length, a condition not found in many cases such as this one. Trying to deal with both simultaneously then creates a chicken-and-egg problem. The third image demonstrates the problem of distortion. While to the human eye the distorted fish contour is still easily recognizable, to a computer using current methods, the distortion completely changes the identifying contour to a state not similar to the query image. These issues provided the main motivation for the development of the proposed approach.
Our goal is to build a representation of shapes that is suitable for shape-based recognition of contour parts. The input is edge images obtained by Canny edge detector. Our experimental setting is composed of the three steps:
1. Building a part segment database for retrieval.
2. Extraction of contour parts from images as a query part.
3. Finding the most similar part quickly and accurately.
| <- Prev | Next -> |