Analysis and evaluation of images of multi-part objects in order to extract the silhouette
Jean-François Bernier
Robert Bergevin (Supervisor)
Problem: A 2D image is composed of many pixels, representing the textures or limits of an object. This forms an ensemble of boundaries where the human brain can easily recognize the main shape of an image, which is not the case for a computer. One thus strives to find the best multi-part object of the scene using a simplified model of the latter so as to reduce the possibilities: all of the boundaries are approximated using lines and arcs of circles.
Motivation: Solving a general problem involving the search for the best object in a scene opens the door for new methods of problem solving. The research work on common characteristics of contours, the ordering of contour parts and the completion of partial images are all similar subjects and can all be at least partially resolved with the work presented herein. Moreover, the process of locating an object in a scene enables one to then carry out other treatments on the object in the real world. If the computer recognizes the zone of interest, all of the current treatment algorithms could be applied, without the need for a hyper-controlled environment (background of a known colour, predetermined lighting, predefined objects to be analyzed, etc.).
Approach: The approach used here consists in choosing the first images to be treated and an efficient generation of the sub-groups on which one must apply the evaluation criteria. This can include, for example, the pretreatment and shape recognition algorithms. A mark is attributed to each sub-group so as to classify them in a logical order. In a more practical manner, the generation will be carried out first with the help of the GraphSearch algorithm, so as to offer the possibility of generating the complete ensemble of solutions, while limiting the search through the application of constraints as soon as it is known that the continuation is non valid. The work in this area is divided into two different approaches: the basic and high-level analysis, i.e. the treatment of pixels on a one by one basis so as to discover certain common characteristics, or the treatment of pixels forming part of a whole and attempting to extract the shapes. Each of these studies begins at a certain level and remains there until the results are obtained. Here the method used allows one to begin a low-level analysis, pixel by pixel, but also to summarize and group these pixels into ensembles, so as to enable a study at a higher level (by ensemble pairs for example). Finally, global evaluation criteria will be applied on all of the pixel groups, thus reaching the summit of the rising levels of complexity.
Challenges: Two major challenges are present in this research: 1) The proper use of efficient criteria enabling one to perfecting distinguish one shape from another. One can consider quality criteria such as the level of noise, angles de suivi (tracking angles?), space between the contours, etc. 2) The explosive combination of possibilities of choices of sub-groups. If, for example, one image contains 100 primitives modeling the contours, and each of these primitives has two extremities, image the number of cases to be treated for the order in which one covers all of the primitives! One must therefore eliminate the majority of the possibilities via certain treatments. Thus, the general idea is to have an algorithm offering the possibility to generate a tree of all of the possible solutions and to trim this tree using certain criteria and constraints.
Expected results: In theory, one expects to obtain more or less complete silhouettes of the main object in the image. It is very possible that the result contains various non-desirable elements, but it will always be possible to then apply other criteria so as to obtain a more-perfect result! We will attempt to compare already existing methods and those proposed here by examining the computation times, the difficulties considered and the memory space required.
Calendar: September 2004 - September 2006
Last modification: 2007/10/01 by jfbernie


©2002-. Computer Vision and Systems Laboratory. All rights reserved