An important component in computer vision is image analysis, i.e. theidentification of objects and their analysis, in terms of shape, colour,texture, etc. Combined with other domains like artificial intelligence,image analysis is in fact the basic block of computer vision, withapplications areas as wide as industrial inspection, guidance, remotesensing, video processing or medicine. Image analysis is actually thegeneral context of this project.
One of the central issues in this field is clearly the introduction of apriori information in the recognition scheme. The appropriate manner ofintroducing it is the central point of this project. Up to now, in theclassical segmentation scheme, the pre-processing and segmentation tasksuse very little a priori knowledge about the problem to solve. A priori isgenerally introduced in the high-level processing step, when objects havealready been segmented, in order to analyze and classify them. This is acrucial point of the pattern recognition problem. In the process of betterrecognizing objects in a complex scene, a priori information about targetobjects has to be included as soon as possible, i.e. in the segmentationprocess. Some segmentation techniques, especially those based onvariational approaches, are very interesting in this context, since theyare intrinsically well suited for including a priori information in theirimplementation.
Our project, which is the continuation of a previous Swiss NSF projectstarted in 2001, will continue and extend the work that we have alreadydone on the inclusion of a priori shape information in active contours. Inthe previous project, we have defined a framework based on a statisticalshape representation by a PCA of the Signed Distance Function of objecttemplates, and on a new functional defined to take this a priori shapeinformation into account in the segmentation by geodesic active contour.In this project, we will concentrate on the following aspects:
- integrate contour, region and shape information into a singlevariational framework for image segmentation by active contours,
- define new shape similarity measures compatible with this framework,that do not assume Gaussian distributions of the shape descriptors, as itis currently done in the PCA approach. Non-parametric classificationschemes will be deeply investigated here,
- develop new texture descriptors, also compatible with the variationalframework, in order to segment objects that are homogeneous not only incoulour/gray level, but also in their texture. Specifically, wavelet-basedtexture description will be studied and incorporated in the variationalframework for image segmentation.
At the end of this project, we will thus have a complete variationalframework able to include a priori information about the shape and thetexture of the objects to segment, into a robust and accurate segmentationscheme.
Many applications can be foreseen for this technique, since segmentationis obviously one of the main issues in computer vision. They include videoobject tracking and image retrieval, multimodal image segmentation andanalysis, medical image