Image Segmentation in the Field of the Logarithmic Image Processing Model
| Autor: | Enguerrand Couka, Josselin Breugnot, Bassam Abdallah, Maxime Carré, Joris Corvo, Michel Jourlin |
|---|---|
| Rok vydání: | 2013 |
| Předmět: |
Segmentation-based object categorization
business.industry Scale-space segmentation Pattern recognition Image processing Image segmentation Image texture Region growing Computer Science::Computer Vision and Pattern Recognition Human visual system model Computer vision Artificial intelligence Range segmentation business Mathematics |
| Popis: | The present paper deals with image segmentation, which constitutes a crucial step in image processing. In fact, the initial grey levels number is generally too large to permit the analysis in good conditions of the considered image and it is necessary to define regions (segments) whose pixels possess some properties in common, in terms of homogeneity, entropy, texture… The segmentation quality is also linked to the pertinence of boundaries separating regions (high level of contrast for example). To address this segmentation goal, a lot of methods exist, generally depending on the choice of some arbitrary tools like metrics, similarity or homogeneity parameters and sometimes on an a priori knowledge concerning the desired number of classes. We have decided to locate our study in the LIP (Logarithmic Image Processing) framework because of this Model compatibility with the Human Visual System. First we propose LIP versions of classical algorithms like multi-thresholding, k-means and region growing (Part 2 and Part 3). For this last technique, we present a “systolic” approach. A special highlight is given on Hierarchical classifications (Part 4), because they suppress some subjective initial hypotheses concerning for example: - the moment where a region becomes inhomogeneous and must be divided - what is the number of significant classes present in the studied image In fact, such methods have the advantage of producing on one hand all the possible segmentations and on the other hand a “cost” function based on an ultra-metric concept which permits to decide what are the most pertinent levels of classification. This 4th part of the paper ends with a novel “Gravitational Clustering” algorithm starting from the universal attraction law of Newton. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|