Section: New Results

Pattern recognition and statistical learning

Participants : Nozha Boujemaa, Michel Crucianu, Donald Geman, Wajih Ouertani, Asma Rejeb Sfar.

Machine identification of biological shapes

Participants : Asma Rejeb Sfar, Donald Geman, Nozha Boujemaa.

The increasing availability of digital images in the traditional sciences, the growing interest in biodiversity and the ongoing shortage of skilled taxonomists combine to make the automated categorization algorithms, increasingly important in many fields such as botany, agriculture and medicine. In this work, we propose a hierarchical coarse-to-fine approach to identify botanical species from a scanned sample of a plant organ, e.g., a leaf or a flower. To this end, we exploit domain-specific knowledge about taxonomy and landmarks. Promising recognition rates are achieved on several leaf datasets. Results have been submitted for publication.

Relevance feedback on partial image query

Participants : Wajih Ouertani, Michel Crucianu, Nozha Boujemaa.

scalability, hashing, SVM, prediction, approximation

Even if cropping an image to perform one-shot partial query filters a considerable amount of senseless regions for target definition, it does not yet clearly illustrates what the user is looking for. Indeed, the user target is either closer to the instance level or to the category level. Then we may have numerous suggested examples within the first response ranks while possibly some of them are totally irrelevant examples.

We claim that a localization interaction is still more appropriate than having a holistic decision about image relevance if it is performed on more examples. We go beyond the first partial query and investigate machine learning process to learn intention iteratively and interactively. Our learning process is based on what user delimit within additional images taken from the first response ranks. Our motivations include dealing with semantic gap revealed by local features hit falling into false regions within retrieved images. Those images might be either totally irrelevant, where all partial zones are out of the interest, or partially relevant, not because of the zones expected by the system (false-localization) but rather because of some missed zones. Through local annotations we expect the ability of redirecting the recognition session to those relevant regions and studying how much we can reduce the semantic gap within interactive localization.

This year, we studied several learning strategies based on several assumptions heuristically extracted on user interaction. The presented strategies have been also combined with features filtering within object representation. The filtering includes grouping contextualizing and varying features set representations.