ressearch project afm

parlab

research> activities

Research projects

Attentive vision

Nabil Ouerhani

Project program

CSEM-IMT common research program

Keywords

Low level vision, early vision, parallel processing, smart camera

Purpose of the project

Largely inspired by human vision, the visual attention paradigm leads to an efficient machine vision architecture that tends to use highly parallel but simple processing to locate few significant positions where sequential, more elaborate processing can then be applied. A first project Multicue Vision was devoted to the development of a realistic attention model and then, to the study of implementation aspects of this model. Now, an implementation of a real-time two-cue attention process is available which confirms the applicability of the pre-attentive part of this vision approach.

This project proposes now to link this pre-attentive stage of processing to a subsequent attentive processing stage. It is mainly concerned with the various links towards higher-level vision. The goal set for the next activities is thus to exploit and extend the now available stage of visual attention for performing subsequent segmentation and recognition processes in a prioritized, sequential way, best suited for implementation efficiency.

Attention guided segmentation

In a previous segmentation approach, spots of attention are used to guide the segmentation, for instance by serving as seeds for the growing of regions [1] . In this new approach, the spots of attention not only guide the location of the seeds, but also guide the selection of features used for segmenting. [2]

Recognition of trafic signs

The recognition system proceeds in two distinct steps. The preattentive step operates under a principle close to visual attention and produces a list of significant spots of attention. Each spot of attention is then analysed separately in a second step of attentive vision.


Original image	Spots of attention	Recognized road signs

Fig. 1 Spot of attention guided road sign recognition

References

[1] Nabil Ouerhani, Neculai Archip, Heinz Hügli & Pierre-Jean Erard, " A Color Image Segmentation Method Based on Seeded Region Growing and Visual Attention", Int. Journal of Image Processing and Communications, vol.8, no1, 2002

[2] Nabil Ouerhani and Heinz Hügli, "MAPS: Multiscale Attention-based Pre-Segmentation of Color Images", Proc. 4th International Conference on Scale-Space theories in Computer Vision 2003, 10-12 June 2003, Isle of Skye, Scotland, UK