Related Experiment Videos
Binding and segmentation of multiple objects through neural oscillators inhibited by contour information
Mauro Ursino1, Giuseppe-Emiliano La Cara, Alessandro Sarti
1Dipartimento di Elettronica, Informatica e Sistemistica, University of Bologna, viale Risorgimento 2, I-40136 Italy. mursino@deis.unibo.it
Biological Cybernetics
|July 2, 2003
Summary
This study introduces a novel neural network model for visual object recognition. The network effectively segments multiple objects in complex scenes using simplified Wilson-Cowan oscillators and contour detection, achieving near-perfect accuracy.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
- Computer Vision
Background:
- Temporal correlation of neuronal activity is a proposed mechanism for multiple object recognition.
- The binding and segmentation problem in visual processing remains a significant challenge.
- Gestalt principles, like connectedness, offer insights into visual organization.
Purpose of the Study:
- To develop and evaluate a neural network model for solving the binding and segmentation problem in visual scenes.
- To implement a system based on Wilson-Cowan oscillators that utilizes contour information for object separation.
- To test the model's performance across diverse visual scenes and under noisy conditions.
Main Methods:
- A two-dimensional network of simplified Wilson-Cowan oscillators was employed.
- Original, time-independent coupling terms facilitated neural binding.
- A two-layer processing approach was used: contour extraction via 'retinal cells' and selective inhibition in the second layer.
- A global inhibitor prevented spurious synchrony between distinct objects.
Main Results:
- The network achieved near 100% accuracy in segmenting objects across 21 different visual scenes with a single parameter set.
- The model demonstrated robustness against dynamical noise.
- It successfully segmented objects with both positive and negative contrast and handled image fragmentation.
- A limitation identified was sensitivity to static noise, suggesting future improvements in contour enhancement.
Conclusions:
- The proposed Wilson-Cowan oscillator network effectively addresses the binding and segmentation problem in visual scenes.
- The model's high accuracy, robustness to noise, and parameter independence highlight its potential for visual processing.
- Further research should focus on enhancing contour detection mechanisms to mitigate static noise sensitivity, aligning with biological visual cortex functions.