A self-organising neural network model of image velocity encoding

Area of Science:

• Computational Neuroscience
• Computer Vision
• Artificial Intelligence

Background:

• The aperture problem in computer vision arises from observing a moving object through a limited aperture, creating ambiguity in motion direction.
• Understanding how biological systems solve the aperture problem is crucial for developing advanced visual processing models.
• Previous models often lacked biological plausibility or failed to fully account for the complexities of visual input.

Purpose of the Study:

• To develop a self-organising neural network capable of mapping image velocities of rigid objects.
• To investigate a computational model that instantiates the Intersection of Constraints solution to the aperture problem.
• To demonstrate the biological plausibility of the developed neural network architecture.

Main Methods:

• Utilized a self-organising neural network architecture.
• Input data consisted of spatial image components in the Fourier domain.
• The network was trained on inputs derived from rigid translation of textures.

Main Results:

• The network successfully mapped image velocities of rigid objects topologically onto a neural layer.
• The network's computation aligns with the Intersection of Constraints solution for the aperture problem.
• Network connectivity emerged organically from exposure to visual input, demonstrating self-organisation.

Conclusions:

• The developed neural network provides a biologically plausible model for solving the aperture problem.
• The findings suggest that complex visual computations can arise from simple, self-organising principles.
• The model's consistency with psychophysical evidence supports its relevance to understanding biological vision.