Retinal Modeling: Segmenting Motion from Spatio-Temporal Inputs using Neural Networks
Abstract
We applied two first-order, linear, time-varying, differential equations to the task of segmenting motion from sequences of images. The equations are modified Grossberg formulas for long-term and short-term memory models characterizing the neurotransmitter and cell-activity levels of a synapse and neuron. We described how a two layered, sensory, neural network can be built using the equations to simulate the amacrine neurons of the retina. The model is defined using adaptive input nodes (adaptive model) and is compared to a similar model without these nodes (O and G model). By replicating the basic amacrine neuron model to form both one- and two-dimensional arrays, we created a novel method for processing images over time and space. To simulate the veto effect observed in shunt inhibitory synaptic junctions, we applied a nonrecurrent, asynchronous, inhibitory region in the receptive field of our amacrine neural model. We show how this effects the performance of the model ill one dimension. In two dimensions we investigate the models' response to synthesized imagery (pristine) and to real, forward looking infrared radar (FLIR) images. The output of our models are further processed through two types of moving-average filters - causal and noncausal.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1992
- Accession Number
- ADA258854
Entities
People
- David E. Swanson
Organizations
- Air Force Institute of Technology