What can engineers learn from insect vision?

Abstract

The mechanisms of insect vision contrast sharply with those of contemporary artificial visual systems and yet present a compelling example of w hat is possible with a stripped-dow n system of minimal weight.The first step is to describe the components with the aid of behaviour and identification of neurons. Electrophysiology has shown that numerous parallel pathways and superimposed maps, with both local and global mechanisms, are essential for the visual processing. Groups of neurons characteristically have partially overlapping fields when plotted in the dimension where they make discriminations. For example, neurons responding to motion of edges have large fields which partially overlap when the response is plotted against angular velocity, position of nodes of expansion, or inclination of moving edges. This exploitation of field overlap is one of m any lessons to be learned. The advantages of 360° vision, and the ancient success story of implementing it with a com pound eye, are related to the stabilization against rotation and the smooth control of locomotion. Form vision and colour vision are then improved at places on the eye where relative motion is reduced, but the 360° vision makes local gain control essential. The analysis of honeybee vision demonstrates that local spatial resolution is excellent but the spatial layout of a pattern, region by region, is not so well discriminated. This result suggests that responses of numerous local templates are lumped together to give an aggregate quality of each local region, just as ratios of responses of different receptor types make colour discrimination independent of pattern intensity and rate of flicker. Extending this ratio mechanism to form vision works in the opposite direction from the combinatorial explosion of exact spatial distributions of pixels in pattern analysis. Insect vision works with groups of neurons which respond simultaneously, and this activity of neuron groups is able to fit the combination of visual features, although each neuron alone is inadequate for a discrimination.