Abstract
This paper describes the Cellular weighted Graph Neuron (CwGN) scheme, which demonstrates real-time transformation-invariant pattern recognition for resource constraint networks such as wireless sensor networks (WSN). This scheme can recognise patterns from a variety of perspectives in a fully distributed manner, which makes it highly responsive, scalable, and energy efficient. Learning operations are decoupled and reduced to relatively simple processing elements to achieve an economical parallelism. We show that by using only local computation, i.e., between adjacent nodes, we can fully describe patterns by their boundaries and then recognise the transformations of these patterns. Our scheme solves the invariant pattern recognition for the first time in a fully distributed manner within a predictable single cycle learning duration which allows the use of the scheme within resource-constrained networks such as wireless sensor networks for event-detection. Theoretical analysis and experimental results demonstrate online recognition of transformed patterns with competitive accuracy and a low energy budget.