SpikeSEG : Spiking segmentation via STDP saliency mapping

Kirkland, Paul and Di Caterina, Gaetano and Soraghan, John and Matich, George; (2020) SpikeSEG : Spiking segmentation via STDP saliency mapping. In: 2020 International Joint Conference on Neural Networks (IJCNN). IEEE, GBR. ISBN 978-1-7281-6926-2 (https://doi.org/10.1109/IJCNN48605.2020.9207075)

[thumbnail of Kirkland-etal-IJCNN-2020-SpikeSEG-spiking-segmentation-via-STDP]
Preview
Text. Filename: Kirkland_etal_IJCNN_2020_SpikeSEG_spiking_segmentation_via_STDP.pdf
Accepted Author Manuscript

Download (6MB)| Preview

Abstract

Taking inspiration from the structure and behaviourof the human visual system and using the Transposed Convo-lution and Saliency Mapping methods of Convolutional NeuralNetworks (CNN), a spiking event-based image segmentationalgorithm, SpikeSEG is proposed. The approach makes use ofboth spike-based imaging and spike-based processing, where theimages are either standard images converted to spiking images orthey are generated directly from a neuromorphic event drivensensor, and then processed using a spiking fully convolutionalneural network. The spiking segmentation method uses the spikeactivations through time within the network to trace back anyoutputs from saliency maps, to the exact pixel location. Thisnot only gives exact pixel locations for spiking segmentation,but with low latency and computational overhead. SpikeSEGis the first spiking event-based segmentation network and overthree experiment test achieves promising results with 96%accuracy overall and a 74% mean intersection over union forthe segmentation, all within an event by event-based framework.