Bio-inspired color image segmentation on the GPU (BioSPCIS)

Abstract

In this paper we introduce a neural architecture for multiple scale color image segmentation on a Graphics Processing Unit (GPU): the BioSPCIS (Bio-Inspired Stream Processing Color Image Segmentation) architecture. BioSPCIS has been designed according to the physiological organization of the cells on the mammalian visual system and psychophysical studies about the interaction of these cells for image segmentation. Quality of the segmentation was measured against hand-labelled segmentations from the Berkeley Segmentation Dataset. Using a stream processing model and hardware suitable for its execution, we are able to compute the activity of several neurons in the visual path system simultaneously. All the 100 test images in the Berkeley database can be processed in 5 minutes using this architecture. © 2011 Springer-Verlag Berlin Heidelberg.