BIOLOGICALLY INSPIRED EDGE DETECTION USING

SPIKING NEURAL NETWORKS AND HEXAGONAL IMAGES

Marine Clogenson

, Dermot Kerr

Martin McGinnity

Sonya Coleman

and Qingxiang Wu

CPE Lyon, Domaine Scientifique de la Doua, BP 82077, 69616, Villeurbanne, France

Intelligent Systems Research Centre, University of Ulster, Magee, Derry, BT48 7JL, U.K.

Keywords: Spiking neural network, Edge detection, Multi-scale hexagonal receptive fields.

Abstract: Inspired by the structure and behaviour of the human visual system, we extend existing work using spiking

neural networks for edge detection with a biologically plausible hexagonal pixel arrangement. Standard

digital images are converted into a hexagonal pixel representation before being processed with a spiking

neural network with scalable hexagonally shaped receptive fields. The performance is compared with

different sized receptive fields implemented on standard rectangular images. Results illustrate that using

hexagonal-shaped receptive fields provides improved performance over a range of scales compared with

standard rectangular shaped receptive fields and images.

1 INTRODUCTION

The human vision system (HVS) processes a visual

scene starting in the retina where light intensity is

converted into nerve signals within the

photoreceptors. The signals are pre-processed and

propagated through the various layers within the

retina. The resulting spike train output from the

retinal ganglion cells travels along the optic nerve

for further processing in the lateral geniculate

nucleus and visual cortex. The powerful

performance of the HVS is achieved through

massive parallel processing using neurons and their

complex interconnections formed by synapses.

Taking inspiration from the HVS research has tried

to improve image processing techniques, via the use

of neural networks (NNs) (Egmont-Petersen et al.,

2002). Spiking neural networks (SNNs) are a class

of NNs that mimic more accurately the biological

information processing in the brain. Using a

temporal coding scheme SNNs improve upon NNs

increasing computational power, speed and therefore

enabling real-time processing (Kunkle and

Merrigan, 2002). SNNs use simple neuronal models

and communicate using spikes in a manner similar

to action potentials found in biological neurons. In

(Wu et al., 2007); (Meftah et al., 2008); (Buhmann

et al., 2005) SNN approaches have been developed

for image segmentation. In (Escobar et al., 2009) a

SNN is used to model two areas of the brain

concerned with motion with the aim of action

recognition. A SNN model is proposed in (Meftah et

al., 2010) that performs segmentation and edge

detection, however images must first be segmented

before the edge detection stage can be performed. A

SNN is proposed in (Hugues et al., 2002) to detect

contours in images through the synchronisation of

integrate and fire neurons using simple synthetic

images. In (Wu et al. 2007) a SNN is proposed for

real-time edge detection. In (Chevallier et al., 2006)

a distributed SNN is proposed for extracting

saliencies in an image and in (Chevallier and

Dahdouh, 2009) a SNN is used to perform

Difference of Gaussian filtering. In (Delorme and

Thorpe, 2003) a SNN is proposed that uses a rank

order coding scheme.

In computer vision the apparent strength of a

feature in an image may depend on the scale at

which the appropriate feature detection operator is

applied and many standard approaches to multi-scale

feature detection have been developed (Lindeberg,

1994). However, scalable feature extraction

algorithms using bio-inspired approaches have been

researched and developed to a much lesser extent

with the exception of the approach by (Gao et al.,

2006) where a NN that simulates the multi-scale

receptive field of the biological vision is proposed.

In this paper we present a novel approach to feature

extraction using scalable receptive fields.

381

Clogenson M., Kerr D., McGinnity M., Coleman S. and Wu Q..

BIOLOGICALLY INSPIRED EDGE DETECTION USING SPIKING NEURAL NETWORKS AND HEXAGONAL IMAGES.

DOI: 10.5220/0003682103810384

In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2011), pages 381-384

ISBN: 978-989-8425-84-3

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Within the fovea the cone photoreceptors are

tightly packed into a hexagonal lattice, resulting in

photoreceptors that are generally surrounded by 6-

neighbours (small irregularities occasionally exist).

Most existing methods for feature extraction are

based on rectangular lattices. Allen (2003) has

shown that curved structures are not well

represented on a rectangular lattice leading us to

question why we use them when nature has chosen a

hexagonal lattice for photoreceptors? Using an

artificial hexagonal sampling lattice, both spatial and

spectral advantages may be derived: namely,

equidistance of all pixel neighbours and improved

spatial isotropy of spectral response. Pixel spatial

equidistance facilitates the implementation of

circular symmetric kernels that are associated with

an increase in accuracy when detecting straight and

curved edges (Allen, 2003). Better spatial sampling

efficiency is achieved by the hexagonal structure

compared with a rectangular grid of similar pixel

separation, leading to improved computational

performance. A hexagonal grid with unit separation

of pixel centres has approximately 13% fewer pixels

than the same image resolution on a rectangular grid

with unit horizontal and vertical separation of pixel

centres (Vitulli, 2002).

In this paper we present a biologically inspired

approach to feature extraction using spiking neural

networks, hexagonal pixel-based images that mimic

the hexagonal arrangement found in the retina, and

scalable hexagonally arranged receptive fields.

2 MODEL IMPLEMENTATION

We use a method proposed in (Middleton and

Sivaswamy, 2001) to create hexagonal pixels (and

images) from clusters of sub-pixels which limits the

loss of image resolution whilst complying with the

main hexagonal properties (Middleton and

Sivaswamy, 2001). The hexagonal pixel is

comprised of 56 sub-pixels closely representing the

shape of a single hexagonal pixel, thus enabling us

to mimic the hexagonal structure used by the HVS

for image capture.

We define our spiking neural network structure

as illustrated in Figure 1. Suppose that the first layer

represents photoreceptors where each pixel in the

hexagonal image corresponds to a photoreceptor. A

receptive field is where a spiking neuron integrates

the spikes from a group of afferent neurons, and in

our model this intermediate layer is composed of

four types of neurons corresponding to four different

receptive fields respectively.

Each of the four parallel arrays of neurons in the

intermediate layer are the same dimension as the

receptor layer with only one neuron in each array

illustrated in Figure 1 for simplicity. Each of these

layers performs the processing for different edge

directions and is connected to the receptor layer by

differing weight matrices. These weight matrices can

be of varying sizes to represent the width of the

receptive field under consideration. We use the

conductance-based integrate-and-fire spiking neuron

model as this offers biological realism whilst

providing a reduction in computational complexity

(Izhikevich, 2004). ‘X’ in the synapse connections

represents an excitatory synapse and ‘Δ’ represents

an inhibitory synapse. Each neuron in the output

layer integrates four corresponding outputs from

intermediate neurons. The firing rate map of the

output layer forms an edge graphic corresponding to

the input image. The receptive fields illustrated in

the intermediate layer in Figure 1 can be receptive

fields of any size and we will use 7, 19, and 37-point

hexagonal receptive fields.

The network model was implemented in Matlab

using the network parameters found in (Wu, et al.,

2007) that are consistent with biological neurons

(Masland, 2001). Synapse strengths can be adjusted

to ensure that the neuron does not fire in response to

a uniform image within its receptive field.

3 RESULTS AND EVALUATION

We present edge detection results at three different

scales using 7- 19- and 37-point sized hexagonal

receptive fields (denoted as HSNN-7, HSNN-19 and

HSNN-37). For comparison we also present results

using the SNN approach in (Wu et al., 2007) which

uses a standard rectangular structure (denoted as

SNN). In Figure 2 we present the edge maps

generated by using the well known Lena image as

input to the SNN models. The edge brightness

increases as the firing rate of the neuron becomes

stronger, thus the firing rate may be set as a

threshold to determine the presence or absence of an

edge. The output from the HSSN is much clearer

and less noisy than the corresponding output from

the SNN.

We evaluate the performance of both the HSNN

and SNN approaches using the Figure of Merit

(FOM) technique (Abdou and Pratt, 1979) in Figure

3. The FOM is compared over a range of noise

levels. Figure 3 illustrates that the HSNN shows

improved performance over the SNN for all edge

types, in particular in areas of high noise and this is

NCTA 2011 - International Conference on Neural Computation Theory and Applications

382

Figure 1: Spiking Neural Network Structure.

(a) SNN (Wu et al., 2007). (b) 7-Point HSNN (c) 19-Point HSNN (d) 37-Point HSNN

Figure 2: Example Edge Detection Outputs.

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 5 10 20 50 100 No

noiseSNR

FoM

HSNN‐7

HSNN‐19

HSNN‐37

SNN

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 5 10 20 50 100 No

noise

SNR

FoM

HSNN‐7

HSNN‐19

HSNN‐37

SNN

(a) Vertical edge (b) Diagonal edge

Figure 3: Figure of Merit result.

becomes more evident as the receptive field size

increases. The simulation is run and spikes are

computed over a time interval of 100ms. Table 1

compares the time to run this simulation and

illustrates an improvement in computation time with

the hexagonal arrangement.

BIOLOGICALLY INSPIRED EDGE DETECTION USING SPIKING NEURAL NETWORKS AND HEXAGONAL

IMAGES

383

Table 1: Algorithm run times (seconds).

RF Size Processing time

SNN 3.92

HSNN 7-Point 3.16

HSNN 19-Point 3.47

HSNN 37-Point 3.78

4 DISCUSSION AND FUTURE

WORK

We present a biologically inspired approach to

feature detection that is mimics the human visual

system. The presented SNN is constructed by a

hierarchical structure that is composed of spiking

neurons with various receptive fields. The input

image has a hexagonal pixel arrangement and

correspondingly the receptive fields used are also

arranged in a hexagonal structure. The spiking

neuron models provide powerful functionality for

integration of inputs and generation of spikes.

Synapses are able to perform different complicated

computations. This paper demonstrates how a

spiking neural network can detect edges in an image

using a hexagonal structure over a wide range of

scales and demonstrates performance and

computational improvements over rectangular pixel-

based SNN approaches.

ACKNOWLEDGEMENTS

This work was supported by the Centre of

Excellence in Intelligent Systems project, funded by

InvestNI and the Integrated Development Fund.

REFERENCES

Abdou, I. E. and Pratt, W. K., (1979). Quantitative design

and evaluation of enhancement/thresholding edge

detectors. Proc. of the IEEE, 67(5) pp. 753-763.

Allen J. D., (2003). Filter Banks for Images on Hexagonal

Grid. Signal Solutions.

Buhmann, J. M., Lange, T. M and Ramacher, U., (2005).

Image segmentation by networks of spiking neurons.

Neural computation, 17, 5, 1010-1031.

Chevallier, S., Tarroux, P. and Paugam-Moisy, H., (2006).

Saliency extraction with a distributed spiking neural

network. Proc. of ESANN, 209-214.

Chevallier, S. and Dahdouh, S., (2009). Difference of

Gaussians Type Neural Image Filtering with Spiking

Neurons”. Proc. of Int. Conf. on Neural Compuation,

Maderia, Portugal.

Delorme, A. and Thorpe, S. J., (2003). SpikeNET: an

event-driven simulation package for modelling large

networks of spiking neurons. Computation in Neural

Systems,14(4), 613-627.

Egmont-Petersen, M., De Ridder, D. and Handels, H.,

(2002). Image processing with neural networks-a

review. Pattern Recognition,35(10), 2279-2301.

Escobar, M. J., Masson, G. S., Vieville, T. and

Kornprobst, P., (2009). Action Recognition Using a

Bio-Inspired Feedforward Spiking Network. IJCV,

82(3), 284-301.

Gao Y., Zhang L., Wang Y. and Jiang J., (2006). Multi-

scale receptive field neural networks for object

tracking. International Congress Series, 1291, 229–

232.

Hugues, E., Guilleux, F. and Rochel, O., (2002). Contour

Detection by synchronization of Integrate and Fire

Neurons. LNCS. 2525, 60-69.

Izhikevich, E. M., (2004). Which model to use for cortical

spiking neurons?. IEEE Trans. on NN, Vol. 15, No. 5.

Kunkle, D. R. and Merrigan, C., (2002). Pulsed neural

networks and their application. Computer Science

Dept., College of Computing and Information

Sciences, Rochester Institute of Technology.

Lindeberg, T., (1994). Scale-Space Theory in Computer

Vision. Kluwer, Academic Publisher/Springer,

Dordrecht, Netherlands.

Masland, R. H., (2001). The fundamental plan of the

retina. Nature Neuroscience, vol. 4, pp. 877-886.

Meftah, B., Lezoray, O. and Benyettou, A., (2010).

Segmentation and Edge Detection based on Spiking

Neural Network Model. Neural Processing Letters.

32(2), 131-146.

Middleton L. and Sivaswamy, J., (2001). Edge Detection

in a Hexagonal-Image Processing Framework. Image

and Vision Computing.19, pp. 1071-1081.

Paugam-Moisy, H. and Bohte, S. M., (2009). Computing

with spiking neuron networks. Handbook of Natural

Computing. 40. Springer, Heidelberg.

Vitulli R., (2002). Aliasing Effects Mitigation by

Optimized Sampling Grids and Impact on Image

Acquisition Chains. Geoscience and Remote Sensing

Symposium, pp. 979-981.

Wu, Q. X., McGinnity, M., Maguire, L., Belatreche, A.

and Glackin, B., (2007). Edge Detection Based on

Spiking Neural Network Model.

Advanced Intelligent

Computing Theories and Applications. With Aspects

of Artificial Intelligence, 26-34.

NCTA 2011 - International Conference on Neural Computation Theory and Applications

384