TY - JOUR
T1 - LEGION-based image segmentation by means of spiking neural networks using normalized synaptic weights implemented on a compact scalable neuromorphic architecture
AU - Sanchez, Giovanny
AU - Madrenas, Jordi
AU - Cosp-Vilella, Jordi
N1 - Publisher Copyright:
© 2019 Elsevier B.V.
PY - 2019/8/4
Y1 - 2019/8/4
N2 - LEGION (Locally Excitatory, Globally Inhibitory Oscillator Network) topology has demonstrated good capabilities in scene segmentation applications. However, the implementation of LEGION algorithm requires machines with high performance to process a set of complex differential equations limiting its use in practical real-time applications. Recently, several authors have proposed alternative methods based on spiking neural networks (SNN) to create oscillatory neural networks with low computational complexity and highly feasible to be implemented on digital hardware to perform adaptive segmentation of images. Nevertheless, existing SNN with LEGION configuration focus on the membrane model leaving aside the behavior of the synapses although they play an important role in the synchronization of several segments by self-adapting their weights. In this work, we propose a SNN-LEGION configuration along with normalized weight of the synapses to self-adapt the SNN network to synchronize several segments of any size and shape at the same time. The proposed SNN-LEGION method involves a global inhibitor, which is in charge of performing the segmentation process between different objects with different sizes and shapes on time. To validate the proposal, the SNN-LEGION method is implemented on an optimized scalable neuromorphic architecture. Our preliminary results demonstrate that the proposed normalization process of the synaptic weights along with the SNN-LEGION configuration keep the capacity of the LEGION network to separate the segments on time, which can be useful in video processing applications such as vision processing systems for mobile robots, offering lower computational complexity and area consumption compared with previously reported solutions.
AB - LEGION (Locally Excitatory, Globally Inhibitory Oscillator Network) topology has demonstrated good capabilities in scene segmentation applications. However, the implementation of LEGION algorithm requires machines with high performance to process a set of complex differential equations limiting its use in practical real-time applications. Recently, several authors have proposed alternative methods based on spiking neural networks (SNN) to create oscillatory neural networks with low computational complexity and highly feasible to be implemented on digital hardware to perform adaptive segmentation of images. Nevertheless, existing SNN with LEGION configuration focus on the membrane model leaving aside the behavior of the synapses although they play an important role in the synchronization of several segments by self-adapting their weights. In this work, we propose a SNN-LEGION configuration along with normalized weight of the synapses to self-adapt the SNN network to synchronize several segments of any size and shape at the same time. The proposed SNN-LEGION method involves a global inhibitor, which is in charge of performing the segmentation process between different objects with different sizes and shapes on time. To validate the proposal, the SNN-LEGION method is implemented on an optimized scalable neuromorphic architecture. Our preliminary results demonstrate that the proposed normalization process of the synaptic weights along with the SNN-LEGION configuration keep the capacity of the LEGION network to separate the segments on time, which can be useful in video processing applications such as vision processing systems for mobile robots, offering lower computational complexity and area consumption compared with previously reported solutions.
KW - FPGA
KW - LEGION
KW - SIMD architecture
KW - Spiking neural networks
UR - http://www.scopus.com/inward/record.url?scp=85065418438&partnerID=8YFLogxK
U2 - 10.1016/j.neucom.2019.04.037
DO - 10.1016/j.neucom.2019.04.037
M3 - Artículo
AN - SCOPUS:85065418438
SN - 0925-2312
VL - 352
SP - 106
EP - 120
JO - Neurocomputing
JF - Neurocomputing
ER -