Back to Search
Start Over
Developmental Network: An Internal Emergent Object Feature Learning.
- Source :
- Neural Processing Letters; Oct2018, Vol. 48 Issue 2, p1135-1159, 25p
- Publication Year :
- 2018
-
Abstract
- Face recognition has great theory research and application value. It is a very complicated problem which often suffers from variations in lighting condition, facial expression, head pose, glasses, background, and so on. This paper realizes an effective recognition for 108 face images of 27 individuals in complex background, through a biological inspired emergent developmental network (DN). To decrease the influence of complex background on the recognition of the foreground object, another biological inspired mechanism—synapse maintenance, which can dynamically determine which synapse should be removed, weaken or strengthened, is introduced to enhance the image recognition rate. To prevent the quick decay of the learning rate with the increasement of the neuron age, simulating the learning principle of the human brain, a new learning rate is proposed to determine the neuron learning process. Moreover, to exploit the network resource efficiently, neuron regenesis mechanism is designed to regulate the neuron resource dynamically. First, we design two kinds of neuron states to depict the neuron action, then, simulating the work mechanism of the human brain to produce new neurons continuously to learn new knowledge, we design the neuron regenesis mechanism to activate the suppressed old neurons in the developmental network to regenerate and learn new feature, thus to enhance the network usage efficiency. In order to demonstrate the effect of DN on face recognition, we compare and analyze the performances of DN with/without synapse maintenance mechanism, with the neuron regenesis mechanism. Experiment results in semi-constrained dataset and unconstrained dataset illustrate that DN with the synapse maintenance and neuron regenesis mechanism can effectively improve the face recognition rate in complex background. Further, comparing the performance of DN with some state-of-the-art algorithms, experimental results demonstrate the superior performance of the proposed method. [ABSTRACT FROM AUTHOR]
- Subjects :
- NEURAL circuitry
ELECTROPHYSIOLOGY
BIOLOGICAL neural networks
LEARNING
SYNAPSES
Subjects
Details
- Language :
- English
- ISSN :
- 13704621
- Volume :
- 48
- Issue :
- 2
- Database :
- Complementary Index
- Journal :
- Neural Processing Letters
- Publication Type :
- Academic Journal
- Accession number :
- 132161636
- Full Text :
- https://doi.org/10.1007/s11063-017-9734-z