Study On Stochastic Resonance Based On Improved FHN Mode And Its Image Processing Applications

The point of view that noise is harmful to the enhancement of signal is overturned bystochastic resonance. The energy of noise is transferred to the signal by nonlinear systems. At thesame time, a growing number of neurophysiological experiments revealed that stochastic resonancein the nervous system may be ubiquitous, and the background noise plays an important role in theflow of neural information. The stochastic resonance in the visual nervous system has become thefocus of the study, and more scholars put their eyes on the applications in signal and imageprocessing. However, these studies focused on the ideal double-sided system or the computationalsimulation model of single neuron, and focused on stochastic resonance method itself, whileignoring the nonlinear system of stochastic resonance. Therefore, this paper put the computingmodel of FitzHugu-Nagnmo(FHN) neural which can fully reflects action potentials as the basic unit.Firstly, the model of FHN neural with self-synaptic properties is proposed, considering theimportant role of self-synaptic structure in the visual perception of the nervous system, and it isapplied to enhance weak signals in strong noise. Second, the model of coupled FHN neuronsisproposed,considering the synapses between neurons, and it is applied to detect weak signals instrong noise to explain the important role of bilateral information flow mechanism in nervoussystem. Finally, the model of FHN neurons based on multiple stochastic resonance mechanismisproposed, simulating vision system’s perception of different levels of detail, and the new model isused to detect colonies images whichcontain multi-contrast edges. The main work and researchresults are as follows(1) The traditional model of FHN neural networks one-sided emphasize the connection mechanismbetween neurons, ignoring the loop between microscopic neurons itself and neural networks,which is called auto-synaptic. Traditional FHN neuron model is improved by simulating theinformation transfer of auto-synaptic structure. The new neural network model is built byauto-synaptic FHN neurons, and it is used to enhance one-dimensional weak signal in strongnoise firstly. The optimal internal noise is found by Gradient-Descent and the result is evaluatedby quantitative indicator. Then it is applied to two-dimensional image signal and taking PSNRas quantitative indicator. The simulation results confirmed that the new model has betterrobustness, higher PSNR, richer details, fewer glitches compared to traditional model(2) Neurons in multiple brain regions are required to work together to complete nervous systemfunction., however traditional series connection and parallel connection are based on one-sideinformation flow, which can’t simulate complex connection between the neurons. Therefore the coupled FHN neuron model is proposed which is based on bilateral information flowmechanism. The output of each neuron is transmitted to another by feedback, which forms theclosed-loop system. Firstly it is used to detect one-dimensional weak signal in strong noise. Andthen it is used to detect two-dimensional image signal in strong noise. The simulation resultsconfirmed that the new model has higher entropy, more edge information,obvious outline andthe edge is more complete, clear, accurate, comparing with the series FHN neuron model andthe parallel FHN neuron model..(3) In fact, the coupled FHN neuron model’s enhance is global, and it is not applicable to detectmulti-level intensity signal. Therefore the new method of weak signal enhancement based onmultiple stochastic resonance mechanism is proposed. Firstly the stronger edge is detected bynoise’s driving, Then put the detect result and original image together, weakening strongeredge’s bad influence on original image. Last the weaker edge is detected by second stochasticresonance mechanism. It is used to detect multi-level intensity signal in strong noise. The newmodel’s robustness is gained by adjusting internal noise and external noise independently.Finally, it is used to detect edge of colony images. The simulation results confirmed that thenew model has higher AUC and entropy, more beautiful ROC, less noise, obvious outline andthe edge is more complete, clear, accurate, comparing with traditional model.