Research On Stochastic Resonance Of Discrete Hopfield Neural Network Character Recognition

With the rapid development of computer software and the acceleration of financial market,the related research on neural networks has led to the development of all aspects of life.Artificial neural network,as a mathematical model simulating human brain,plays an increasingly important role and provides more and more convenience for human’s life.Its structure,fault-tolerance and nonlinear characteristics attracts the attention of researchers in various fields.It is noted that,with the development of computer software nowadays,the recognition of handwritten numerals is of great significance,and its application will result in huge social and economic benefits.Numerous studies have shown that noise widely exists in the nervous system and can enhance the ability of neurons to process information,namely stochastic resonance phenomenon.This phenomenon reflects that noise can enhance,rather than degrade,the ability of neural networks to perceive external stimuli,which to some extent means an information processing technology that can improve the quality of human life.In this thesis,the random Hopfield neural network for handwritten digits recognition is taken as an example,and the application of Hopfield neural network in image recognition is explored according to the steady-state property of Hopfield dynamics.The recognition scheme is mapping the digit information into the order of state vectors of Hopfield neural network.In the simulation,the handwritten digit corrupted by noise is scanned and transformed into the binary signal that is inputted into the neural network.The Hopfield neural network of a given weight matrix has two stable state vectors,which will evolve into the network output with the help of added noise.Then,the digit information is decoded by the inner product of the network outputs and the stored state vector.Next,the order of state vectors of Hopfield neural network is further mapped to digital images.The experimental results show that the bit error rate of digital image recognition is negatively correlated with the amplitude of modulated signal,the inter-symbol time interval and the number of network neuron coupling.Moreover,with the increase of noise intensity,the phenomenon of aperiodic stochastic resonance occurs.When the non-zero optimal noise intensity reaches the minimum,the recovery of digital image becomes clearer,which is a generalized stochastic resonance effect.The manifestation of phenomena is also an important non-linear phenomenon in the complex network.The results provide experimental basis for further research on adaptive weight matrix of Hopfield neural network,and are of great significancefor the study of the role of random factors(e.g.noise)in the associative memory of neural networks.