Study On RMB Number Recognition Based On Genetic Algorithm Evolution Of Neural Network

1.IntroductionOptical Character Recognition(OCR) means such as by scanning the optical input to a variety of notes,bills,newspapers,periodicals,manuscripts or other printed image into the text information,and then use technical language to identify the image information into a computer that can try to import skilled.It is an important branch of pattern recognition.In the financial sector,RMB is my legal tender currency notes each have a serial number.Therefore the serial number of the logo as a note.However,for character recognition in different sectors have different needs,the industry must be the development and promotion of industry-oriented.At the rapid development of China's financial industry today,how to manage the flow of the RMB are questions to be settled urgently.For specific industries,according to their characteristics, for the construction of a complete set of character recognition solutions,is the only way to create a corresponding economic benefits and social benefits.Neural network pattern recognition approach is emerging in recent years the field of pattern recognition of a new research direction.Artificial neural network is composed of many highly connected posed by artificial neural,nerve tissue to mimic biology.Neural network has the ability to study non-linear function,by training the weights and threshold updates.As the network to run parallel processing,each sample is studying and computing at the same time,there is some error will not affect the network's output,so,the neural network has excellent fault tolerance.Therefore,it is widely used in pattern recognition,classification, function approximation and so on.However,the study of artificial neural networks are used in gradient descent method,there is the inevitable slow convergence and easy to fall into local minimum points disadvantage. Genetic algorithm is a randomized algorithm,it is based on the evolution of rules of survival of the fittest,the species included operation based on genetic,and continuously generate new species and the species is being optimized,at the same time to the overall approach to parallel heuristic search optimization search of the best individual stocks,in order to achieve the optimal solution to meet the requirements.Genetic algorithms have good global search capabilities,information processing implied parallelism,robustness,scale, and so on can be characterized by fine.This shows that genetic algorithms are good at global search element and neural network in the local search is more effective,so the artificial neural network and genetic algorithm hybrid combining training,set up the genetic algotithms evolution of artificial neural network identification number of RMB is a feasible path.2.Research ContentThis algorithm is based on image processing and pattem recognition theory,and its research purpose is to recognize RMB number.The algorithm describes the while process from image sampling to recognition results.GA improves ANN for RMB number recognition problem based on ANN and GA basic theory.The main research contents are as follows:(1) Introduction and implementation of Digital image processing methods.This paper introduces sampling,gray value,binary method and the edge detection method in detail. RMB during the circulation are inductivded,then median filter is chosen to filter out RMB number image noise through experiments.(2) Image skew correction,characters segmentation and normalization are studied. Rotation method corrects skew image.Vertical projection method separates the RMB number image from the background image.Bilinear interpolation method normalized characters sizes.The method based on characters dot matrix outside frame normalized characters positions.(3) BP NN principle,Artificial Neuron Model,Artificial Neural Network Topology rules and learning rules are studied.BP NN model and learning method are analyzed. Mathematical Derivation of BP algorithm is discussed in detail,and this paper does further researches on improvement of BP algorithm shortcomings.(4) Principle of genetic algorithm is studied.Process of genetic algorithm is analyzed.Genetic operators of selection,crossover and mutation are introduced and analyzed.Solving process of genetic algorithm application is also introduced and analyzed, and the application of genetic algorithm in ANN is introduced.(5) Based on the genetic evolution of BP network model of the hybrid training algorithm,for BP network deficiencies,the algorithm is a fully integrated global search genetic algorithm and BP characteristics of a good local search algorithm performance.It is divided into two parts:BP network part and the part of the genetic algorithm.BP network components:first of all,note the number in accordance with the characteristics of the network to determine the structure of hidden layer nodes,as well as the activation function,and the design of the letters and numbers on the two sub-networks and encoded each character.Genetic algorithm components:According to BP network structure of BP network weights and thresholds for real-coded,defined the fitness function,the roulette wheel selection,using the optimal preservation strategy,adaptive crossover,adaptive mutation and other operations,search to the global optimum individual,be decoded as a BP neural network's initial weights and thresholds.(6) GA-BP network training.In this paper,the GA-BP network genetic algorithm neural network initial weighwork to identify the initial weights and thresholds to improve the network convergence speed and stability.The simulation results show that:the algorithm is compared with the original BP algorithm,the network convergence rate stability and convergence has been greatly improved.In the steepest descent method using the BP algorithm,GA-BP network model than the BP network model improve on the convergence rate of 22.57%;GA-BP network model for the convergence of 91 percent success rate,and the convergence of BP network a success rate of 84%fully GA-BP network in the stability of the convergence speed and convergence on the network should be superior to BP. Convergence stability,GA-BP network training convergence success rate of 91%;and 84% for the BP network.(7) GA-BP algorithm will be used to identify the number of RMB.The use of GA for BP network weights and thresholds to optimize the future will be decoded network weights and thresholds used as the initial BP network weights and thresholds.Then,you can go directly through a set of training sample for study and training.And test the new picture in the known network simulation,And compare with the original BP network without optimizing of genetic algorithm.The simulation results show that genetic algorithm optimization by BP network was significantly less than the original BP network on identification accuracy increased to 95%recognition rate.3.ConclusionIn this paper,the design of the genetic evolution of the BP neural network identification RMB number has a more direct identification with BP networks with higher recognition performance,the network convergence rate and the stability has all increased The overall recognition rate of 95%can be achieved.