Extreme Learning Machine Applied In Handwritten Numeral Recognition

Handwritten numeral recognition means recognizing the handwritten digits bycomputer automatically, which belongs to one of the research areas of machinevision.Considering the global generic of Arabic numerals and a wide range ofapplications of the handwritten digital, handwritten numeral recognition has been theconcern of many experts and scholars. But different people in different regions havedifferent writing habits, which leads to causing many handwritten digits variants, andadds to recognition difficulty. Handwritten numeral recognition technology is still animportant research direction in machine vision.Neural network classifier which has high tolerance to the noice data has beenwidely used. However, traditional neural network has many problems such as longtraining time, easy to fall into local optimum problem and so on. The classifier trainedby traditional neural network is not ideal. Extreme Learning Machine networkparameters given randomly and its training process does not require iterativeadjustment. Finally, extreme Learning Machine network’s training speed is very fastbut will not fall into local optima. It has been proven its superiority in terms of patternrecognition. This paper focuses on the advantages and disadvantages of the classifierby the experiments. For the insufficient of Extreme Learning Machine NetworkClassifier at generalization and stability, this paper presents a particle swarm ExtremeLearning Machine Network, and compares the difference between the two networksby the experiments.The key technology of handwritten numeral recognition includes featureextraction and pattern recognition. For this two aspects, many work have been doneas follows in this paper.1. For capturing the useful information to the maximum limit, so that we caneasily get feature extraction, in this paper, a handwritten digital image is preprocessed:gray, binary expansion corrosion, calibration and normalization.2. In a comprehensive study of handwritten digits mesh and structural features, in this paper, handwritten digits are extracted contour characteristics, Fourier transformfeatures, invariant moment characteristics, thirteen points of the grid characteristicsand so on. Considering the nimiety of dimensions of the original features, which areapplied directly to the classifier training， which will cause that the classifier trainingspeed is too slow and the recognition accuracy is not high. So, in this paper, a featuredimension reduction methods-principal component analysis is applied to dimensionreduction for the original features.3. A particle swarm algorithm for extreme learning machine is proposed. Ahandwritten numeral recognition classifier is established by the characteristics above.Compared with the traditional classification by testing on the NIST database, we canfind that the algorithm has better performance applied to handwritten digitrecognition.4. A handwritten numeral recognition system is established by Matlab GUI, sowe can capture handwritten digital image and use the classifier established to identify.By the system, we can verify the correctness of the algorithm which can providesguidance for the recognition of handwritten digits.