Abstract: Handwritten character recognition(HCR) is an important branch of image recognition,which recognizes the handwritten characters with the data mining and machine learning technologies.Currently,the HCR methods mainly focus on the improvements of different deep learning models,where the multiple-layer extreme learning machine(ML-ELM) has attracted the wide attention from the academia and industry due to its faster training speed and better recognition performance than deep belief net(DBN) and deep Boltzmann machine(DBM).However,the recognition performance of ML-ELM is severely influenced by the random weights when determining the input weights for each hidden-layer.This paper first proposes a decomposition ELM(DE-ELM) which is a shallow ELM training scheme based on the hidden-layer output matrix decomposition and then applies DE-ELM to deal with HCR problems,i.e.,handwritten digits in MNIST,handwritten digits and English letters in EMNIST,handwritten Japanese characters in KMNIST and K49-MNIST.In comparison with ML-ELM,DE-ELM reduces the randomness of ELM-based HCR model.Meanwhile,DE-ELM can obtain higher recognition accuracy than ML-ELM with the same training time and faster training speed than ML-ELM with the equal recognition accuracy.Experimental results demonstrate the feasibility and effectiveness of the proposed DE-ELM when dealing with HCR problems.

Key words: Handwritten character recognition, Extreme learning machine, Multiple layer extreme learning machine, Deep lear-ning, Feature extraction