Research Of Handwritten Digital Recognition Based On Quantum Neural Networks

The handwritten digital recognition is an important research problem of pattern recognition and image processing fields, and the study works have great society and economy benefits. How to improve the recognition rate and recognition reliability is a research hotspot. The traditional recognition methods are always based on the extraction and analysis of character contour feature, but the methods only depend on the image processing widely exist low recognition rate problem. As a new method of pattern recognition, the artificial neural network has some unique virtues compared with traditional methods: perfect fault tolerance ability, powerful classification ability, parallel management and self-learning ability.In the paper, we apply the Quantum Neural Network (QNN) based on the multi-level transfer function to study the handwritten digital recognition problem. QNN is a kind of neuro-fuzzy system by merging neural modeling with fuzzy-theoretic concepts, because of inherently fuzzy characters, QNN can represent and quantify the uncertainty inherent in the training data and encode the sample information into discrete levels of certainty or uncertainty, thereby reduce the uncertainty of pattern recognition and improve the veracity of pattern recognition.A pattern recognition algorithm based on the multi-level transfer function QNN is presented for the pattern recognition with overlapping classes. By theory analysis and experiments, the results show that comparing with the BP network, the QNN network not only can overcome BP network's limitations, but also has excellent classification ability in pattern recognition with uncertainty and overlapping classes.Aiming at the problem that the confused numeral pairs influence the recognition rate in the handwritten digital recognition processing, a handwritten digital recognition system: Multi-layer quantum neural network Recognition System (MLQNN) based on multi-level transfer functions QNN and multi-layer classifiers is proposed. Handwritten digital recognition experiments are performed by using data from MNIST database. Experiment results indicate the proposed MLQNN recognition system can achieves excellent performance in terms of recognition rates and recognition reliability.In additions, we apply the digital samples from the MNIST database to train and test, and there are binarization, segmentation and normalization image preprocessing steps; we use pixel features as direct input of recognition system.