Dissertation > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device

A Multistage Neural Network System for Handwritten Numeral Recognition

Author ChenBaiZuo
Tutor JiaHuaDing
School Southwestern University of Finance and Economics
Course Applied Computer Technology
Keywords Handwritten numeral recognition Feature Extraction BP neural network Classifier Fusion
CLC TP391.41
Type Master's thesis
Year 2008
Downloads 89
Quotes 1
Download Dissertation

Handwritten numeral recognition is a technology machine or computer to automatically recognize handwritten Arabic numerals, a branch of the optical character recognition technology. Due to the the Arabic numerals world versatility, and the identification and handling of digital are often the core of some of the automation system and key, so the study of handwritten numeral recognition versatility, and is of great significance. This paper presents a multi-stage neural network-based integrated simulation system of handwritten numeral recognition handwritten character recognition technology to investigate and research-oriented offline handwritten isolated numeral recognition. The entire system consists of four modules: the first image pre-processing module and the character feature extraction module as the basis of the pattern data input; Secondly, the number of BP neural network as to identify the core classifier, that classification module; then these neural classifier the output of the fusion decision to obtain a final recognition result, i.e. a decision module. The main research work focused on the identification of feature extraction and fusion of multiple classifiers. In the experimental analysis of the development potential of a variety of feature extraction methods in feature extraction based on the characteristics of a global and two local features as the main feature, and the combination of a series of additional features of the low-dimensional combination method. The experiments show that the combination of three groups of eigenvectors handwritten digits with high classification ability. The fusion of multiple classifiers fusion method proposed in this paper and other fusion experimental analysis and comparison. The experiments show that the use of multi-classifier fusion strategy can achieve high-precision recognition, and the fusion method proposed in this paper is slightly better than the other fusion methods. The first chapter describes the application prospects of the handwritten numeral recognition, current research and research methods, and describes the common methods of pattern recognition, illustrates the difficulty of the handwritten numeral recognition and its broad application prospects. The second chapter describes the pretreatment technology handwritten digit recognition, image smoothing, binarization, image standardization, refinement technology, this paper, refined processing algorithms. The third chapter describes the feature extraction technology, and feature extraction algorithm used in the identification system of this article, and experimental analysis to select the best combinations of features. The fourth chapter describes the principles and neural network algorithm, and pointed out that the internal mechanism of the neural network for handwritten numeral recognition and unique advantages. Finally, the simulation of this paper, the digital identification system on MNIST handwritten digital image library were three single classifier to identify experimental and multi-level classifier fusion experiments. The experiments show that the best single classifier recognition rate of 98.14%, while multi-classifier fusion recognition performance than a single classifier systems, the highest recognition rate of 98.47%. Finally, when added to about 4% of the data identification strategies to give a 99.60% recognition accuracy, the error rate was 0.38%, to have practical value.

Related Dissertations
More Dissertations