Enhancement of parallel thinning algorithm for handwritten characters using neural network

Abstract

Thinning is a well known pre-processing step in many image analysis techniques and it has been applied in a wide variety of applications in the fields of pattern recognition and machine vision. Thinning can be applied onto various images to produce one-pixel width skeletons that represent a good abstraction of the image shape. This project aims to improve a parallel thinning algorithm that satisfies two fundamental requirements of thinning, namely the processing speed and the quality of skeletons. This work is an attempt to devise an effective thinning method by applying a neural network model, used for handwritten characters. Parker’s parallel thinning algorithm is chosen in this project because it satisfies the requirements of producing good quality skeletons although it does not provide an efficient processing speed. Thus, this project presents a framework for the implementation of a multilayer perceptron neural network with backpropagation algorithm, in the Parker thinning algorithm in order to produce a fast fully parallel thinning algorithm. The proposed thinning algorithm is tested on a set of binary images of isolated handwritten characters, obtained from CEDAR and MNIST databases. Comparison is conducted on the proposed thinning algorithm, Zhang- Suen thinning algorithm and Parker thinning algorithm. The analysis of these three thinning algorithms is based on two parameters, namely topological analysis and implementation speed analysis. The experimental results show that the proposed thinning algorithm produced acceptable, one-pixel width skeletons that preserve connectivity. It also works faster than Parker thinning algorithm but it cannot deals with the necking problem. The skeletons produced also retain the general abstraction of the global shape of the handwritten characters images.