Handwritten recognition using neural networks based on multiple feature extraction algorithms

Abstract EN

Handwritten letters are vague in nature as there may not always be sharp perfectly straight lines, and curves not necessarily be clearly, unlike the printed letters.

Furthermore, letter can be drawn in different sizes.

Therefore, the proposed used of four methods based on two algorithms for feature extraction.

The first algorithm discrete wavelet transform and second algorithm is Radon transform, by experiments was found that the feature extraction using radon transform is more precise than wavelet transform.

The proposed technique computes Radon projections in different orientations and captures the directional features of letter images.

The obtain recognition rate was 93 % when used for train artificial neural networks.

Comparative between four methods to choose the best has been done.

Training method is based on Back-Propagation learning algorithm used by feed forward Neural Network.