Search

A new formation of a neural network called a Density Based Random Vector Functional Link Network (DBRVFLN) is introduced to solve high dimensional real-world problems. It is a hybrid technique which uses the combination of a priori knowledge of the problem and randomness to prepare unknown factors. Simple but powerful feature extraction methods for handwritten digit recognition and script recognition are introduced. Handwritten digit recognition systems using neural networks are introduced. For the script recognition task, a global approach which uses whole sets of features of the image and an analytical approach from nonsegmented sequence of features via letters to a word using neural networks are designed and explored. The recognition systems are based on conventional preprocessing methodologies, novel feature extraction and reduction algorithms and quadratic approaches of neural networks such as Radial Basis Function Neural Network(RBNN) and DBRVFLN. The recognition systems are tested using unconstrained real-world databases. In the handwritten digit recognition task, the performance of the recognition system using DBRVFLN is better than that of RBNN if there is enough priori knowledge. To attain 1% substitution error rates, the current recognition system needs to tolerate rejection rat es of about 11%∼12%. In the global approach of script recognition task, the ability to construct a filter for one word is tested. While keeping a very low substitution error rate of 0.74%, the recognition system which uses DBRVFLN rejects 11.48% and has better performance with 2.78% in the rejection ratio than the system using RBNN even if they have the same number of enhancement nodes. The experimental results show that the random vector enhancements for the unknown factor in DBRVFLN act very nicely as decision enhancers and that they do improve the classification performances in comparison with RBNN, in the same recognition system