Multiple-Classifier Fusion Using Spatial Features for Partially Occluded Handwritten Digit Recognition

The subject of “handwritten digit recognition” is a great concern and has many applications in various fields. Although highly restricted forms of digit recognition are widely utilized, reading incomplete and occluded digit image is still a challenge for both academia and industries. In this paper, we attack the problems of recognizing occluded handwritten digits by finding the influence of small patches in digit images to the recognition results. We apply one-hidden-layer neural networks to train and validate each patch independently and enhance the performance of each classifier by the results of its correlated patches. This method allows us to restrict the effect of false information solely into areas that small patches lay on and then correct recognition results by their neighbors. The result of the proposed method shows a noticeable improvement in the stable ability of recognition model with different kinds of simulated distortions.