A Novel Softmax Regression Enhancement for Handwritten Digits Recognition using Tensor Flow Library

Abstract

Background and Aim: Handwritten Digit Recognition has a wide variety of applications in postal mail order, phone records search, automatic car number plate recognition, and in the medical sector that observed how Machine Learning makes the daily tasks simpler and more efficient. This paper aims to improve the classification accuracy of existing handwritten digit systems, thus improve their efficiency. Methodology: The proposed system consists of an enhanced decision function by adding a “Bias Probability” Function. The function adds negative weights to the output classes (0-9) that have a high positive bias and add a positive weight to the output classes that have a high negative bias to neutralize the effect of this high negative bias. Therefore, the Bayesian Classifier function has been enhanced thereby improving the accuracy of classification, which will further improve the performance of the multiclass probability categorization. Result: An increase of 5.6% was observed in the overall accuracy of handwritten digit classification using the Modified National Institute of Standards and Technology (MNIST) Dataset. Conclusion: From the results, it is clear that the proposed system, enhances the main decision function to further improve the accuracy with no significant increase in the processing time.