Hybrid Architecture for Handwritten Numeral Recognition
2017
This paper presents a unique recognition system based on two modelsa Random Forest and a Multi-Layer Perceptron.This hybrid architecture is used to classify handwritten digits, taken from the MNIST dataset.The system has two outcomesprediction or no prediction.If both models provide the same output, system provides a prediction, else it does not.The Random Forest and Multi-Layer Perceptron are used specifically as they have shown high accuracies individually for the dataset, that is, 97.06% and 97.87% respectively.This would allow for lower rates of output mismatches.The paper also demonstrates that excluding certain pixels (features) of the image which have low variance helps increase accuracy and improve speed of computation.The proposed architecture has helped us to keep false predictions under 1% when used for the test set provided.The hybrid architecture performs better than the individual architectures alone.