A Classifier for Reducing Numerical Errors using Ensemble Method
Yuta Suzuki, Toi Tsuneda, Daiki Kuyoshi, Satoshi Yamane · 2021 IEEE 10th Global Conference on Consumer Electronics (GCCE) · 2021
Nowadays, CNN have shown high identification accuracy in handwritten numeral recognition, but normal CNN learns image features from the pixel values of the image, so they could not learn the numerical information of the image. In this study, we propose a method to reduce the numerical error in handwritten numeric images in addition to improving the identification accuracy. Our proposed method improves performance by using an ensemble method that combines a CNN that learns to reduce numerical errors with a normal CNN. The results show that the proposed method produces higher identification rates and smaller numerical errors than the baseline method on the MNIST and KANNADA-MNIST datasets.