Peningkatan Akurasi Klasifikasi Digit melalui Modifikasi CNN dengan Batch Normalization, Dropout, dan Data Augmentation

Abstract

Digit image classification is a fundamental task in pattern recognition and computer vision. This study aims to improve the performance of a baseline Convolutional Neural Network (CNN) by applying several architectural modifications, including data augmentation, batch normalization, dropout, and deeper convolutional layers. The digit dataset was divided into training and validation sets with an 80:20 ratio. Augmentation techniques such as random rotation, translation, and zoom were applied to enhance data variability and reduce overfitting. Batch normalization was used to stabilize training, while dropout served as a regularization method to prevent overfitting in dense layers. Experimental results demonstrate that the modified CNN achieved high validation accuracy and produced a confusion matrix indicating consistent classification across all digit classes. These findings suggest that enhancing a simple CNN architecture can significantly improve performance without requiring complex transfer learning approaches. This method is suitable for small to medium-scale image classification tasks.