Handwritten Digit Recognition Using CNN with Average Pooling and Global Average Pooling

Recognition of handwritten digits is crucial because it enables computers to comprehend and understand human handwriting. The automatic sorting of mail, processing of financial records, and reading of prescriptions are only a few examples of its many practical uses. It can also be used in the classroom for exam grading. The automating process with handwritten digit recognition eliminates mistakes and saves time. The major goal of this paper is to examine several existing model configurations and give trustworthy and effective ways for detecting transcribed numerical data. This paper highlights Convolutional Neural Network (CCN) with Average Pooling (AP) and Global Average Pooling (GAP) layer for proper extraction of insights from the feature maps. Results indicate that the proposed Convolutional Neural Network classifier with AP and GAP layer outperformed existing neural network with significantly increased computational efficiency without sacrificing execution. This paper development is based on the Modified National Institute of Standards and Technology, or MNIST dataset, which is a well-known handwritten digits dataset in the machine learning and computer vision fields. Each handwritten digit is a 28x28 grayscale image, and there are 70,000 of them in total numbers from 0 to 9. Thus, the classification model is of classes 10. 60,000 images are there in the training dataset and 10,000 images are there in the testing dataset. The proposed CNN Model with AP and GAP Layer achieves a classification accuracy of 99% on the MNIST Dataset.