A Lightweight CNN Baseline for Offline Handwritten Digit Recognition on MNIST: Design and Evaluation

Mohammed, Raid Ismael

doi:https://doi.org/10.33103/2617-3352.1517

	A Lightweight CNN Baseline for Offline Handwritten Digit Recognition on MNIST: Design and Evaluation
IRAQI JOURNAL OF COMPUTERS, COMMUNICATIONS, CONTROL AND SYSTEMS ENGINEERING
Volume 26, Issue 1, April 2026 PDF (3.37 M)
Document Type: Research Paper
DOI: https://doi.org/10.33103/2617-3352.1517
Author
Raid Ismael Mohammed^*
Islamic University of Lebanon , Faculty of Engineering, Khalde, Lebanon
Abstract
Handwritten digit recognition is a classic computer vision task, often used as a benchmark for image classification models. This paper presents a lightweight convolutional neural network (CNN) baseline for oﬄine handwritten digit recognition using the MNIST dataset. The proposed CNN has a simple architecture (two convolutional layers with max-pooling, followed by a flat- tening and two fully connected layers) and is designed to be computationally eﬃcient while achieving high accuracy. We train the model with standard settings (Adam optimizer, batch size 128,∼10 epochs) and obtain∼98–99% recognition accuracy on MNIST. We provide a detailed evaluation including training curves and a confusion matrix to analyze performance per class. Despite its simplicity (only∼37k parameters), the model’s accuracy approaches that of more complex networks, confirming that even a small CNN can serve as a strong baseline for MNIST. We discuss the model’s performance relative to other approaches (e.g., SVM, KNN, deeper CNNs), its limitations on this virtually solved task, and directions for future work such as applying the design to more complex datasets or optimizing it for embedded deployment.loyment.
Keywords
MNIST; Convolutional neural networks; Lightweight models; Handwritten digit recognition; Deep learning

Statistics Article View: 20 PDF Download: 13