RTL Design of a CNN-Based FPGA Accelerator for Handwritten Digit Recognition
International Journal of Communication Systems and Network Technologies · 30.04.2026 · 15 (1) · 68-81
DOI: 10.18486/ijcsnt/15.1.005 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Communication Systems and Network Technologies, online erschienen am 30.04.2026. Verfasst von Vivek Rathi und 3 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Handwritten digit recognition plays a key role in areas like mail sorting, banking, and document processing. Software-based Convolutional Neural Networks (CNNs) achieve high recognition accuracy, but their computing power and energy needs limit their use in embedded systems. This paper presents the design and RTL-level simulation of a CNN-based hardware accelerator for handwritten digit recognition targeting a Xilinx Kintex-7 FPGA. The accelerator implements a modified LeNet-5 CNN architecture entirely in Verilog using fixed-point arithmetic. It consists of convolution layers followed by ReLU activation and max pooling, along with a fully connected output layer. The CNN model is trained offline using Python, and the trained and quantized weights and biases are converted to fixed-point representation and mapped directly into the RTL design. A streaming, line-buffer-based architecture is employed to efficiently generate convolution windows and enable continuous data flow through the network. The design is verified at the RTL level using cycle-accurate simulation with MNIST test images, ensuring functional correctness of all network layers. The proposed accelerator achieves a classification accuracy of 91% on 1,000 MNIST test images while utilizing 12,709 Slice LUTs and 238 DSP slices, with only 796 trainable parameters. These results demonstrate that the RTL-based CNN accelerator provides an effective trade-off between recognition accuracy and hardware resource utilization, making it suitable for embedded FPGA-based inference. Hardware implementation and on-board validation are planned as future work.
Zitieren
APA
Rathi, V., Kulkarni, V., Ingle, H., & Agnihotri, M. P. (2026). RTL Design of a CNN-Based FPGA Accelerator for Handwritten Digit Recognition. International Journal of Communication Systems and Network Technologies, 15(1), 68-81. https://doi.org/10.18486/ijcsnt/15.1.005
Vancouver
Rathi V, Kulkarni V, Ingle H, Agnihotri MP. RTL Design of a CNN-Based FPGA Accelerator for Handwritten Digit Recognition. International Journal of Communication Systems and Network Technologies. 2026;15(1):68-81. doi:10.18486/ijcsnt/15.1.005
BibTeX
@article{Rathi2026,
author = {Rathi, Vivek and Kulkarni, Vedant and Ingle, Harshal and Agnihotri, Mrs. P.S.},
title = {RTL Design of a CNN-Based FPGA Accelerator for Handwritten Digit Recognition},
journal = {International Journal of Communication Systems and Network Technologies},
year = {2026},
volume = {15},
number = {1},
pages = {68--81},
doi = {10.18486/ijcsnt/15.1.005}
}
EUVIMED