EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Efficient CNN Hardware Accelerator for Traffic Sign Classification

Jihoon Yu, Heesu Lee, Miso Shin, Eunji Park, Sunhee Kim

International Journal on Advanced Science, Engineering and Information Technology · 28.06.2026 · 16 (3) · 1052-1059

DOI: 10.18517/ijaseit.16.3.14128 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal on Advanced Science, Engineering and Information Technology, online erschienen am 28.06.2026. Verfasst von Jihoon Yu und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the commercialization of autonomous driving technology, driving convenience has significantly improved. However, securing reliable recognition performance in limited computational resources and memory environments remains a critical challenge. Embedded vision systems, in particular, must simultaneously meet constraints such as power consumption, memory capacity, and computational complexity while maintaining high accuracy. So, it is essential to optimize neural networks considering hardware implementation. The purpose of this study is to design a CNN-based traffic sign recognition model that considers the balance between classification accuracy and memory efficiency for the embedded vision system and to analyze its performance. To this end, the CNN model was trained using the GTSRB dataset in the Python-TensorFlow environment. The reference model achieved classification accuracy of about 95% using floating-point parameters. In order to reduce computational complexity and memory usage, an average pooling layer was added to the network structure, and all model parameters were quantized in the 8-bit Q4.4 fixed-point format. Training was performed for a total of 50 epochs using an Adam optimizer, with a learning rate of 0.001 and a mini-batch size of 8. Experimental results show that a model using the proposed optimization technique reduces memory usage by approximately 97% compared to a floating-point-based model while maintaining a classification accuracy of approximately 93.5%. In addition, hardware implementation in Verilog confirmed that quantization and structural simplification had only a limited impact on classification performance. This study shows that hardware-friendly CNN optimization can effectively improve memory efficiency without significantly reducing recognition accuracy.

Zitieren

APA

Yu, J., Lee, H., Shin, M., Park, E., & Kim, S. (2026). Efficient CNN Hardware Accelerator for Traffic Sign Classification. International Journal on Advanced Science, Engineering and Information Technology, 16(3), 1052-1059. https://doi.org/10.18517/ijaseit.16.3.14128

Vancouver

Yu J, Lee H, Shin M, Park E, Kim S. Efficient CNN Hardware Accelerator for Traffic Sign Classification. International Journal on Advanced Science, Engineering and Information Technology. 2026;16(3):1052-1059. doi:10.18517/ijaseit.16.3.14128

BibTeX
@article{Yu2026,
  author = {Yu, Jihoon and Lee, Heesu and Shin, Miso and Park, Eunji and Kim, Sunhee},
  title = {Efficient CNN Hardware Accelerator for Traffic Sign Classification},
  journal = {International Journal on Advanced Science, Engineering and Information Technology},
  year = {2026},
  volume = {16},
  number = {3},
  pages = {1052--1059},
  doi = {10.18517/ijaseit.16.3.14128}
}
Kostenlos starten →