Neural Networks Based on a Learnable Two-Dimensional Separable Transform for Image Classification: Theory and Hardware Implementation on FPGA

Objectives. Development of methods for design compact and efficient neural networks for image recognition tasks, as well as their hardware implementation based on FPGA.

Methods. The paper proposes the concept of a learnable two-dimensional separable transformation (LST) for designing feedforward neural networks for image recognition tasks. A feature of the LST is the sequential  processing of image rows by a fully connected layer, after which the resulting representation is processed by columns using second fully connected layer. In the proposed architecture of a feedforward neural network, the LST is considered as a feature extractor. The hardware implementation of LST-based neural network is based on the concept of in-place computing (shared memory for storing source and intermediate data), as well as using a single set of computing cores to calculate all layers of the neural network.

Results. A family of compact neural network architectures LST-1 is proposed, differing in the image  embedding size. Experiments on the classification of MNIST handwritten digits have shown the high efficiency of these models: the LST-1-28 network achieves 98.37 % accuracy with 9.5 K parameters, and the more compact LST-1-8 shows 96.53 % accuracy with 1.1 K parameters. Testing of the LST-1-28 hardware implementation  confirms the architecture's resistance to parameter quantization errors.

Conclusion. The proposed concept of a learnable two-dimensional separable transformation provides the  design of compact and efficient neural network architectures characterized by: a small number of learnable parameters, high recognition accuracy, and the regular structure of the algorithm, which makes it possible to obtain their  effective implementations based on FPGAs.

1. Park J., Sung W. FPGA based implementation of deep neural networks using on-chip memory only. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Shanghai, China, 20–25 March 2016. Shanghai, 2016, pp. 1011–1015.

2. Medus L. D., Iakymchuk T., Frances-Villora J. V., Bataller-Mompean M., Rosado-Munoz A. A novel systolic parallel hardware architecture for the FPGA acceleration of feedforward neural networks. IEEE Access, 2019, vol. 7, pp. 76084–76103.

3. Han S., Mao H., Dally W. J. Deep compression: сompressing deep neural networks with pruning, trained quantization and Huffman coding. International Conference on Learning Representations (ICLR), San Juan, Puerto Rico, 2–4 May 2016. San Juan, 2016, pp. 1–14.

4. Samragh M., Ghasemzadeh M., Koushanfar F. Customizing neural networks for efficient FPGA implementation. Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM), Napa, CA, USA, 30 April – 2 May 2017. Napa, 2017, pp. 85–92.

5. Usatyuk V., Egorov S. Boosting DNN efficiency: replacing FC layers with graph embeddings for hardware acceleration. International Conference on Digital Signal Processing and its Applications (DSPA), Moscow, Russia, 26–28 March 2025. Moscow, 2025, pp. 1–6.

6. Kwon J., Kim S. Design of a low-area digit recognition accelerator using MNIST database. JOIV: International Journal on Informatics Visualization, 2022, vol. 6, no. 1, pp. 53–59.

7. Westby I., Yang X., Liu T., Xu H. FPGA acceleration on a multilayer perceptron neural network for digit recognition. The Journal of Supercomputing, 2021, vol. 77, no. 12, pp. 14356–14373.

8. Vashkevich M., Krivalcevich E. Compact and efficient neural networks for image recognition based on learned 2D separable transform. International Conference on Digital Signal Processing and its Applications (DSPA), Moscow, Russia, 26–28 March 2025. Moscow, 2025, pp. 1–5.

9. Starovoitov V. V., Golub Yu. I. Data normalization in machine learning. Informatika [Informatics], 2021, vol. 18, no. 3, рр. 83–96 (In Russ.).

10. Loshchilov I., Hutter F. SGDR: Stochastic Gradient Descent with Warm Restarts, 2016. Available at: https://arxiv.org/abs/1608.03983 (accessed 01.08.2025).

11. Srivastava N., Hinton G., Krizhevsky A. Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 2014, vol. 15, no. 1, pp. 1929–1958.

12. Liang S., Yin S., Liu L., Luk W., Wei S. FP-BNN: Binarized neural network on FPGA. Neurocomputing, 2018, vol. 275, pp. 1072–1086.

13. Huynh T. V. Deep neural network accelerator based on FPGA. NAFOSTED Conference on Information and Computer Science, Hanoi, Vietnam, 24–25 November 2017. Hanoi, 2017, pp. 254–257.

14. Krivalсevich E. A., Vashkevich M. I. Investigation of hardware implementation of a feedforward neural network for handwritten digit recognition based on FPGA. Doklady BGUIR, 2025, vol. 23, no. 2, pp. 101–108 (In Russ.).

15. Solovyev R. A., Kustov A. G., Ruhlov V. S., Shchelokov A. N., Puzyrkov D. V. Hardware implementation of a convolutional neural network in FPGA based on fixed point calculations. Izvestija Juzhnogo federal'nogo universiteta. Tehnicheskie nauki [Proceedings of Southern Federal University. Engineering Sciences], 2017, vol. 192, no. 7, pp. 186–197 (In Russ.).