A spatial linear image filtering algorithm based on a separable binomial kernel for streaming implementation in programmable logic

The effect of the Gaussian filter kernel size on the object detection performance of a lightweight convolutional neural network detector in a low-light video scene is studied. A separable filter model oriented towards streaming implementation in programmable logic is considered, and the hardware implementation options are compared in terms of the number of operations per pixel. The aim of the work is an experimental assessment of the influence of the kernel size on detection latency and detection stability, together with a verification that the detection performance is preserved when moving to an implementation without hardware multipliers. Processing was carried out frame by frame for a video sequence of 1351 frames at two values of the detector confidence threshold. It is established that Gaussian filtering acts mainly by increasing detection stability rather than by improving early detection: the number of frames with detection grows by 6.7 % at the baseline confidence threshold and by 25 % at the strict one, whereas the moment of first detection is practically unchanged. At the strict threshold and a kernel size exceeding three samples, a degradation of early detection is observed, caused by excessive smoothing of an object of small angular size; this sets the upper bound of the admissible parameter region. It is shown that the closeness of the binomial approximation to the exact Gaussian filter is preserved at the level of the responses of a nonlinear detector as well: the discrepancy in the detection moment does not exceed one frame, and in the number of frames with detection it is 0.7 %. The parameter selection problem is formalised as maximisation of detection stability under a constraint on detection latency. The optimum is a three-sample kernel with coefficients one, two, one and normalisation by a two-bit shift, requiring four adders and two line buffers with zero consumption of digital signal processing blocks.

1. Medvedeva E.V., Perevoshchikova A.A. Research of neural network algorithms for recognizing railway infrastructure objects in video images. Computer Optics. 2025;49(3):443–450. (In Russ.). https://doi.org/10.18287/2412-6179-CO-1563

2. Redmon J., Farhadi A. YOLOv3: An Incremental Improvement. arXiv. URL: https://arxiv.org/abs/1804.02767 [Accessed 15th May 2026].

3. Araki Y., Matsuda M., Manabe T., Ishizuka Y., Shibata Y. FPGA Implementation of an Object Recognition System with Low Power Consumption Using a YOLOv3-tiny-based CNN. In: Complex, Intelligent and Software Intensive Systems: Proceedings of the 16th International Conference on Complex, Intelligent and Software Intensive Systems (CISIS-2022), 29 June – 01 July 2022, Virtual Event. Cham: Springer; 2022. P. 223–233. https://doi.org/10.1007/978-3-031-08812-4_22

4. Gonzalez R.C., Woods R.E. Digital Image Processing. New York: Pearson; 2018. 1022 p.

5. Mostafa Sh.Sh., Sousa L.N., Ferreira N.F., et al. Full image-processing pipeline in field-programmable gate array for a small endoscopic camera. Journal of Electronic Imaging. 2017;26(1):013005. https://doi.org/10.1117/1.JEI.26.1.013005

6. Sever K.O., Alekseev K.N., Turulin I.I. Implementation of an efficient separable vector digital filter on FPGA. Izvestiya SFedU. Engineering Sciences. 2024;(3):158–168. (In Russ.). https://doi.org/10.18522/2311-3103-2024-3-158-168

7. Aubury M., Luk W. Binomial filters. Journal of VLSI Signal Processing Systems for Signal, Image, and Video Technology. 1996;12(1):35–50. https://doi.org/10.1007/BF00936945

8. Zhao M., Shi P., Xu X., Xu X., Liu W., Yang H. Improving the accuracy of an R-CNN-based crack identification system using different preprocessing algorithms. Sensors. 2022;22(18):7089. https://doi.org/10.3390/s22187089

9. Gholinavaz S., Saeedi N., Samadi Gharehveran S. Robustness analysis of YOLO and faster R-CNN for object detection in realistic weather scenarios with noise augmentation. Scientific Reports. 2025;15(1):44888. https://doi.org/10.1038/s41598-025-28737-5

10. Yano T., Kuroki Y. Fast implementation of Gaussian filter by parallel processing of binominal filter. In: 2016 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS), 24–27 October 2016, Phuket, Thailand. IEEE; 2016. https://doi.org/10.1109/ISPACS.2016.7824738

11. Cambuim L., Barros E. FPGA-based pedestrian detection for collision prediction system. Sensors. 2022;22(12):4421. https://doi.org/10.3390/s22124421

12. Mursaev A.Kh. Hardware implementation of median filtering. Nauchnoe priborostroenie. 2013;23(3):114–119. (In Russ.).

13. Rodríguez-Andina J.J., Moure M.J., Valdés M.D. Features, design tools, and application domains of FPGAs. IEEE Transactions on Industrial Electronics. 2007;54(4):1810–1823. https://doi.org/10.1109/TIE.2007.898279

14. Wang J., Yang P., Liu Y., et al. Research on improved YOLOv5 for low-light environment object detection. Electronics. 2023;12(14):3089. https://doi.org/10.3390/electronics12143089

15. Gladkov L.A., Gladkova N.V. Algorithm for synthesis of combinational logic circuits based on the evolutionary approach. Izvestiya SFedU. Engineering Sciences. 2023;(6):24–34. (In Russ.). https://doi.org/10.18522/2311-3103-2023-6-24-34

16. Li X., Li Z., Zhou L., Huang Zh. FOLD: Low-level image enhancement for low-light object detection based on FPGA MPSoC. Electronics. 2024;13(1):230. https://doi.org/10.3390/electronics13010230