EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Unified Single-Pass Streaming Data-Flow Accelerator for Multi-Output Image Augmentation on FPGA MPSoCs

Fakhra Aftab, Majida Kazmi

IEEE Access · 2026 · 14 · 92856-92875

DOI: 10.1109/access.2026.3705025 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Fakhra Aftab und 1 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Ministry of Science and Technology (MoST) Endowment Fund.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Data augmentation is widely used to improve the robustness and generalization of machine vision models, particularly in industrial inspection systems, where variability in lighting, orientation, and object appearance is common. Conventional software-based augmentation introduces significant pre-processing overhead in high-resolution and high-throughput vision pipelines. Despite its importance, relatively little attention has been paid to accelerating this stage, as most existing research focuses on deep learning inference acceleration rather than data preparation. This work presents a hardware-accelerated image augmentation architecture implemented using the Xilinx Vitis Unified Software Platform (v2023.2) targeting the ZCU102 Evaluation Board. The proposed design employs a single-pass streaming data-flow architecture with on-chip FIFOs, enabling concurrent execution of memory access, geometric transformation, and photometric processing stages. Multiple AXI4 memory-mapped interfaces are utilized to support parallel output streams and reduce memory contention. The accelerator achieves effective streaming behavior close to one pixel per clock cycle after pipeline fill, operating at a target frequency of 150 MHz, processes a $700\times 400$ image in 1.868 ms, corresponding to kernel-level and system-level throughput of approximately 535 FPS and 280 FPS respectively. A roof-line-based performance model is proposed to characterize the compute and memory bounds of the design, revealing that system-level performance is primarily constrained by external memory bandwidth and data movement overhead. The post-synthesis results show modest resource utilization, including 12.01% BRAM, 0.12% DSPs, 3.38% FFs, and 6.29% LUTs, with a total on-chip power consumption of 4.17 W and an energy cost of 27.8 nJ per pixel. These results demonstrate that the proposed architecture enables efficient and low-latency augmentation for real-time vision pipelines.

Förderung

  • Ministry of Science and Technology (MoST) Endowment Fund

Zitieren

APA

Aftab, F., & Kazmi, M. (2026). A Unified Single-Pass Streaming Data-Flow Accelerator for Multi-Output Image Augmentation on FPGA MPSoCs. IEEE Access, 14, 92856-92875. https://doi.org/10.1109/access.2026.3705025

Vancouver

Aftab F, Kazmi M. A Unified Single-Pass Streaming Data-Flow Accelerator for Multi-Output Image Augmentation on FPGA MPSoCs. IEEE Access. 2026;14:92856-92875. doi:10.1109/access.2026.3705025

BibTeX
@article{Aftab2026,
  author = {Aftab, Fakhra and Kazmi, Majida},
  title = {A Unified Single-Pass Streaming Data-Flow Accelerator for Multi-Output Image Augmentation on FPGA MPSoCs},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {92856--92875},
  doi = {10.1109/access.2026.3705025}
}
Kostenlos starten →