High-Throughput Line Buffer Microarchitecture for Arbitrary Sized Streaming Image Processing | JoVE Visualize

Area of Science:

Computer Engineering
Digital Signal Processing
Hardware Architecture

Background:

Real-time image processing demands high-throughput and low-latency solutions, often implemented on Field-Programmable Gate Arrays (FPGAs).
Existing FPGA streaming architectures struggle with seamless, uninterrupted image size and resolution changes, crucial for applications like region of interest (ROI) detection.
Run-time reconfiguration in current solutions leads to interruptions, hindering continuous processing at high frame rates.

Purpose of the Study:

To propose a dynamically-programmable buffer architecture (D-SWIM) for FPGAs that enables arbitrary image size processing without interruptions.
To redefine on-chip memory organization and control for adaptive hardware accommodating variable image dimensions.
To achieve seamless image size-switching in real-time image processing streams.

Main Methods:

Developed the dynamically-programmable buffer architecture (D-SWIM) based on the Stream-Windowing Interleaved Memory (SWIM) architecture.
Implemented D-SWIM to redefine on-chip memory organization and control for dynamic adaptation to arbitrary image sizes.
Evaluated D-SWIM performance using real-world applications like 2D-Convolution and Harris Corner Detector.

Main Results:

D-SWIM enables arbitrary image size adaptation with sub-100 ns delay, ensuring minimal interruption.
Achieved dynamic programmability with only slight logic overhead, saving up to 56% of BRAM resources compared to prior SWIM.
Reached a maximum operating frequency of 329.5 MHz, reduced power consumption by 45.7%, and achieved pixel throughputs of 4.5 Giga Pixel/s and 4.2 Giga Pixel/s.
Improved hardware efficiency up to 30x compared to prior streaming frameworks.

Conclusions:

D-SWIM provides a breakthrough in FPGA-based real-time image processing by enabling seamless, dynamic image size-switching.
The architecture offers significant improvements in resource utilization, power efficiency, and processing speed.
D-SWIM is highly suitable for vision-guided intelligent applications requiring flexible and uninterrupted image stream processing.