An FPGA-Based YOLOv5 Accelerator for Real-Time Industrial Vision Applications.

Area of Science:

• Computer Vision
• Hardware Acceleration
• Deep Learning

Background:

• The You Only Look Once (YOLO) object detection network is widely used in industry for its speed and accuracy.
• Existing 8-bit quantization FPGA accelerators face challenges meeting demands for higher performance.
• Industrial automation requires faster and more accurate object detection systems.

Purpose of the Study:

• To present a novel 4-bit quantization-based neural network accelerator for the YOLOv5 model.
• To enhance real-time processing capabilities and maintain high detection accuracy.
• To improve performance and resource efficiency in industrial object detection systems.

Main Methods:

• Implemented an optimized 4-bit quantization scheme for the entire YOLO network, including the first layer.
• Developed a unified Digital Signal Processor (DSP) packing scheme.
• Introduced a novel parity adder tree architecture to support the quantization strategies.

Main Results:

• Achieved a 4-bit quantization with only a 1.5% degradation in mean Average Precision (mAP).
• Reduced on-chip DSP utilization by 50% through the proposed hardware architecture.
• The FPGA accelerator achieved 808.6 GOPS throughput and 0.49 GOPS/DSP efficiency for YOLOv5s.

Conclusions:

• The proposed 4-bit quantization accelerator significantly enhances YOLOv5 performance for industrial object detection.
• The novel hardware architecture offers substantial improvements in speed, efficiency, and resource utilization.
• This work provides a viable solution for next-generation industrial automation requiring high-performance real-time object detection.