YOLOv4-Tiny-Based Coal Gangue Image Recognition and FPGA Implementation.

Area of Science:

• Computer Vision
• Embedded Systems
• Artificial Intelligence

Background:

• Deep learning for coal gangue identification typically requires high-power CPUs/GPUs, limiting deployment in underground mines due to size, heat, and power demands.
• Existing methods face challenges in real-time application within the harsh and resource-constrained environments of coal mines.

Purpose of the Study:

• To develop and deploy an efficient coal gangue identification method on a low-power FPGA platform.
• To overcome the limitations of high-power hardware in underground coal mine environments.

Main Methods:

• A YOLOv4-tiny model was trained and optimized using 16-bit fixed-point quantization and layer integration.
• Custom convolution and pooling IP kernels were designed for FPGA acceleration, incorporating parallelism, pipelining, and ping-pong operations.
• A complete FPGA hardware system was implemented for the entire identification algorithm.

Main Results:

• The FPGA-based system achieved a mean Average Precision (mAP) of 96.56% for coal gangue recognition.
• Image recognition speed was 0.376 seconds per image, falling between CPU and GPU performance.
• Hardware power consumption was remarkably low at 2.86 W, with energy efficiency ratios 10.42x (vs. CPU) and 3.47x (vs. GPU).

Conclusions:

• The proposed YOLOv4-tiny based FPGA system offers a viable, low-power solution for coal gangue identification in underground mines.
• The method balances recognition accuracy with significant improvements in power efficiency and hardware footprint.
• This approach enhances the feasibility of deploying advanced AI for safety and operational improvements in challenging mining conditions.