RRBM-YOLO: Research on Efficient and Lightweight Convolutional Neural Networks for Underground Coal Gangue Identification

Yutong Wang, Ziming Kou, Cong Han, Yuchen QinSensors 2024 · DOI

TL;DR

This paper introduces RRBM-YOLO, a novel, lightweight, and efficient deep learning model for identifying coal gangue in challenging underground mining environments. By combining a robust YOLOv8 backbone with custom enhancements like dark light image preprocessing and specialized modules, RRBM-YOLO significantly improves detection accuracy and speed, making it suitable for real-world industrial applications.

What Problem Does This Paper Address?

Underground coal mining operations generate vast amounts of coal gangue, a waste material that poses significant environmental risks if not properly managed. Accurate and efficient identification of coal gangue is the crucial first step in its separation and management. However, existing methods face several challenges:

  • Harsh Environments: Underground mines are characterized by poor lighting, dust, and complex geological formations, making image acquisition and analysis difficult.
  • Low Efficiency: Traditional manual inspection is labor-intensive and prone to errors.
  • Complex Algorithms & High Hardware Demands: Many advanced computer vision techniques require substantial computational resources, making them impractical for deployment in resource-constrained mining settings.
  • Lack of Public Datasets: There's a scarcity of publicly available datasets specifically for coal gangue detection in these challenging conditions.

This paper aims to bridge this gap by developing a "hardware-friendly" coal gangue identification system that is accurate, efficient, and robust to the difficult environmental conditions.

Methods

The proposed RRBM-YOLO model builds upon the state-of-the-art YOLOv8 object detection framework, incorporating several key innovations:

  • Retinexformer for Dim Light Enhancement: To address the poor lighting conditions in underground mines, the Retinexformer algorithm is used for preprocessing images. This method enhances illumination and reduces noise and artifacts, making details more visible.
  • Lightweight Backbone: The core of the model utilizes YOLOv8, known for its efficiency.
  • RepGhost Module: This lightweight module is integrated to reduce computational complexity and inference time without sacrificing feature extraction capabilities. It achieves this through structural reparameterization.
  • BiFPN (Bi-directional Feature Pyramid Network): This module enhances feature fusion by incorporating both top-down and bottom-up pathways, allowing for better aggregation of multi-scale features.
  • MCA (Multidimensional Collaborative Attention Module): This novel attention mechanism focuses on capturing interactions across channel, height, and width dimensions, improving feature representation and model convergence with minimal overhead.
  • Custom Dataset Creation: Recognizing the lack of public datasets, the authors collected and curated their own dataset under various lighting conditions.

Datasets

The paper introduces a custom dataset for coal gangue identification. This dataset was created by collecting images in two scenarios:

  • Normal Lighting: Standard image capture.
  • Simulated Underground Lighting: Images captured under poor lighting conditions mimicking those found in mines.

The dataset comprises six categories: Normal_light (coal, coal gangue mixture, gangue) and Low_light (gangue, coal, coal gangue mixture), and High_light (gangue, coal, coal gangue mixture). A total of 8400 image samples were collected and annotated, divided into training (5880 images), validation (1680 images), and testing (840 images) sets in a 7:2:1 ratio. The paper also mentions testing the model on public datasets like CrowdHuman and COCO to evaluate its generalization capabilities.

Key Insights and Ideas

  • Dark Light Enhancement is Crucial: The success of the model heavily relies on effectively handling low-light conditions, highlighting the importance of specialized preprocessing techniques like Retinexformer.
  • Lightweight Modules are Key for Industrial Deployment: Integrating modules like RepGhost is essential for achieving "hardware-friendly" models that can run efficiently on less powerful hardware, a common constraint in industrial settings.
  • Multi-dimensional Attention Boosts Performance: The MCA module's ability to consider spatial and channel dimensions simultaneously offers a significant advantage in capturing complex feature interactions, leading to improved accuracy.
  • Custom Datasets are Necessary for Niche Problems: The creation of a tailored dataset was vital for training and evaluating a model that performs well in the specific, challenging environment of underground coal mines.

Key Results

The RRBM-YOLO model demonstrated significant improvements over the original YOLOv8 and other baseline models:

  • Accuracy: Achieved a precision of 0.988.
  • mAP: [email protected] increased by 10.49% and [email protected]:0.95 increased by 36.62% compared to the original YOLOv8.
  • Inference Speed: Reached 8.1 GFLOPS, representing a substantial speedup (92.32% higher than original YOLOv8).
  • Generalization: Showed strong performance across various datasets, including CrowdHuman, COCO, and rock datasets, indicating good generalization ability.
  • Ablation Studies: Adding individual modules (Retinexformer, BiFPN, RepGhost, MCA) showed performance gains, with the combined model achieving the best results. The MCA module, in particular, showed a significant increase in [email protected] and [email protected]:0.95 compared to other attention mechanisms.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the model is "hardware friendly," further optimization might be needed for extremely resource-constrained environments.
  • The dataset, while custom-built, might still benefit from further expansion and diversity.
  • Future Work:
    • Expand and enhance the dataset by analyzing the effects of natural changes like mine humidity and temperature on coal gangue characteristics.
  • Develop methods to differentiate "average coal" (coal-gangue blend) from pure coal streams for even more precise classification.
  • Further improve the model's precision and robustness to enhance its application potential in industrial practice.

Summary updated

Abstract

Coal gangue identification is the primary step in coal flow initial screening, which mainly faces problems such as low identification efficiency, complex algorithms, and high hardware requirements. In response to the above, this article proposes a new "hardware friendly" coal gangue image recognition algorithm, RRBM-YOLO, which is combined with dark light enhancement. Specifically, coal gangue image samples were customized in two scenarios: normal lighting and simulated underground lighting with poor lighting conditions. The images were preprocessed using the dim light enhancement algorithm Retinexformer, with YOLOv8 as the backbone network. The lightweight module RepGhost, the repeated weighted bi-directional feature extraction module BiFPN, and the multi-dimensional attention mechanism MCA were integrated, and different datasets were replaced to enhance the adaptability of the model and improve its generalization ability. The findings from the experiment indicate that the precision of the proposed model is as high as 0.988, the [email protected](%) value and [email protected]:0.95(%) values increased by 10.49% and 36.62% compared to the original YOLOv8 model, and the inference speed reached 8.1GFLOPS. This indicates that RRBM-YOLO can attain an optimal equilibrium between detection precision and inference velocity, with excellent accuracy, robustness, and industrial application potential.
BibTeX
@inproceedings{wang2024rrbm,
  author = {Yutong Wang and Ziming Kou and Cong Han and Yuchen Qin},
  title = {{RRBM-YOLO: Research on Efficient and Lightweight Convolutional Neural Networks for Underground Coal Gangue Identification}},
  booktitle = {Sensors},
  year = {2024},
  doi = {10.3390/s24216943},
}
Figure