Improved YOLOv8-Based Target Precision Detection Algorithm for Train Wheel Tread Defects

Yu Wen, Xiaorong Gao, Lin Luo, Jinlong Li2024 · DOI

TL;DR

This paper tackles the critical challenge of accurately detecting defects on train wheel treads, a task complicated by water stains and small, subtle flaws. The researchers enhance the YOLOv8 object detection model by optimizing its detection layers for small targets, incorporating a novel attention mechanism, and employing a more robust loss function (SIoU), achieving state-of-the-art performance on their custom dataset.

What Problem Does This Paper Address?

The safety of high-speed train operations hinges on the integrity of train wheels, particularly the tread surface. Defects like bruises, peeling, and scratches can compromise safety, necessitating timely detection and maintenance. Traditional image-based detection methods often struggle with misidentifying water stains as defects and failing to detect small, subtle flaws. While deep learning models like YOLOv8 have shown promise, they can still be susceptible to these issues, especially in real-world scenarios with varying lighting and the presence of water. This paper aims to improve the robustness and accuracy of YOLOv8 for train wheel tread defect detection, specifically addressing the challenges posed by water stains and small defects.

Methods

The core of this work is the enhancement of the YOLOv8 architecture. The authors introduce several key modifications:

  • Optimized Detection Layers: To mitigate the misdetection of water stains, the authors modify the network's detection layers. They remove a large-target detection layer and add a shallow, small-target detection layer. This shift in focus allows the network to better distinguish between water stains and actual defects, as shallow layers are found to be more effective at filtering out water stains.
  • Reconstructed SPPCSPC Module with SimAM: To improve the detection of small targets, the original SPPF (Spatial Pyramid Pooling - Fast) module is replaced with a reconstructed SPPCSPC (Spatial Pyramid Pooling - Concurrent Spatial Pyramid Convolution) module. This new module integrates spatial pyramid pooling with cross-stage partial channel connections. Crucially, they also introduce the SimAM (Simple Attention Mechanism) into this module. SimAM assigns 3D attention weights to neurons, helping to differentiate between complex backgrounds and small-target defects, thereby reducing feature loss and improving robustness.
  • SIoU Loss Function: The standard loss function is replaced with SIoU (Soft IoU). Unlike previous IoU-based losses (like CIoU) that primarily consider overlap, distance, and aspect ratio, SIoU also incorporates angular loss. This accounts for the orientation difference between predicted and ground truth bounding boxes, leading to more accurate localization and faster convergence.

Datasets

The study utilizes a custom dataset comprising 6547 images of train wheel treads. These images were captured under varying conditions, including different lighting and the presence of water stains, making them representative of real-world challenges. The dataset is partitioned into training (70%), validation (20%), and testing (10%) sets.

Key Insights and Ideas

  • Shallow is Better for Water Stains: The analysis using Grad-CAM heatmaps reveals that shallow network layers are more effective at distinguishing between water stains and defects. This insight directly informs the architectural change of removing large-target detection layers and adding shallow ones.
  • Attention is Crucial for Small Defects: The integration of SimAM highlights the importance of attention mechanisms in focusing the network's resources on subtle features, especially when dealing with small defects against complex backgrounds.
  • Orientation Matters for Bounding Boxes: The adoption of SIoU underscores that accurately predicting the bounding box orientation is as critical as overlap and distance for precise object localization, particularly in challenging detection scenarios.

Key Results

The enhanced YOLOv8 model achieved impressive results on the tread defect detection task:

  • High Accuracy: The model achieved a precision of 96.30%, a recall of 95.31%, and a mean average precision ([email protected]) of 96.95%.
  • Significant Improvement: Compared to the original YOLOv8, the improved model shows substantial gains, with mAP increasing by 8.76%.
  • Reduced False Alarms: The false alarm rate for tread defects was reduced from 16.64% in the original YOLOv8 to 4.69% in the improved version.
  • Competitive Performance: When compared to other mainstream algorithms like Faster R-CNN, SSD, YOLOv3, and YOLOv5s, the proposed method demonstrated superior accuracy and mAP while maintaining a reasonable parameter count.

Limitations and Future Work

The authors acknowledge that while their model significantly improves detection accuracy, further work is needed to enhance real-time detection speed. They also suggest that future research will focus on further improving detection accuracy. The current study relies on a custom dataset, and while it's representative, broader datasets could further validate the model's generalizability.

Summary updated

BibTeX
@misc{wen2024improved,
  author = {Yu Wen and Xiaorong Gao and Lin Luo and Jinlong Li},
  title = {{Improved YOLOv8-Based Target Precision Detection Algorithm for Train Wheel Tread Defects}},
  year = {2024},
  doi = {10.3390/s24113477},
}
Figure