TL;DR
This paper introduces YOLO-LWTD, a novel lightweight algorithm for railway track feature detection. By integrating StarNet for enhanced feature extraction, an Extended Path Aggregation Network (EPAN) for efficient multi-scale fusion, and a simplified detection head (Detect-LADH), the model achieves improved accuracy and significantly faster inference speeds with a drastically reduced model size, making it ideal for embedded and real-time applications.
What Problem Does This Paper Address?
Intelligent operation and maintenance of rail transportation are increasingly reliant on accurate track inspection. However, existing methods often suffer from high computational loads, slow real-time performance, and large model sizes, hindering their deployment on resource-constrained embedded systems. A key challenge is the cumulative error in mileage-based measurements, which can lead to significant deviations in identifying track component positions and assessing their condition. While real-time error correction mechanisms exist, they depend on precise track feature detection for mileage calibration. Therefore, there's a critical need for lightweight, efficient, and accurate track feature detection algorithms that can operate effectively in real-world railway environments.
Methods
The proposed YOLO-LWTD algorithm builds upon the YOLO11n architecture, introducing several key modifications:
- StarNet Backbone: The original backbone network is replaced with StarNet. StarNet utilizes element-wise multiplication operations to create a mapping from low-dimensional to high-dimensional, non-linear feature spaces. This approach enhances feature characterization and discriminative ability without increasing network width, contributing to a more lightweight model.
- Extended Path Aggregation Network (EPAN) Neck: The standard PANet neck is optimized into an EPAN. This involves introducing additional feature-processing modules and a cross-layer jump connection, inspired by residual structures. EPAN aims to improve the fusion of multi-scale features, particularly enhancing the capture of detailed low-level features and global context information, which was a limitation in previous PANet implementations for track inspection.
- C3K2-Light Module: A lightweight version of the C3K2 module, incorporating Partial Convolution (PConv), is used. PConv selectively processes input channels, reducing computational complexity and FLOPs while maintaining feature extraction capabilities. This module is integrated into both the EPAN neck and the detection head.
- Detect-LADH Head: A lighter and more efficient detection head, Detect-LADH, replaces the original decoupled head. It employs an asymmetric multi-level compression strategy and depthwise separable convolutions (DWConv) to reduce computational complexity and parameter count while improving detection accuracy and avoiding task interference between classification and regression.
Datasets
The study utilized a self-built dataset specifically tailored for railway inspection.
- Custom Dataset: Images were acquired from a ballasted track field environment using a camera mounted on a track inspection train at a 45° angle to focus on fasteners and sleepers. The dataset consists of 566 original images (1408 × 1024 pixels).
- Data Augmentation: To enhance dataset robustness and cover diverse scenarios, five augmentation techniques were applied: horizontal flip, luminance adjustment (-30% to +30%), Gaussian noise injection, random rotation, and atmospheric interference simulation (fogging effect).
- Annotated Dataset: After augmentation, the dataset comprised 3396 images, with 3396 instances of sleepers and 3396 instances of fasteners. This dataset was split into training (70%), validation (10%), and test (20%) sets.
Key Insights and Ideas
- Element-wise Multiplication for Lightweight Feature Enhancement: The integration of StarNet's element-wise multiplication demonstrates that complex feature interactions can be achieved without significantly increasing model size, offering a powerful technique for lightweighting backbones.
- EPAN's Superior Multi-Scale Fusion for Track Scenes: The EPAN architecture effectively addresses the limitations of traditional PANet in fusing low-level and high-level features, which is crucial for accurately detecting track components under varying conditions.
- PConv's Efficiency Gains: The C3K2-Light module, leveraging PConv, highlights the effectiveness of selective channel processing in reducing computational load while maintaining performance, a key enabler for real-time applications.
- Asymmetric Head for Task Specialization: The Detect-LADH head's asymmetric design and use of DWConv show promise in optimizing detection heads for specific tasks like track inspection, balancing accuracy and efficiency.
Key Results
The YOLO-LWTD model demonstrated significant improvements over the baseline YOLO11n:
- Accuracy: Achieved a mean average precision (MAP) of 84.9%, outperforming the baseline YOLO11n (84.1%) by 0.8 percentage points. Precision (P) and recall (R) also saw improvements, reaching 98.2% and 73.0%, respectively.
- Lightweighting: The model size was reduced to a remarkable 1.5 MB, representing a 71.1% lightweight rate compared to the original model.
- Speed: Inference speed increased to 163 FPS, a 38.1% improvement over the baseline.
- Comparative Performance: In comparison tests with other state-of-the-art models (YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv12, YOLOv13, and RT-DETR), YOLO-LWTD achieved competitive accuracy while offering significantly lower model size and higher inference speeds.
- Backbone and Neck Efficiency: Ablation studies confirmed StarNet's efficiency in terms of parameters, model size, and speed, and EPAN's superiority in balancing accuracy and overall efficiency for multi-scale feature fusion.
Limitations and Future Work
- Dataset Limitations: The authors acknowledge that the custom dataset, while tailored, may still have limitations in covering all possible track scenarios. The lack of a widely recognized public benchmark dataset for track features presents a challenge for comprehensive evaluation and generalization.
- Future Directions:
- Dataset Expansion: Continuously expanding the scale and diversity of datasets and building an open-access benchmark dataset for track features to foster collaborative development.
- Rigorous Validation: Applying more rigorous statistical methods, such as cross-validation, to further validate model generalization capabilities on larger datasets.
- Heterogeneous Challenges: Addressing complex railway inspection scenarios like dynamic environmental changes, extreme weather, and high-speed perspectives through robust perception and intelligent recognition technologies.
- Multi-Source Fusion: Exploring deep integration and collaborative analysis of multi-source heterogeneous sensing methods (e.g., machine vision, 3D laser scanning, infrared imaging, radar, acoustic detection, inertial units) to develop next-generation intelligent track inspection systems.
Summary updated