TL;DR
This paper introduces YOLOv5-MRL, a lightweight object detection model designed for efficient and accurate meter pointer recognition in substations. By combining GhostNet for a lighter backbone and soft pruning for parameter reduction, the model achieves a significant reduction in size and inference time while maintaining high detection accuracy, making it suitable for deployment on edge devices.
What Problem Does This Paper Address?
Traditional object detection models, while accurate, are often too large and computationally expensive for deployment on resource-constrained devices like substation monitoring robots. This poses a challenge for real-time, automated meter reading in power systems, where efficiency and speed are crucial. Existing lightweight models have not been extensively explored for this specific application, and there's a need for a solution that balances detection accuracy with real-time performance and minimal model size.
Methods
The core of this work is the development of the YOLOv5-MRL model, which builds upon the YOLOv5 architecture. Key methodological contributions include:
- GhostNet Backbone: The standard CSPDarknet53 backbone of YOLOv5 is replaced with GhostNet. GhostNet uses a "ghost module" that generates more feature maps with fewer parameters and computations by dividing normal convolution into two parts: one for essential feature maps and another for generating "ghost" feature maps through depth-wise separable convolutions. This significantly reduces computational complexity and model size.
- Soft Pruning of Convolution Kernels: To further compress the model, a soft pruning (SFP) technique is applied to the convolutional kernels. Unlike hard pruning, SFP allows pruned kernels to participate in subsequent training iterations, helping to retain more of the network's expressive capacity and reduce the risk of overfitting. This process iteratively removes redundant kernels based on their L2 norm.
- Dial External Circle Fitting: For the final meter reading, a dial external circle fitting method is employed. This method, based on fitting a circle to the dial, uses the detected pointer and scale positions to calculate the angle between them and determine the meter reading. This approach is noted for its applicability to various meter types and its ability to achieve fast and accurate readings.
Datasets
The study utilizes a newly constructed "Lighting-Meter" dataset. This dataset comprises:
- 2000 lightning rod images.
- 32,000 labeled targets.
- Four types of images: JCQ-8 lightning rod detector, JCQ-8H lightning rod detector, JCQ3A lightning rod monitor, and JCQ3A-20/2000 lightning rod monitor.
- Labels include pointer position, pointer center, and various scale points (0 to 6).
- The dataset was split into a 7:3 ratio for training (1400 images) and testing (600 images).
- Data augmentation techniques such as scaling, flipping, panning, cropping, and color brightness transformation were used to enhance the dataset's size, diversity, and robustness.
Key Insights and Ideas
- Lightweighting is Crucial for Edge Deployment: The paper strongly emphasizes that for practical deployment on substation robots and other edge devices, model size and inference speed are paramount, often more so than marginal gains in accuracy.
- GhostNet and Soft Pruning Synergize Well: The combination of GhostNet for a lighter backbone and soft pruning for parameter reduction proved effective in achieving significant model compression without drastically sacrificing accuracy.
- Soft Pruning Preserves Network Capacity: The choice of soft pruning over hard pruning is highlighted as a key factor in maintaining the model's expressive power and robustness, preventing a severe drop in performance.
- Dial Fitting is Robust: The dial external circle fitting method is presented as a reliable post-detection step for accurately reading meter values, adaptable to different meter designs.
Key Results
- Model Size: The YOLOv5-MRL model achieved a remarkably small size of 5.5 MB, representing a 98.48% reduction compared to the original YOLOv5s model.
- Inference Speed: The model boasts an inference time of 5.0 ms/frame, meeting real-time detection requirements. This is a significant improvement over many other models, including the original YOLOv5s (29 ms).
- Detection Accuracy: The model achieved a mean average precision (mAP) of 96.9% on the self-built dataset. While slightly lower than some original YOLOv5s models, it is considered a good trade-off for the substantial gains in speed and size.
- Parameter Reduction: The number of parameters was reduced by 47.2% compared to the original YOLOv5s.
- Comparison: YOLOv5-MRL outperformed several other models in terms of a balance between size, speed, and accuracy, particularly for applications requiring lightweight solutions.
Limitations and Future Work
- Limited to Single Scene: The current YOLOv5-MRL model is optimized for a single scene (lightning rod meter reading).
- Anchor Parameters: The paper notes that the algorithm's reliance on anchor parameters might affect detection speed, suggesting room for optimization.
- Future Work:
- Optimize with YOLOv8: Explore integrating the latest YOLOv8 algorithm for further improvements.
- Expand Dataset: Incorporate more diverse data to enhance the generalization ability of the model for a generalized meter reading solution.
- Enhance Interpretability: Improve the interpretability of the algorithm to boost the robustness of the meter reading model.
Summary updated