TL;DR
This paper introduces SE-ResNet18, a novel lightweight deep learning model for detecting surface defects in aluminum-based composite core conductors (ACCC) using X-ray images. By adapting ResNet18 for single-channel X-ray data and integrating a Squeeze-and-Excitation (SE) attention mechanism, the model achieves a remarkable 98.39% accuracy with real-time inference speeds, addressing key challenges in industrial defect detection.
What Problem Does This Paper Address?
The reliable detection of surface defects in Aluminum-based Composite Core Conductors (ACCC) is crucial for the safety and efficiency of power transmission grids. These defects, such as splitting, fracture, sawing, and displacement, can compromise the conductor's mechanical integrity and lead to catastrophic failures. However, existing X-ray-based defect detection methods face several hurdles:
- Scarcity and Imbalance of Defect Data: Obtaining large, balanced datasets of rare defect types is challenging, limiting the generalization capabilities of traditional models.
- Real-time Performance Demands: Industrial online inspection requires rapid inference, which is often not met by high-precision but computationally expensive models.
- Single-Channel X-ray Images: X-ray images are grayscale (single-channel), whereas many standard deep learning models are designed for RGB (three-channel) images. Directly applying these models requires inefficient channel conversion and can lead to suboptimal feature representation.
This paper aims to overcome these limitations by developing a lightweight, accurate, and efficient model specifically tailored for single-channel ACCC X-ray defect detection.
Methods
The core of this work is the SE-ResNet18 model, a custom-designed architecture built upon the ResNet18 backbone. Key methodological innovations include:
- Single-Channel Input Adaptation: Instead of converting grayscale X-ray images to RGB, the ResNet18's initial convolutional layer is modified to accept a single input channel (7x7x1x64 kernel). This preserves native image information, avoids computational waste, and prevents feature dilution.
- Integrated Squeeze-and-Excitation (SE) Attention Mechanism: SE modules are embedded within each ResNet18 residual block. These modules adaptively recalibrate channel-wise feature responses, amplifying subtle defect features while suppressing irrelevant background noise. This is crucial for detecting small, low-contrast defects.
- Lightweight Architecture: The model leverages the inherent efficiency of ResNet18 and a streamlined output layer (replacing the ImageNet-specific head with a Global Average Pooling and a small fully connected layer) to keep the parameter count low (12.0 million).
- Targeted Data Augmentation and Training Strategies: To combat data scarcity and class imbalance, the study employs various augmentation techniques (flipping, rotation, scaling, etc.) with specific emphasis on the "Splitting" defect. A robust training strategy involving AdamW optimizer, learning rate scheduling (warm-up + cosine annealing), gradient clipping, and early stopping is also utilized.
Datasets
The study utilizes a publicly released X-ray scan dataset of ACCC conductors collected from industrial processes. This dataset was further processed for quality screening, inter-frame redundancy removal, and category balancing.
- ACCC X-ray Scan Dataset: This dataset comprises 2488 images featuring five categories: No Defect (Class_1), Splitting (Class_0), Sawing (Class_2), Fracture (Class_3), and Displacement (Class_4).
- Data Split: The dataset was partitioned into training (1741 images), validation (249 images), and test (498 images) sets at a 7:1:2 ratio.
Key Insights and Ideas
- Synergy is Key: The paper strongly emphasizes that the model's success stems from the orchestrated integration of single-channel adaptation, SE attention, and optimized training, rather than a single novel component.
- Single-Channel Processing is Not Just Efficient, It's Better: Directly processing single-channel X-ray data, rather than converting to RGB, not only saves computation but also preserves crucial information that might be lost or diluted in the conversion process.
- Attention for the "Hard Cases": The SE attention mechanism proves particularly effective in boosting the recall for challenging, small-sample defects like "Splitting," which are often missed by models without such fine-grained feature amplification.
- Lightweight Doesn't Mean Low Accuracy: The SE-ResNet18 demonstrates that it's possible to achieve state-of-the-art accuracy with a significantly smaller model and faster inference, making it practical for industrial deployment.
Key Results
- High Accuracy: The SE-ResNet18 model achieved an impressive 98.39% accuracy on the test set.
- Excellent Macro-F1 Score: A Macro-F1 score of 0.996 indicates strong performance across all defect categories, effectively addressing the challenge of rare defects.
- High Recall for Difficult Defects: The recall for the "Splitting" defect category (Class_0) was significantly improved from 88.2% (baseline ResNet18) to 94.0%.
- Lightweight and Fast: The model has only 12.0 million parameters and achieves an inference speed of 0.44 ms/image (over 2200 images/second), making it suitable for real-time industrial applications.
- Robustness: The training and validation curves showed stable convergence with no significant overfitting, indicating effective generalization.
- Comparison: SE-ResNet18 outperformed other lightweight models like SqueezeNet1.1 and AlexNet, and achieved a superior balance of accuracy and efficiency compared to high-performance models like Inception-ResNet-V2 and DenseNet121.
Limitations and Future Work
- Minute Defects: The model still struggles with identifying extremely minute defects (e.g., <3 pixels) due to their low signal-to-noise ratio and limited feature space representation. Future work could explore dual spatial-channel attention mechanisms (like CBAM) for better ultra-small defect capture.
- Single Perspective: The current approach uses a single X-ray view, which might miss oblique or concealed defects. Future research could investigate multi-view image acquisition and dual-branch networks to integrate information from different angles.
- Environmental Robustness: The model's performance in extreme industrial environments (e.g., high noise, surface contamination) needs further validation. Adversarial training and specialized datasets could enhance environmental adaptability.
- Lack of Microstructural Validation: The use of a public dataset prevented direct microstructural analysis (e.g., SEM) of defect samples. However, the defined defect categories are strongly correlated with established failure mechanisms.
- Classification Only: The current study focuses on defect type classification. Future work could integrate object detection and image segmentation to develop a defect severity grading model, moving from "detection" to "assessment."
Summary updated