Technical Summary: High-Accuracy Surface Defect Detection for Metal Strips via Lightweight YOLOv5s-SC2 with LAMP Pruning

1. Problem Statement

The paper addresses the critical challenges in industrial quality control for metal strips, specifically the detection of surface defects such as cracks, patches, scratches, bumps, and pits. These defects, caused by processing technologies or environmental factors, significantly impact product quality, corrosion resistance, and wear resistance.

While traditional non-destructive testing methods (e.g., magnetic leakage, ultrasonic detection) offer flexibility, they suffer from limitations such as material specificity, insensitivity to shallow pits, and leakage due to surface roughness. Machine vision based on deep learning has emerged as a superior alternative due to its high accuracy and adaptability. However, existing deep learning models face specific hurdles in this domain:

• Diverse Defect Types: Defects vary significantly in morphology (point-like, irregular).
• Scale Variations: Defects range from large patches to tiny scratches.
• Small Target Miss Rates: Standard models often fail to detect minute defects due to information loss in deep feature maps.
• Deployment Constraints: Industrial edge devices have limited computational resources, memory, and power, making the deployment of heavy, high-accuracy models difficult.

2. Methodology

The authors propose YOLOv5s-SC2, a lightweight, high-accuracy object detection model based on the YOLOv5s architecture, coupled with a structural compression strategy using Layer-Adaptive Magnitude-Based Pruning (LAMP).

2.1. Anchor Optimization and Loss Function

• K-means++ Clustering: Replaces the standard K-means algorithm to generate anchor boxes. K-means++ improves cluster center initialization, avoiding local optima and producing anchors better suited to the specific size and shape distribution of steel surface defects in the NEU-DET dataset.
• Shape-IoU Loss: Introduces the Shape-IoU loss function to replace standard IoU-based losses. Unlike traditional methods that focus solely on overlap, Shape-IoU explicitly accounts for the geometric properties (width and height) of the bounding box. This is crucial for defects with irregular shapes or extreme aspect ratios, guiding the model to predict boxes that align more closely with true defect geometries.

2.2. Backbone Network Improvements (YOLOv5s-SC)

The backbone is modified to enhance feature extraction and gradient flow while maintaining efficiency:

• C2f Module: Replaces the original C3 module. The C2f (CSP Bottleneck with 2 convolutions) module, inspired by ELAN (Efficient Layer Aggregation Network), utilizes a split-and-merge strategy with more layer-hopping connections. This enriches gradient flow and improves the retention of small target information that is often lost in deep networks.
• Switchable Atrous Convolution (SAC): Integrated to address the trade-off between receptive field and computational cost. SAC applies different atrous rates to the same input features and fuses them via a switching function. This allows the network to adapt to multi-scale features flexibly without increasing computational complexity, effectively capturing contextual information and preventing the loss of small target details during pooling.

2.3. Neck Network Enhancements (YOLOv5s-SC2)

The neck network is upgraded to improve semantic fusion and focus on critical regions:

• RepGFPN (Re-parameterized Generalized Feature Pyramid Network): Replaces the standard PANet. RepGFPN optimizes channel dimensions for different scales, removes redundant feature fusion operations, and simplifies up-sampling operators. It enhances interaction between high-level semantic and low-level spatial information without adding significant computational burden.
• SE Attention Mechanism: The Squeeze-and-Excitation (SE) module is integrated before feature fusion blocks. It performs a "squeeze" operation to aggregate global spatial information into channel descriptors and an "excitation" operation to generate channel-wise weights. This allows the model to selectively emphasize important defect features while suppressing irrelevant background noise.

2.4. Model Pruning for Edge Deployment

To enable deployment on resource-constrained devices, the authors apply LAMP Pruning:

• Mechanism: LAMP calculates a score for each weight based on its magnitude relative to other weights in the same layer, allowing for adaptive pruning rates.
• Structural Constraints: A specific filtering mechanism is implemented to skip modules containing split and concatenation operations (e.g., in C2f and RepGFPN) to prevent feature map mismatches. Pruning is applied only to independent convolutional layers.
• Fine-tuning: After pruning, the model undergoes fine-tuning to recover any minor accuracy losses.

3. Key Contributions

1. Optimized Anchor and Loss: The integration of K-means++ for anchor generation and Shape-IoU loss significantly improves bounding box regression accuracy for irregular and small defects.
2. Enhanced Backbone (YOLOv5s-SC): The combination of C2f and SAC modules enriches gradient flow and multi-scale feature extraction, achieving a baseline mAP of 82.4% (a 2.5% improvement over standard YOLOv5s).
3. Advanced Neck (YOLOv5s-SC2): The fusion of RepGFPN and the SE attention mechanism further boosts performance to 84.3% mAP, a 4.4 percentage point improvement over the baseline.
4. Lightweight Deployment Strategy: The application of LAMP pruning successfully compresses the model for edge devices. At a 2.5× acceleration setting, the model size is reduced to 56.2%, computational cost to 39.4%, and memory usage to 56.9% of the original, with a negligible mAP drop of only 0.2 percentage points.

4. Experimental Results

Experiments were conducted on the NEU-DET dataset, which contains 1,800 grayscale images of six defect types (rolled-in scale, patches, cracks, pits, inclusions, scratches).

• Performance: The final YOLOv5s-SC2 model achieved a mean Average Precision (mAP) of 84.3%, Precision of 82.4%, and Recall of 75.3%.
• Ablation Studies:
  • The C2f and SAC modules improved mAP to 82.4%.
  • Adding RepGFPN increased mAP to 83.3%.
  • The final addition of the SE attention mechanism yielded the peak performance of 84.3%.
  • Visualizations (heatmaps) confirmed that the proposed model focuses more accurately on defect regions and suppresses background noise compared to the baseline.
• Pruning Efficiency:
  • 2.0× Pruning: mAP 84.1%, Params 7.97M.
  • 2.5× Pruning: mAP 84.1%, Params 7.04M, GFLOPs 7.6. This was identified as the optimal balance.
  • 3.0× Pruning: mAP 83.8%, Params 6.48M.

5. Significance and Claims

The paper claims that the proposed YOLOv5s-SC2 offers an effective and lightweight solution for industrial surface defect detection. By systematically addressing the challenges of diverse defect types and scale variations through architectural improvements (C2f, SAC, RepGFPN, SE) and optimizing for deployment via LAMP pruning, the model achieves a superior balance between accuracy and efficiency.

The authors emphasize that the 2.5× pruned model provides a practical solution for resource-constrained edge devices, reducing computational costs to less than 40% of the original while maintaining detection accuracy that significantly outperforms the baseline YOLOv5s. This makes the approach suitable for real-time, online industrial inspection scenarios where both high precision and low latency are required.