Abstract
To improve the precision of defect categorization and localization in images, this paper proposes an approach for detecting surface defects in hot-rolled steel strips. The approach uses an improved YOLOv5 network model to overcome the issues of inadequate feature extraction capacity and suboptimal feature integration when identifying surface defects on steel strips. The proposed method achieves higher detection accuracy and localization precision, making it more competitive and applicable in real production. Firstly, the multi-scale feature fusion (MSF) strategy is utilized to fuse shallow and deep features effectively and enrich detailed information relevant to target defects. Secondly, the CSPLayer Res2Attention block (CRA block) residual module is introduced to reduce the loss of defect information during hierarchical transmission, thereby enhancing the extraction of fine-grained features and improving the perception of details and global features. Finally, the experimental results indicate that the mAP on the NEU-DET and GC10-DET datasets approaches 78.5% and 67.3%, respectively, which is 4.9% and 2.1% higher than that of the baseline. Meanwhile, it has higher precision and more precise localization capabilities than other methods. Furthermore, it also achieves 59.2% mAP on the APDDD dataset, indicating its potential for growth in further domains.
Introduction
In industrial steel manufacturing, identifying surface defects such as crazing, inclusions, or rolled-in scale is a critical yet challenging task. These defects are often small, irregular in shape, and frequently overlap, making them difficult for standard computer vision models to detect reliably. While deep learning has revolutionized defect detection, standard Convolutional Neural Networks (CNNs) often face a trade-off: as networks deepen to understand complex patterns, they lose the fine-grained, high-resolution details found in earlier layers. Missing a tiny crack can lead to catastrophic structural failures later in the supply chain. This research proposes an enhanced YOLOv5s model designed to capture both global context and minute details without excessive computational cost.
Methodology
The researchers selected YOLOv5s as their baseline due to its lightweight architecture and real-time processing capabilities. They introduced two primary architectural innovations to address the specific challenges of steel surface inspection.
Multi-Scale Feature Fusion (MSF)
Standard object detection models often ignore the "P2" layer—the initial high-resolution feature map—because it is computationally expensive to process. This paper explicitly fuses the P2 layer back into the detection heads. This strategy ensures that shallow details, such as sharp edges and tiny spots, are preserved and combined with deep semantic information, preventing the loss of critical fine-grained features during hierarchical transmission.
CSPLayer Res2Attention (CRA) Block
The CRA block is a custom residual module that replaces standard bottlenecks. It integrates three key concepts:
- Res2Net-style Splitting: Feature channels are split into smaller groups and processed hierarchically, creating a multi-scale receptive field within a single layer.
- CBAM (Convolutional Block Attention Module): This applies both channel and spatial attention, directing the model to focus on the most relevant parts of the image and specific features while filtering out industrial noise like steam or oil spots.
- CSP (Cross Stage Partial) Structure: This maintains efficiency by reducing redundant gradient information.
Experimental Results
The model was rigorously tested on three distinct datasets: NEU-DET (1,800 grayscale images), GC10-DET (2,300 images with complex backgrounds), and APDDD (Aluminum Profile Surface Detection Database) to test scalability.
Performance Metrics
The improved model significantly outperformed the baseline YOLOv5s and other state-of-the-art architectures:
- NEU-DET: Achieved 78.5% mAP, a 4.9% increase over the baseline.
- GC10-DET: Reached 67.3% mAP, a 2.1% improvement.
- APDDD: Achieved 59.2% mAP, demonstrating the model's potential for cross-domain application in other metal manufacturing processes.
While the added architectural complexity reduced the frame rate from 112 FPS to 54 FPS on the NEU-DET dataset, the performance remains well above the 30 FPS threshold required for real-time industrial monitoring.
Conclusion
By re-integrating high-resolution features from the P2 layer and utilizing the multi-scale receptive fields of the CRA block, the proposed model provides a robust solution for industrial defect detection. It effectively filters out background noise and captures the minute details necessary for high-precision steel inspection. Future work will focus on field deployment in live factory environments and using data augmentation to address unbalanced defect classes.
Summary updated