A lightweight algorithm utilizing improved YOLOv8 for detecting surface defects on steel has been presented, addressing significant challenges inherent to traditional methods. This innovative approach leverages cutting-edge advancements such as GhostNet, MultiPath Coordinate Attention (MPCA), and Simplified IoU (SIoU) to improve both the accuracy and efficiency of defect detection.
Steel is one of the most prevalent materials used across various industries, fundamentally supporting manufacturing and construction processes. Despite the advancements made over the years, imperfections such as cracks, scratches, and folds can often go undetected, leading to potential safety hazards and economic loss. Existing approaches typically involve manual inspection, which is notoriously inefficient and subjective, resulting in errors and missed defects.
The research team, comprising Ma, Zhao, Wan, and colleagues from Northeastern University, sought to develop a solution to these longstanding issues. They implemented an enhanced YOLOv8 model, which stands out for its ability to detect defects effectively even under challenging operational constraints. By integrating GhostNet as the backbone of their neural network, the model reduces overall parameters and associated computational complexity. This optimization is key, as it allows for deployment on systems with limited processing capabilities.
The strategic use of the MPCA attention mechanism enhances the model's capacity to extract significant features from varying scales within images, which is particularly advantageous for identifying defects of different sizes. The MPCA mechanism efficiently deals with complex backgrounds and lighting conditions, ensuring higher accuracy by focusing on the relevant parts of the image.
To bolster the model's performance, the researchers replaced the traditional CIoU loss function with SIoU. This new loss function considers the directional discrepancies between predicted and actual defect frames, enhancing regression performance and optimizing detection efficacy. The authors remarked, "This integration strikes a balance between detection accuracy and speed, offering a novel solution..."
Results from the experiments show remarkable improvements: the YOLOv8n algorithm experienced a 37% reduction in calculations and achieved an increase of 1.2% in mean average precision (mAP). Notably, the model operates with only 2.04 million parameters, making it substantially lighter than previous iterations, increasing its viability for real-time applications.
The experiments were conducted using the NEU-DET dataset—a widely recognized benchmark containing six types of steel surface defects, with each category represented by 300 images. The evaluation metrics included mean average precision (mAP), parameters, recall, precision, and floating-point operations per second (GFLOPS), ensuring rigorous assessment of the proposed model.
Further comparative tests indicated the model's superior performance against other traditional methods, achieving the best performance metrics with mAP of 0.786 and 171.5 frames per second (FPS). This positions the improved YOLOv8n model as not only accurate but also efficient—critical factors for real-world applications and deployment.
Importantly, the overall results point to the method's robustness and adaptability, capable of meeting the demands of environments where computational resources are constrained, such as embedded systems and mobile devices. The approach effectively balances the need for lightweight design with the accuracy necessary for successful defect detection.
Going forward, researchers aim to refine this methodology, particularly focusing on enhancing the detection accuracy of cracks and ensuring broad applicability across different defect types. The continued exploration of AI-driven solutions like this reflects the promising intersection of industrial practices and advanced machine learning techniques, paving the way for smarter and more reliable manufacturing processes.