EAD-YOLOv10: Lightweight Steel Surface Defect Detection Algorithm Research Based on YOLOv10 Improvement

Abstract

In response to the issues of low detection accuracy (DA), slow speed, and missed detections caused by the complex texture background and diverse shapes of surface defects (SD) in steel, this paper designs an improved lightweight YOLOv10 model called EAD-YOLOv10. By incorporating an Adaptive Downsampling (Adown) structure to replace traditional convolutional layers, this module reduces the model’s computational complexity and accelerates inference speed while enhancing the retention of semantic information through a lightweight design. A Dynamic Upsample (DySample) module is embedded in the neck network, which adaptively adjusts sampling weights based on the occurrence frequency of various samples in the dataset, improving the model’s recognition accuracy for minority class samples. Finally, the designed C2f_EMSCP method is integrated into the backbone and neck networks, effectively merging multi-scale convolutional networks and position-aware modules to enhance the model’s sensitivity and detection capability for multi-scale targets by fusing feature maps of different scales. The experimental results show that the improved EAD-YOLOv10 network model achieves an average precision of 94.2% for detecting six types of defects in the NEU-DET dataset, which is an improvement of 7.6% over the baseline model, with a 9.75% reduction in model size and a 12.5% decrease in computational load, outperforming other mainstream object detection algorithms and meeting the requirements for SD detection in industrial production. Additionally, generalization experiments on actual industrial defect datasets confirm that this algorithm has good generalization capabilities.