Steel Surface Defect Detection with the Efficient-Net Backbone in YOLOv8

Steel surface defect detection remains a critical challenge in industrial quality control systems, where traditional inspection methods often fail to achieve the precision required for real-time manufacturing environments. This paper presents a new YOLOv8-based architecture that integrates Efficient-Net-B6 as the backbone network to improve steel surface defect detection capabilities. The proposed framework incorporates Mobile Bottleneck Convolution (MBConv) blocks, multiscale feature fusion pathways, and optimized C2f modules to enhance feature extraction efficiency even though conventional single-stage detectors struggle with small defect recognition. Despite the complexity of industrial environments, the modified architecture demonstrates superior performance across multiple defect categories. Experimental validation in the NEU-DET and GC10-DET datasets reveals substantial improvements over the baseline models, including YOLOv8, YOLOv9, YOLOv10, and YOLOv11. For the NEU-DET dataset, the enhanced model achieves a 99,5% value in mAP@50 and a 99% value in mAP@0.50:0.95, while for the GC10-DET dataset the improved YOLOv8 shows a 99.4% value in mAP@0.50 and a 94.6% value score in mAP@0.50:0.95. The integration of Squeeze-and-Excitation (SE) attention mechanisms within the MBConv blocks enables selective feature emphasis, proving particularly effective for detecting minute surface anomalies. These results establish our YOLOv8-EfficientNet framework as a viable solution for automated steel surface inspection systems.