广西科技大学学报2025，Vol.36Issue(5)：65-72,8.DOI:10.16375/j.cnki.cn45-1395/t.2025.05.009
面向轻量化网络的硅钢片缺陷检测算法
Detection algorithm for silicon steel sheet defects oriented towards lightweight networks
摘要
Abstract
To tackle the problems of low accuracy,sluggish speed,and demanding computational power,and difficulty in model deployment in industrial silicon steel surface defect detection,a lightweight object detection network,YOLOv7-GSS,was proposed.This algorithm was based on YOLOv7-tiny as the foundation model.Initially,the ELAN module of its backbone network was reconstructed by introducing lightweight GSConv convolution and parameter-free SimAM attention mechanism,forming the ELAN-GSS module to reduce model complexity while maintaining accuracy.Subsequently,the improved model performed channel pruning to eliminate surplus channels and decrease model parameters for enhanced lightweight characteristics.Finally,knowledge distillation was applied to assist in training the pruned network to improve accuracy.Experimental results prove that in comparison to the YOLOv7-tiny,the YOLOv7-GSS network achieves a 1.564%increase in mAP@0.5,an 80.399%reduction in parameters,a 79.389%decrease in floating-point operations per second,and a 44.751%increase in FPS.In comparison to other object detection models,the YOLOv7-GSS model strikes a good balance between detection speed and accuracy,meeting the real-time and accurate detection requirements for silicon steel plates in production workshop scenarios.This model provides favorable conditions for embedding into mobile devices.关键词
缺陷检测/YOLOv7-tiny/注意力机制/剪枝/知识蒸馏Key words
defect detection/YOLOv7-tiny/attention mechanism/prune/knowledge distillation分类
计算机与自动化引用本文复制引用
李克讷,陈福丁,李永革,陈健民..面向轻量化网络的硅钢片缺陷检测算法[J].广西科技大学学报,2025,36(5):65-72,8.基金项目
国家自然科学基金项目(61663003) （61663003）
广西科技大学博士基金项目(院科博12Z05)资助 （院科博12Z05）
