针对钢铁连铸产线板坯检测精度低和速度慢等问题,提出了一种结合机器视觉、图像处理和深度学习的 DMS-YOLOv8(YOLOv8 with depthwise separable convolution,multi-pooling,and SE(squeeze-and-excitation)-EMA(exponential moving average))算法.该算法通过构造基于反向残差和多尺度池化的深度可分离卷来替换标准卷积,减轻了冗余网络的负担,降低了内存使用,并提高了计算速度;通过混合注意力机制SE-EMA使模型在处理输入数据时可以有选择地关注和加权不同部分的信息,提高了模型的表达能力;最后,通过在自制的板坯数据集及PASCAL VOC2012数据集上进行对比分析及消融实验,验证了本文所提方法在实时运行工况下能有效提升板坯检测精度,并保证一定的快速性,为后续热轧流程的高效生产奠定了基础.
DMS-YOLOv8 slab detection algorithm based on improved depthwise separable convolution and hybrid attention mechanism
To address the issues of low detection accuracy and slow speed in the inspection of slab on steel continuous casting production lines,an algorithm named DMS-YOLOv8(YOLOv8 with depth-wise separable convolution,multi-pooling,and SE-EMA)that combines machine vision,image pro-cessing,and deep learning was proposed.This algorithm replaces standard convolution with depthwise separable convolution based on reverse residual and multi-scale pooling,reducing the burden of re-dundant networks,decreasing memory usage,and improving computational speed.The mixed attention mechanism SE-EMA allows the model to selectively focus on and weight information from different parts when processing input data,enhancing the model's expressive power.Finally,by conducting comparative analysis and ablation experiments on the self-made slab dataset and the PASCAL VOC2012 dataset,the study validates that the proposed method can effectively enhance slab detection accuracy in real-time operating conditions while ensuring a certain level of efficiency,laying the foun-dation for efficient production in subsequent hot rolling processes.
NETL
NSTL
万方数据
