基于模型自优化的带钢表面缺陷语义分割研究

Research on Semantic Segmentation of Strip Steel Surface Defects Based on Model Self-Optimization

针对当前带钢表面缺陷检测方法存在着检测速度低、检测判别特征提取不充分以及模型人工调参主观性强等技术性瓶颈问题,开展了基于模型自优化的带钢表面缺陷语义分割方法研究。在模型优化上,使用实体卷积替代原有的膨胀卷积,解决了边缘伪成像的问题,并且使用轻量化通道注意力机制模块,捕获了通道之间的依赖关系。构建了基于智能优化算法的关键超参数优化策略,使用改进全局搜索能力的麻雀搜索算法对模型整体的超参数组合进行寻优,选择影响效果最好的超参数,最终实现了自适应优化的带钢缺陷检测功能。在东北大学热轧带钢表面缺陷数据集上进行了实验,通过实验验证了该方法对夹杂、斑点和划伤等表面缺陷自动提取的可行性和有效性,满足了低配置、高性能的检测需求。

Aiming at the technical bottlenecks of the current strip surface defect detection methods,such as low detection speed,insufficient detection and discriminative feature extraction,and strong subjectivity of model manual parameter adjustment,this paper carried out a research on the semantic segmentation method of strip surface defects based on model self-optimization.In terms of model optimization,this paper uses entity convolution to replace the original dilated convolution to solve the problem of edge false imaging and uses a lightweight channel attention mechanism module to capture the dependencies between channels.This paper also builds a key hyperparameter optimization strategy based on the intelligent optimization algorithm,uses the sparrow search algorithm that improves the global search ability to optimize the overall hyperparameter combination of the model,selects the hyperparameters with the best effect,and finally realizes the adaptive optimization of the function of strip defect detection.In this paper,experiments are carried out on the data set of surface defects of hot-rolled strip steel from Northeastern University,and the feasibility and effectiveness of the method for automatic extraction of surface defects such as inclusions,spots and scratches are verified by experiments,which can meet the detection requirements of low configuration and high performance.