基于边缘感知深度残差网络的带钢表面缺陷检测

Boundary-Aware Deeply Residual Network for Salient Object Detection of Strip Steel Surface Defects

基于深度学习的显著目标检测方法已被用于带钢表面缺陷检测中,但仍存在模型收敛速度慢、检测结果边缘不清晰等问题。针对现有问题,本文提出了基于边缘感知深度残差网络(boundary-aware deeply residual network, BADRNet),以此进行带钢表面缺陷的显著目标检测。将边缘信息引入至缺陷检测任务中,解决了因目标尺寸多样性带来的检测结果边缘不清晰的问题;通过在边缘提取、显著特征融合部分采用具有残差结构的3个卷积层作为基本块,提高了训练效率且保持原有的检测精度不变。在公开的SD-saliency-900数据集上的实验结果表明,所提模型相比于现有模型,在6个评价指标上均取得了最优效果。BADRNet比当前最优的EDRNet在S-measure指标上提升了1.6%,同时对于缺陷区域边缘的检测效果具有明显提升。

Deep learning-based salient object detection has been used in strip steel surface defects,but there are still some problems such as slow model training speed and unclear boundary of detection results.To address these issues,we proposed a boundary-aware deeply residual network(BADRNet)for salient object detection of strip steel surface defects.Boundary features are introduced into the steel surface defects to solve the problem of unclear boundary of detection results caused by varying object sizes.Three convolution layers with residual structure are used as basic blocks for boundary extraction and salient feature aggregation,improving training efficiency while maintaining original detection accuracy.Experimental results on the public strip steel benchmark dataset,SD-saliency-900,show that our model outperforms existing models in all six evaluation indicators.The proposed BADRNet improves the S-measure performance by 1.6%,and significantly enhances the detection effect on the defect area.