Multi-task guided adaptive multi-scale feature fusion method for surface defect segmentation in cold-rolled strips
LUO Xiaoyi
ZHANG Feiyu
DU Wangzhe
HAN Jianchao
LI Xuwei
XU Wei
Abstract：Surface defect detection of cold-rolled strips is crucial for industrial quality control.To address challenges including small defect omission,class imbalance,and low efficiency in ultra-high-resolution image processing,this study proposes Multi-scale Feature Adaptive Fusion-based Multi-task Segmentation Network(MFAFMSNet),an enhanced architecture integrating multi-scale context aggregation and attention-guided feature refinement.The method features:1)A residual network-based encoder with multi-dilation-rate spatial pyramid modules for cross-scale context fusion;2)A decoder incorporating multi-scale squeeze-and-excitation attention to enhance small defect detection;3)A multi-task framework with auxiliary defect classification to improve inter-class discrimination;4)A dynamic weighted hybrid Loss combining Dice and Focal Losses with adaptive weight adjustment for class imbalance mitigation.Experimental results on a constructed cold-rolled strip defect dataset show the model achieves state-of-the-art performance,with mIoU and Dice coefficients surpassing baseline models by 5.9％and 5.8％respectively.
Keywords：defect segmentationmulti-scale feature fusionattention mechanismdeep learningmulti-task learning
Publication Date：2025-07-20
Online Publishing Date：2025-09-08(First online date of this platform, not the publication date of the document)
Pages：7( 7-13 )