基于多尺度特征融合与分块注意力的齿轮表面缺陷分割算法

打开文本图片集

关键词：表面缺陷检测；图像分割；UNet;分块注意力；齿轮中图分类号：TP394.1；TH132.429 文献标识码：Adoi：10.37188/0PE.20253322.3536 CSTR：32169.14.OPE.20253322.3536

Abstract：To address the limitations of traditional segmentation models in handling complex background interference and subtle defect regions of gears-particularly their insuficient feature-representation capability and poor robustness—this paper proposed a novel segmentation network based on multi-scale feature fusion and block-wise atention，aiming to enhance the representation of gear visual features and improve the detection performance of fine defects. First，a multi-scale feature-enhancement module replaced the standard downsampling blocks in the UNet encoder； it leveraged a parallel multi-branch convolutional structure to collaboratively extract multi-scale and multi-directional features，thereby enhancing the model' s perception of both local details and global context.Second，a block-wise feature-focusing module was introduced after downsampling; it employed a block-wise multi-head attention mechanism to independently analyze local regions，significantly improving sensitivity to minute defects and local texture variations. Finally，a weighted hybrid loss function was designed by combining Dice loss，binary cross-entropy （BCE） loss，and a gradient-diference constraint，efectively mitigating the class-imbalance issue and optimizing the quality of segmentation boundaries.Experimental results on both a self-constructed and public gear defect dataset demonstrate that the proposed method outperforms UNet and other state-of-the-art models in various gear defect segmentation tasks，achieving accuracy rates of 91.27% and 85.88% ，respectively. The results validate the effectiveness and superiority of the proposed approach for precise detection and segmentation of surface defects in gears.

Key words： surface defect detection; image segmentation; UNet; block-wise attention； gear

1引言

随着智能制造的持续深入与产业升级，工业领域对机械部件的质量要求日益严苛。（剩余15922字）