摘要
针对烧结钕铁硼产品表面缺陷检测中的复杂背景干扰问题,提出了一种高效的检测方法FFA-YOLO。首先,在骨干网络模块中引入FFA-Net对图像进行去噪处理;其次,在颈部网络层中使用SEAM模块,通过深度可分离卷积学习空间维度与通道的相关性,从而提升在复杂背景下的检测性能。最后,提出了C2f_DSConv模块,自适应聚焦于细长及曲折的局部结构,提高了分割图像的精确性。实验结果表明,该模型在烧结钕铁硼数据集上较基准模型的mAP@0.5和mAP@0.5∶0.95分别提升了3%和6.6%,模型大小为49.6 MB,GFLOPs为78.7,在保持基准模型轻量化的同时,显著提高了检测性能。
FFA-YOLO, an efficient detection method, addresses the challenge of complex background interference in surface defect detection for sintered NdFeB products. First, FFA-Net is introduced in the backbone network module to denoise the images. Second, the SEAM module is applied in the neck network layer, where depthwise separable convolution is employed to learn the correlation between spatial dimensions and channels, thus improving detection performance under complex backgrounds. Finally, the C2f_DSConv module is proposed, which adaptively focuses on elongated and tortuous local structures to enhance the accuracy of image segmentation. Experimental results demonstrate that, compared to the baseline model, this model improves mAP@0.5 and mAP@0.5∶0.95 by 3% and 6.6%, respectively, on the sintered NdFeB dataset. The model size is 49.6 MB, and the GFLOPs is 78.7, achieving a significant improvement in detection performance while maintaining the lightweight parameters of the baseline model.
作者
李俊峰
陈谦
郝京波
杜阳
梁真
LI Junfeng;CHEN Qian;HAO Jingbo;DU Yang;LIANG Zhen(Advanced Technology and Materials Co.,Ltd.,Beijing 100081,China;AT and M North Technology Co.,Ltd.,Baotou 014000,Inner Mongolia,China)
出处
《金属功能材料》
2025年第3期91-96,共6页
Metallic Functional Materials
基金
高丰度稀土元素高综合性能永磁材料项目(TC220H06G)。
作者简介
李俊峰(1984-),男,大学本科,高级工程师,E-mail:lijunf@atmcn.com;通信作者:陈谦(1997-),男,硕士,E-mail:chenqian@atmcn.com。
