宁夏师范学院学报

为了提高磁瓦表面缺陷检测的精度，提出基于YOLOv8n的磁瓦缺陷检测算法YOLOv8-OSE.采用全维度动态卷积模块替代颈部网络中的普通卷积，增强模型的特征提取能力，减少模型的计算量；其次，在颈部网络中加入置换注意力机制，通过结合通道和空间注意力，增强模型对小目标的检测效果，使模型更加精准地定位目标位置；最后，为了解决缺陷样本不平衡和宽高比问题，引入EIoU损失函数，提高模型的收敛速度和回归精度.在磁瓦数据集上的实验结果表明，YOLOv8-OSE算法的平均精度mAP@0.5达到87.6%,mAP@0.5∶0.95达到51.6%,相较于YOLOv8n算法，分别提升了4.5%和2.6%.同时，模型的参数量和计算量基本保持不变.与当前主流的检测算法相比，YOLOv8-OSE算法的检测精度具有一定优势，能够在保持低算力的同时满足高精度的部署需求.

In order to improve the accuracy of magnetic tile surface defect detection, a magnetic tile defect detection algorithm based on YOLOv8n is proposed.The primary contribution of this study lies in the omni-dimensional dynamic convolution module is used to replace the ordinary convolution in the neck network, thereby improving the model′s feature extraction capabilities and reduce computational load.Additionally, a shuffle attention mechanism is integrated into the neck network to improve the detection performance for small targets by combining channel and spatial attention, thus enhancing the model more accurate in locating the target positions.Finally, to address the issues of imbalanced defect samples and aspect ratios, the EIoU loss function is introduced to accelerate the convergence speed and improve regression accuracy of the model.Experimental results on the magnetic tile dataset show that the average precision mAP@0.5 of the YOLOv8-OSE algorithm reaches 87.6%,and mAP@0.5:0.95 reaches 51.6%,representing improvements of 4.5% and 2.6% respectively compared to the YOLOv8n algorithm.Meanwhile, the number of parameters and computational load of the model remain essentially unchanged.Compared to current mainstream detection algorithms, the YOLOv8-OSE algorithm demonstrates a certain advantage in detection accuracy, meeting high precision deployment requirements while maintaining low computational power.