基于YOLOv8-SDB的石化储罐焊缝表面缺陷识别算法

YOLOv8-SDB-based algorithm for identifying surface defects in petrochemical storage tank welds

摘要:

针对目前石化储罐表面焊缝缺陷检测效率低、小目标易出现漏检、误检等问题，提出一种改进YOLOv8n的焊缝缺陷检测方法：YOLOv8-SDB。首先，在主干网络中引入空间深度转换卷积(symmetric positive definite convolution,SPDConv)模块，减少浅层特征提取过程中细节信息的丢失，捕获更加丰富的空间和通道信息；其次，利用加权双向特征金字塔网络（bi-directional feature pyramid network,BiFPN）融合多层次特征，提高特征表达能力；再次，采用轻量化且高效的上采样算子DySample（dynamic sampling），提高模型特征重建能力并减少计算复杂度；最后，使用WIoU(weighted intersection over union)损失函数加快边界回归损失收敛速度，提高回归精度。将改进后的算法在焊缝缺陷数据集上进行实验，实验结果表明，YOLOv8-SDB算法的检测准确率为86.2%，召回率为79.4%，平均精度为84%。较YOLOv8n算法分别提高了3.4%、2.8%和3.9%。

Abstract:

At present, the petrochemical storage tank surface weld defect detection efficiency is low, and small targets are prone to leakage and misdetection. To address these problems, this paper proposes an improved weld defect detection method for YOLOv8n: YOLOv8-SDB.Firstly, the SPDConv(symmetric positive definite convolution) module is introduced into the backbone network to reduce the loss of detail information in the shallow feature extraction process and capture richer spatial and channel information;Secondly, BiFPN (weighted bi-directional feature pyramid network) is utilized to fuse multi-level features and improve the feature expression capability; Again, a lightweight and efficient upsampling operator, DySample (dynamic sampling), is used to improve the model feature reconstruction capability and reduce the computational complexity;Finally, the WIoU (wise intersection over union) loss function is employed to accelerate the convergence of the boundary regression loss and improve the regression accuracy.The improved algorithm is experimented on the weld defect dataset, and the experimental results show that the detection precision of the YOLOv8-SDB algorithm is 86.2%, the recall rate is 79.4%, and the average precision is 84%. This is an improvement of 3.4%, 2.8% and 3.9% over the YOLOv8n algorithm, respectively.