基于改进YOLOv8曲轴表面缺陷检测算法

Based on the Improved YOLOv8 Crankshaft Surface Defect Detection Algorithm

针对曲轴表面小目标缺陷检测难度大、缺陷背景复杂和检测速度慢等问题,提出一种改进曲轴表面缺陷检测的算法RB-YOLOv8。首先,用RepViT模块取代了传统的C2f模块,有助于减少网络的计算负担并加快其运行速度;接着,通过优化双向特征融合模块BiFPN及增加小目标检测层,改善小目标缺陷识别的能力;然后,利用BiFormer注意力机制强化模型的抗干扰能力和解决缺陷背景复杂的难题,提高检测准确率;最后,使用MPDIoU损失函数调整,从而进一步提升检测的精准度。实验结果表明,所提出的算法的检测精度可以达到98.4%,模型大小缩减为2.797 MB,同时使每秒帧数(FPS)达到了169 f/s,成功地实现了对曲轴表面的缺陷检测。

A proposed improved method for the detection of crankshaft RB-YOLOv8 surface defects was presented,due to the complexity and difficulty posed by complex defect backgrounds as well as the slow rate at which small target flaws can be detected.To begin,the C2f module of the classic backbone network is supplanted by RepViT.The lightweight model is brought in to decrease the computational intricacy of the network and enhance detection velocity.Additionally,BiFPN bidirectional feature fusion module is advanced and a Small target detection layer is incorporated,thereby augmenting the detection capability of minor target flaws.Introducing the BiFormer attention mechanism to bolster the model′s robustness and enhance its detection accuracy in order to address complex defect backgrounds,MPDIoU loss function was then utilized for further enhancement.The experiment results demonstrate that the mAP value of the three defective samples is 98.4%.the number of model parameters is reduced to 2.797 MB,and the FPS is increased to 169 f/s,which can accurately detect the surface defects of the crankshaft,and has practical application value.