基于改进YOLOv7的溴铅铯薄膜表面缺陷检测算法

CsPbBr3 thin film surface defect detection algorithm based on improved YOLOv7

为了实现溴铅铯薄膜表面缺陷的自动检测和提高缺陷检测的精度,提出一种改进的目标检测算法YOLOv7DBS。首先,在主干网络中设计了DCN-ELAN模块,引入了可变形卷积DCNv3,增强主干网络对复杂形状缺陷特征的提取能力;其次,在颈部网络中引入BiFormer注意力机制,降低背景对于缺陷检测的干扰,提升网络的检测精度;最后,引入了一种具有角度损失的新型损失函数SIoU替代原始损失函数CIoU,以增强预测框与真实框的匹配程度,从而提高缺陷检测的准确性。实验结果表明,提出的YOLOv7-DBS改进算法相较于YOLOv7基准模型具有更低的参数量和计算量,分别降低了0.17×10^(6)和3×10^(9),对于杂质和不均匀缺陷的检测准确率分别提高了2%和5%,同时模型的mAP@0.5提高了1.68%。这表明提出的改进模型更适合用于溴铅铯薄膜表面缺陷智能检测。

An improved object detection algorithm YOLOv7-DBS is proposed to realize the automatic detection of surface defects of CsPbBr3 thin films and improve the accuracy of defect detection.A DCN-ELAN module is designed in the backbone network.A deformable convolution DCNv3 is introduced to enhance the backbone network ability of feature extracting for complex-shaped defects.A BiFormer attention mechanism is introduced in the neck network to reduce background interference in defect detection and improve the detection accuracy of the network.A novel loss function SIoU with angular loss is introduced to replace the original CIoU loss function in order to enhance the matching degree between predicted boxes and truth boxes,which thereby improves the accuracy of defect detection.The experimental results show that the improved algorithm YOLOv7-DBS has a small number of parameters and lower computational complexity in comparison with the baseline model YOLOv7.Its quantity of the parameters is reduced by 0.17×106 and its computational complexity is reduced by 3×109,respectively.Its detection accuracy for impurities and uneven defects is increased by 2%and 5%,respectively,and its mAP@0.5 is increased by 1.68%.The results indicate that the proposed model is more suitable for the intelligent detection of surface defects of CsPbBr3 thin films.