摘要
针对纺织品表面缺陷检测算法速度低、泛化性能差及鲁棒性弱等问题,课题组提出了一种基于改进Yolo v5的织物表面缺陷检测方法。在Yolo v5骨干网络的基础上增加了卷积注意力模块,增强目标检测网络对特征图中重要信息提取并弱化无关特征;针对网络特征融合阶段特征尺度不一致造成的冲突问题,引入自适应空间特征融合的方法;在训练过程中使用迁移学习加快训练速度。实验结果表明:与未改进的Yolo v5算法相比,所提出的检测框架能够有效提高网络精度达98.8%,检测速度达83帧/s。该检测方法能满足实际工业要求。
Afabric defect detection method based on improved Yolo v5 was proposed in response to the problems of low speed,poor generalization performance and weak robustness of the textile surface defect algorithm.The convolutional attention module was added to the Yolo v5 backbone network to enhance the target detection network for extracting important information in the feature maps and weakening irrelevant features.For the conflict problem caused by inconsistent feature scales in the feature fusion stage of the network,an adaptive spatial feature fusion method was introduced.Transfer learning was used to accelerate the training speed during the training process.Experimental results show that the proposed detection framework can effectively improve the network accuracy up to 98.8%and detection speed up to 83 FPS compared with the unimproved Yolo v5 algorithm.The proposed detection method can meet the practical industrial requirements.
作者
王恩芝
张团善
刘亚
WANG Enzhi;ZHANG Tuanshan;LIU Ya(School of Mechanical and Electrical Engineering,Xi'an Polytechnice University Xi'an 710613,China)
出处
《轻工机械》
CAS
2022年第2期54-60,共7页
Light Industry Machinery
关键词
缺陷检测
深度学习
目标识别
卷积注意力机制
自适应空间特征融合
defect detection
deep learning
object recognition
convolutional block attention
adaptive spatial feature fusion
