基于改进YOLO模型的工业铝片缺陷检测  

Industrial Aluminum Sheet Defect Detection Based on Improved YOLO Model

作　　者：徐红牛 余华云[1] XU Hongniu;YU Huayun(School of Computer Science,Yangtze University,Jingzhou 434023,China)

机构地区：[1]长江大学计算机科学学院,荆州434023

出　　处：《组合机床与自动化加工技术》2023年第9期106-111,共6页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(61440023);中国高校产学研创新基金-新一代信息技术创新项目(2020ITA03012)。

摘　　要：针对目前铝片表面缺陷的目标检测存在很多问题,包括现场大规模算法和计算设备的不适用性,以及检测速度和精度之间的平衡等,提出了一种基于注意力机制的新颖轻量级检测方法。在YOLOv4框架的基础上提出GBANet主干网络,其基于一个新的卷积Ghost模块构建并将改进的注意力模块嵌入在堆叠的Ghost块中。对颈部网络进行了特征融合的重新设计和轻量化,增加感受野,通过SPPF-PANet模块简化网络并通过改进anchor box和损失函数等措施增强模型对缺陷对象精确性。实验表明,所提方法较原YOLOv4提高1.06%的mAP,检测速度达到了36.6 fps,模型体积减少了82.72%,并能有效识别铝型材表面不同种类的缺陷。所提方法能够满足铝型材工厂生产现场缺陷检测要求。Many problems exist for the current target detection of aluminum sheet surface defects,including the unsuitability of large-scale algorithms and computing devices in the field,and the balance between detection speed and accuracy.In this paper,a novel lightweight detection method based on attention mechanism is proposed,focusing on industrial applications of aluminum sheet defect detection.The GBANet backbone network is proposed based on the YOLOv4 framework,which is constructed based on a new convolutional Ghost module and embeds an improved attention module in the stacked Ghost blocks.The neck network is redesigned and lightened by feature fusion,the perceptual field is increased,the network is simplified by the SPPF-PANet module and the accuracy of the model for defective objects is enhanced by measures such as improved anchor box and loss function.Experiments show that the proposed method improves the mAP by 1.06%compared with the original YOLOv4,achieves a detection speed of 36.6 fps,reduces the model volume by 82.72%,and can effectively identify different kinds of defects on the surface of aluminum profiles.The proposed method can meet the requirements of defect detection in the production site of aluminum profile factories.

关 键 词：目标检测 铝材表面缺陷 YOLOv4 注意力机制 卷积神经网络 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]