基于GCNet网络的玻璃瓶缺陷检测算法OA
Defect detection algorithm for glass bottles based on GCNet network
针对当前玻璃瓶缺陷检测深度学习模型计算量大、参数量多、难以部署的问题,探索一种高效的轻量化解决方案.针对此问题,结合YOLOv8的网络结构设计特征提取网络GCNet.首先,采用GhostConv代替标准卷积;其次,为了降低YOLOv8瓶颈层的参数量和计算量,设计瓶颈层卷积模块,并重新搭建瓶颈层;最后,结合C2f模块的结构设计构建新的CM模块.新的特征提取网络较原YOLOv8网络有着更低的参数量.在特征融合部分采用重新构建的重复加权双向特征融合金字塔结构,解决随着网络层数的加深导致特征信息丢失的问题.同时针对边界框的回归问题上,结合WIoU与Inner-ShapeIoU,提高模型的回归收敛速度.结果表明,相较于YOLOv8算法,由上述方法组成的YOLOv8-DB参数量降低了45.8%,计算量下降了11.9%,精度提升了0.4%.改进后的模型能够有效地降低占用的计算资源,更好地适用于特定工业检测环境.
Aiming at the problems of high computational cost,large parameter volume,and difficult deployment in current deep learning models for glass bottle defect detection,an efficient lightweight solution was explored.To solve this problem,the feature extraction network GCNet was designed by combining the network structure of YOLOv8.First,GhostConv was used to replace standard convolution.In order to reduce the number of parameters and calculation amount of YOLOv8 bottleneck layer,the convolution module of bottleneck layer was designed,and the bottleneck layer was rebuilt.A new CM module was built based on the structure design of C2f module.The new feature extraction network had a lower parameter count than the original YOLOv8 network.In the feature fusion part,a reconstructed double-weighted bidirectional feature fusion pyramid structure was used to solve the problem of feature information loss with the deepening of network layers.At the same time,for the boundary box regression problem,the combination of WIoU and Inner-ShapeIoU improved the regression convergence speed of the model.The results show that compared with the YOLOv8 algorithm,the YOLOv8-DB composed of the above,the number of parameters is reduced by 45.8%,the calculation amount is reduced by 11.9%,and the accuracy is increased by 0.4%.The improved model can effectively reduce the consumption of computing resources,and is better suitable for specific industrial detection environments.
傅莉;房志磊;席剑辉;任艳
沈阳航空航天大学 自动化学院,沈阳 110136沈阳航空航天大学 自动化学院,沈阳 110136沈阳航空航天大学 自动化学院,沈阳 110136沈阳航空航天大学 人工智能学院,沈阳 110136
信息技术与安全科学
深度学习缺陷检测多尺度卷积WIoU轻量化
deep learningdefect detectionmulti-scale convolutionWIoUlight weight
《沈阳航空航天大学学报》 2026 (1)
48-55,8
国家自然科学基金(项目编号:61602321).
评论