基于注意力机制和U-net网络的漆面图像分割方法

Paint Image Segmentation Method Based on Serial Dual Attention Mechanism and U-net Network

自动除漆小车的视觉导航过程中,对漆面的分割精度和实时性有着较高的要求。为此,提出一种改进的U-net模型,用于视觉导航中的漆面分割任务。首先,搜集、制作了漆面数据集,并通过数据增强进行了数据集扩充。其次,通过引入Efficientnet-B0编码器和Focal Loss、特征融合环节嵌入串行双注意力机制,对U-net模型做了三处改进。最后,利用改进的模型对漆面进行分割,并与其他模型进行对比实验。实验结果表明,论文改进后的U-net模型,mPA比起U-net和ResU-net分别提升了2.63%和2.19%,达到了97.8%,PA比起U-net和ResU-net分别提升了2.88%和6.12%,达到了90.79%;在分割时间方面,比起DeepLabV3和DeepLabV3+分别提升了91.94%和90.17%,达到了0.1039s。论文提出改进的U-net模型,很好地兼顾了分割精度和实时性的要求。

In the process of visual navigation of automatic paint removing vehicle,the precision and real-time of paint surface segmentation are required.Therefore,this paper proposes an improved U-net model for paint segmentation in visual navigation.Firstly,the paint surface data set is collected and made,and the data set is expanded through data enhancement.Secondly,the U-net model is improved in three aspects by introducing efficientnet-B0 encoder,Focal Loss and feature fusion to embed serial double attention mechanism.Finally,the improved model is used to segment the paint surface,and the experiments are compared with other models.The experimental results show that the mPA of the improved U-net model is improved by 2.63% and 2.19% compared with U-net and ResU-net respectively,reaching 97.8%,and PA is improved by 2.88% and 6.12% compared with U-net and ResU-net respectively,reaching 90.79%.In terms of segmentation time,compared with DeepLabV3 and DeepLabV3+,the segmentation time is improved by 91.94% and 90.17% respectively,reaching 0.1039s.In this paper,an improved U-net model is proposed to meet the requirements of segmentation accuracy and real-time performance.