基于改进SegFormer网络的路面裂缝分割算法

An improved pavement crack segmentation algorithmbased on SegFormer network

针对现有网络执行路面裂缝分割任务时,特征利用不充分、高层语义信息提取不足的问题,提出了一种改进SegFormer网络的路面裂缝分割算法.首先,为充分利用提取到的特征,摒弃了原网络以多层感知机(Multilayer Perceptron, MLP)作为解码器的方案,重新设计了融合不同尺度特征的解码器,并在特征融合时引入注意力模块为其提供信息融合指导,加强了高低层特征间的联系;其次,为弥补高层语义信息不足,设计了结合部分卷积(Partial Convolution, PConv)的空间高效卷积池化模块(Space Efficient Convolutional Pooling Module, SECPM),提升了模型对不同尺度裂缝的分割性能;最后,针对路面裂缝不受位置、形状等方面限制的特点,提出了一种新的数据增强方法,提高了模型的泛化性能.在公开数据集Crack500进行实验,相较于原网络,改进模型的F1和mIoU分别提升了1.03%、1.32%,本文提出的方法能更好地适应路面裂缝分割任务.

Aiming at the problems of inadequate feature utilization and insufficient extraction of high-level semantic information in pavement crack segmentation by existing networks,an improved pavement crack segmentation algorithm based on SegFormer network was proposed.Firstly,in order to make full use of the extracted features,the original scheme of using Multilayer Perceptron(MLP)as decoder was abandoned,and the decoder fused different scale features was redesigned.The attention module was introduced to provide information fusion guidance during feature fusion,and the relationship between high and low features was strengthened.Secondly,in order to make up for the lack of high-level semantic information,a Space Efficient Convolutional Pooling Module(SECPM)combined with Partial Convolution(PConv)was designed.The segmentation performance of the model for cracks of different scales is improved.Finally,a new data enhancement method was proposed to improve the generalization performance of the pavement crack,which was not limited by location,shape,etc.Experiments were carried out on Crack500,and compared with the original network,F1 and mIoU of the improved model improved by 1.03%and 1.32%,respectively.The method proposed in this paper can better adapt to the task of pavement crack segmentation.