基于改进DeepLabv3+网络的风机叶片分割算法研究

Research on fan blade segmentation algorithm based on improved DeepLabv3+network

为了提高风机叶片图像的分割质量,提出了一种改进DeepLabv3+网络的风机叶片分割算法。由于无人机采集风机叶片图像具有背景复杂和叶片占比差异较大的问题,提出的算法在DeepLabv3+网络的基础上改进了ASPP模块和Decoder模块。DASPP通过级联多个空洞卷积层,使用密集连接的方式将每个空洞卷积层的输出传递给后续的空洞卷积层,通过一系列的特征连接编码不同尺度的中间特征,获得了更大范围的感受野。在Decoder阶段添加多层特征融合,以恢复在降采样过程中丢失的细节信息和各级特征。通过对风机叶片数据集进行实验,MIoU值达到了0.9913,PA值达到了0.9968,实验表明该设计的算法对风机叶片的分割效果优于DeepLabv3+网络,具有更好的细节信息。

In order to improve the segmentation quality of fan blade image,this paper proposes a fan blade segmentation algorithm based on improved DeepLabv3+network.Due to the problems of background assistance and large difference in the proportion of blades collected by UAV,the algorithm proposed in this paper improves the ASPP module and decoder module based on the DeepLabv3+network.DSAPP concatenates multiple hole convolutions,and transfers the output of each hole convolution layer to the subsequent hole convolution layer by using dense connection.Through a series of feature connections,DSAPP encodes intermediate features of different scales,and obtains a larger range of receptive fields.In the decoder stage,multi-layer feature fusion is added to recover the detail information and all levels of features lost in the down sampling process.Through the experiment of fan blade data set,the MIoU value reaches 0.9913,PA value reaches 0.9968.The experimental results show that the segmentation effect of the algorithm designed in this paper is better than that of DeepLabv3+network,and has better detail information.