基于改进Deeplabv3+的雾气图像分割

Fog image segmentation based on improved DeepLabv3 Plus

针对传统雾气图像分割算法耗时长、分割结果存在凹陷等问题,提出一种基于DeepLabv3+的雾气图像分割算法。改进算法将DeepLabv3+原结构编码器的Backbone替换为更轻量的Mobilenetv2网络;将解码器的特征融合结构进行重新设计,同时加入注意力通道模块、边缘细化模块,通过消融实验得到分割效果最佳的雾气图像分割网络。实验结果表明,改进算法对雾气图像的分割精度MIOU可达90.31%,优于以ResNet101为基础的DeepLabv3+算法,且分割速度提高了57.26%,模型容量减少了92.62%。

In view of the time-consuming and denting of the traditional fog image segmentation algorithm,a fog image segmentation algorithm based on DeepLabv3 plus is proposed.The improved algorithm replaces the Backbone of DeepLabv3 plus original structure encoder with a lighter Mobiletv2 network,redesigns the feature fusion structure of the decoder,introduces the attention channel module,edge refinement module,and obtains the fog image segmentation network with the best segmentation effect through ablation experiment.Experimental results show that the MIOU of the improved algorithm can be 90.31%,which is better than the DeepLabv3 plus algorithm based on ResNet101,and the segmentation speed is improved by 57.26 percent and the model capacity is reduced by 92.62 percent.