基于DeepLabv3+改进在图像分割中的应用

Application of DeepLabv3+ Based Improvement in Image Segmentation

在图像语义分割中,针对物体边界分割不完整、模型参数量较多等问题,为满足遥感图像语义分割任务对精度和部署便捷的要求,对DeepLabv3+网络进行了改进。将MobileNetV3替换成Xception作为特征提取主干网络,使模型轻量化,再将空洞空间卷积池化金字塔(Atrous Spatial Pyramid Pooling,ASPP)中第一个分支中1×1卷积换成ACmix,以获取更多的特征。实验表明,与典型的语义分割模型相比,改进的模型不仅有较少的模型参数,而且提高了分割精度和训练效率。

In image semantic segmentation,the object boundary segmentation is incomplete and the model parameters are large.In order to meet the requirements of remote sensing image semantic segmentation task for accuracy and convenient deployment,DeepLabv3+network is improved.MobileNetV3 replaces Xception as the feature extraction backbone network to make the model lightweight,and then converts the 1’1 convolution in the first branch of Atrous Spatial Pyramid Pooling(ASPP) into ACmix to obtain more features.Experiments show that compared with the typical semantic segmentation model,the enhanced model not only has fewer model parameters,but also improves the segmentation accuracy and training efficiency.