基于注意力机制改进的DeepLabV3+遥感图像分割算法

Improved DeepLabV3+remote sensing image segmentation algorithm based on attention mechanisms

DeepLabV3+分割算法具有高效的编解码结构,常用在图像分割任务中。针对DeepLabV3+高分辨率遥感图像语义分割中存在的分割目标边缘不精确和孔洞缺陷问题,提出了一种基于注意力机制改进的DeepLabV3+遥感图像分割算法。构建ECBA(Efficient Convolutional Block Attention Module)注意力机制,将ECBA添加至DeepLabV3+主干网络Xception,增强其特征提取能力,得到注意力加权的高层特征。同时,将ECBA添加至编码器和解码器的连接支路,得到注意力加权后的低层特征。解码器将两种特征进行特征融合,以增强网络对不同分割目标的边缘以及同一目标内部的感知。实验结果表明,改进后的算法在ISPRS Potsdam数据集上的平均交并比(mean Intersection over Union,mIoU)和F1指数分别达到了79.80%和75.88%,比DeepLabV3+算法提高了11.06%和6.32%。

DeepLabV3+has an efficient coding and decoding structure and is commonly used in image segmentation tasks.Aiming at the problems of imprecise segmentation target edges and hole defects in the semantic segmentation of DeepLabV3+high-resolution remote sensing images,an improved DeepLabV3+remote sensing image segmentation algorithm based on the attention mechanism is proposed.The ECBA(Efficient Convolutional Block Attention Module)attention mechanism is constructed,and ECBA is added to the DeepLabV3+backbone network Xception to enhance its feature extraction capability and obtain the attention-weighted high-level features.At the same time,ECBA is added to the encoder and decoder connection branch to get the attention-weighted low-level features.The decoder performs feature fusion of the two features to enhance the network's perception of the edges of different segmented targets as well as the interior of the same target.The experimental results show that the improved algorithm achieves a mean Intersection over Union(mIoU)and F1 index of 79.80%and 75.88%on the ISPRS Potsdam dataset,which is 11.06%and 6.32%better than the DeepLabV3+algorithm respectively.