基于改进DeeplabV3+的遥感图像分割算法

Remote sensing image segmentation algorithm based on improved DeeplabV3+

针对高分辨率遥感图像语义分割存在地物边缘分割不连续、小目标分割精度不高的缺陷,本文提出一种基于改进DeeplabV3+的遥感图像分割算法,该算法首先使用分散注意力网络ResNeSt替换DeeplabV3+原始主干网络Xeception,以提取更丰富的深层语义信息,从而提高图像分割精度;其次引入坐标注意力机制(CA),有效获得更精确的分割目标位置信息,使得分割目标边缘更加连续;最后在解码层中采用级联特征融合方法(CFF)提高网络的语义信息表征能力。试验结果表明,该算法在中国南方某城市的高清遥感图像数据集分割任务上mIoU高达97.07%,相比原始DeepLabV3+模型提高了3.39%,能够更好地利用图像语义特征信息,为解译遥感图像语义信息提供一种新的思路。

Aiming at discontinuous object edge segmentation in high-resolution remote sensing image semantic segmentation and low accuracy of small object segmentation, this paper proposes a remote sensing image segmentation algorithm based on improved DeeplabV3+. The algorithm first adopts distraction network called ResNeSt instead of the DeeplabV3+ original backbone network Xeception to extract richer deep semantic information, thereby improving the accuracy of image segmentation;secondly, the Coordinate Attention(CA) mechanism is introduced to effectively obtain more accurate target location information of segmentation to make the segmentation target edge more continuous;finally, the cascade feature fusion method(CFF) is adopted in the decoding layer to improve the semantic information representation ability of the network. The experimental results show that the algorithm has a high mIoU of 97.07% on the high-definition remote sensing image dataset of a city in southern China, which is 3.39% higher than that of the original model and a reflection of better utilization of image semantic feature information. This provides a new way of thinking for remote sensing image semantic information.