边缘优化和注意力融合的遥感地物分割算法

Remote Sensing Ground Object Segmentation Algorithm Based on Edge Optimization and Attention Fusion

针对遥感地物图像种类众多且目标边缘较复杂的特点,以及现有分割网络中局部卷积的感受野有限,对图像上下文信息利用不足,导致分割目标边缘模糊以及分割精度低等问题,提出一种基于UNet3+网络的遥感地物分割算法。在解码过程中引入相似性感知点关联算子作为上采样方式,通过聚合特征金字塔中的多个建议,改善目标边界细节的分割能力;在编码过程中引入选择性内核模块,优化下采样方式,以实现神经元的自适应感受野大小,充分地获取目标特征的多尺度信息,精准捕捉有用的细节语义信息;在跳跃连接阶段添加双多尺度注意力模块,对不同尺度的特征进行加权融合,使模型更好地关注局部细节和全局上下文信息。在WHDLD、ISPRS Potsdam数据集上的实验表明,改进算法的平均交并比分别达到了64.4%、75.4%,较基线模型分别提升了约2.6个百分点、3.2个百分点,同时验证了改进算法在分割边缘模糊问题上的有效性。

Considering the characteristics of remote sensing land cover images with a wide variety of types and complex object edges as well as the limited receptive field of local convolutions in existing segmentation networks resulting in inadequate utilization of contextual information,leading to issues such as blurred object edges and low segmentation accuracy,this paper proposes a remote sensing land cover segmentation algorithm based on the UNet3+network architecture.Firstly,during the decoding process,a similarity-aware point affiliation operator is introduced as an upsampling method.This operator aggregates multiple proposals from the feature pyramid to enhance the segmentation capability for object boundary details.Secondly,in the encoding process,a selective kernel module is introduced to optimize the downsampling approach.This module enables neurons to achieve an adaptive receptive field size,facilitating the acquisition of multi-scale information from target features and precise capture of valuable detailed semantic information.Finally,in the skip-connection phase,a dual multi-scale attention module is added to perform weighted fusion of features from different scales,enabling the model to better focus on both local details and global contextual information.Experimental results on the WHDLD and ISPRS Potsdam datasets demonstrate that the proposed algorithm achieves mean intersection over union(MIoU)improvements of 64.4%and 75.4%respectively,compared to baseline models,the improvement is about 2.6 percentage points and 3.2 percentage points respectively.This also validates the effectiveness of the proposed algorithm in addressing the issue of blurry segmentation edges.