基于改进Deeplabv3+的遥感图像语义分割研究

Semantic Segmentation of Remote Sensing Image Based on Improved Deeplabv3+

遥感图像分割在城市规划、土地资源管理、交通规划等领域中发挥了重要作用,但是遥感图像中环境复杂、地物尺度差别较大,给遥感图像的分割带来一定难度。采用一个结合通道注意力的特征融合模块,替代Deeplabv3+网络中的解码器部分,以自适应地从高层特征引导的低层特征中选择有用的空间细节特征,并筛选相关干扰信息。其中通过通道注意力机制来得到加权后的高层特征,有利于提取全局上下文以及更有效的语义信息,并利用加权后的高层特征指导提取精细化的低层特征信息,以保留更多的图像边缘、纹理等信息。在INRIA Aerial Image高分辨率遥感图像数据集上进行训练和测试,并与相关模型进行对比,研究结果表明,改进后的Deeplabv3+网络在遥感图像分割中性能优异,改善了目标边缘以及小尺度目标物体的分割效果,具有一定的研究和应用价值。

Remote sensing image segmentation plays an important role in urban planning, land resource management, traffic planning, and other fields. But the complex environment and large differences in the scale of remote sensing images bring some difficulties to the segmentation of remote sensing images. A feature fusion module combined with channel attention is proposed to replace the decoder in the Deeplabv3+ network to adaptively select useful spatial detail features from low-level features guided by high-level features, and screen the relevant interference information. Among them, the weighted high-level features are obtained through the channel attention mechanism, which is conducive to the extraction of global contexts and more effective semantic information. To retain more image edge, texture, and other information the weighted high-level features are used to guide the extraction of refined low-level feature information. Training and testing are carried out on the INRIA Aerial Image high-resolution remote sensing image data set, and compared with related models. The experimental results show that the improved Deeplabv3+network has an excellent performance in remote sensing image segmentation, improving the segmentation effect of target edge and small-scale objects, which has certain research and application value.