V型Transformer的遥感影像障碍物提取方法

The method of obstacle extraction about remote sensing image based on V⁃Transformer

遥感影像中的障碍物是地震采集观测系统变观的重要依据之一。传统的人工提取障碍物方法效率低,且易受人为因素影响,难以保证结果的一致性,不适用于复杂地表环境及数量庞大的障碍物。当前通用的卷积神经网络自动提取障碍物方法,由于卷积核的尺寸受限,无法直接进行远距离的语义交互,也不能准确提取具有较大跨度且部分被遮蔽的障碍物(乡间道路、河流等)。为此,提出了基于V型全自注意力网络(MTNet)提取遥感影像障碍物的方法。首先,MTNet采用端到端的V型编码器—解码器结构,通过跳跃连接实现信息交互;其次,用具有远距离建模能力的Mix-Transformer模块取代传统卷积层,提取和重建更准确的障碍物多尺度特征;最后,用轻量的块扩展层取代转置卷积,实现上采样和图像分割,重建障碍物信息。实验结果表明,该网络分割障碍物的精度和速度显著优于现有方法,尤其在道路识别方面,优势更明显。

Obstacles in remote sensing images are the most important bases for the variable geometry of observa⁃tion systems in seismic exploration.The traditional manual obstacle extraction methods are inefficient and sus⁃ceptible to human factors,and difficult to ensure result consistency,making them unsuitable for complex sur⁃face environments and large numbers of obstacles.Current generalized methods for automatic obstacle extrac⁃tion with convolutional neural networks are limited by the size of convolution kernels,unable to directly per⁃form semantic interactions over long distances,and fail to accurately extract obstacles with large spans that are partially occluded(country roads,rivers,etc.).Therefore,this study proposes a V⁃shaped fully self⁃attention network(MTNet)to extract obstacles from remote sensing images.Firstly,MTNet adopts an end⁃to⁃end V⁃shaped encoder⁃decoder structure to realize information interaction through skip connections;Secondly,the tra⁃ditional convolutional layer is replaced by the Mix⁃Transformer block with long⁃range modeling capability to ex⁃tract and reconstruct more accurate multi⁃scale features of the obstacle;Finally,the transposed convolution is replaced by the light⁃weight block extending layer for upsampling and image segmentation to reconstruct the ob⁃stacle information.Experimental results show that the network significantly outperforms existing methods in terms of accuracy and speed in segmenting obstacles,especially in road recognition.