基于Word Embedding的遥感影像检测分割

遥感影像检测分割技术通常需提取影像特征并通过深度学习算法挖掘影像的深层特征来实现.然而传统特征（如颜色特征、纹理特征、空间关系特征等）不能充分描述影像语义信息，而单一结构或串联算法无法充分挖掘影像的深层特征和上下文语义信息.针对上述问题，本文通过词嵌入将空间关系特征映射成实数密集向量，与颜色、纹理特征的结合.其次，本文构建基于注意力机制下图卷积网络和独立循环神经网络的遥感影像检测分割并联算法（Attention Graph Convolution Networks and Independently Recurrent Neural Network，ATGIR）.该算法首先通过注意力机制对结合后的特征进行概率权重分配；然后利用图卷积网络（GCNs）算法对高权重的特征进一步挖掘并生成方向标签，同时使用独立循环神经网络（IndRNN）算法挖掘影像特征中的上下文信息，最后用Sigmoid分类器完成影像检测分割任务.以胡杨林遥感影像检测分割任务为例，我们验证了提出的特征提取方法和ATGIR算法能有效提升胡杨林检测分割任务的性能.

Remote Sensing Image Detection and Segmentation Based on Word Embedding

Remote sensing image detection and segmentation technology usually needs to extract image features and mine the deep features of images through deep learning algorithm.However,traditional imaging features (e.g.,color,texture,spatial relationship) cannot fully reflect the semantic information of the images,while single/sequential algorithm cannot fully exploit the deep features and the contextual semantic information of the images.Aiming at the above challenges,in this paper,the spatial relation features are mapped into real dense vectors by word embedding,which are combined with color and texture features.Further,we propose a new parallel algorithm referred to as attention graph convolution networks and independently recurrent neural network (ATGIR) based on graph convolution network and independent recurrent neural network under attention mechanism for remote sensing image detection and segmentation.Our algorithm first assigns probabilistic weights to the combined features based on attention mechanism;then extracts deep features based on the features with high weights to generate labels with directions by using graph convolution network (GCNs) algorithms,extracts contextual semantic information of the images by using the independently recurrent neural network (IndRNN) algorithm;finally,our algorithm realizes image detection and segmentation by using Sigmoid.For remote sensing image detection and segmentation of populous euphratica forest as an instance,we prove that our feature extraction method and proposed ATGIR algorithm can effectively improve the detection and segmentation tasks.