Multi-scale Features And Cross-sample Perception Based Remote Sensing Scene Graph Aggregation Mechanism

With the rapid development of deep learning,remote sensing object recognition based on deep learning has gradually been applied to tasks such as resource and environmental surveys,disaster emergency rescue,and land use planning.Among them,the recognition of complex geographic scenes with training samples as the basic unit has become hotspots in related fields.The emergence of graph neural network and its application in the field of automatic recognition of remote sensing objects has realized the transformation from traditional pixel-wise semantic segmentation to instance object segmentation,and greatly improved the efficiency of geographic object recognition.However,limited by the memory and computing power of GPU,the receptive field of the convolutional layer in the model and the samples used for model training can only be set to a limited size,which leads to the following problems in the application of graph neural network to complex geographic object recognition:(1)The fixed convolutional receptive field is difficult to extract the accurate features of multi-scale geographic objects;(2)The semantics of the relationship between objects is not clear;(3)The sample edge objects lack complete self-spectrum and contextual information.Aiming at the above problems,this paper proposes a scene graph aggregation model based on gated multi-scale feature fusion and cross-sample perception.By designing a multi-scale convolutional neural network and a scale-gated feature mechanism,the model extracts multi-scale features for geographic objects of different scales and performs adaptive feature scale fusion to obtain scale-adapted object features.The integrated GAT graph aggregation process provides clearer semantic guidance for the information aggregation of geographic objects;and through the cross-sample perception mechanism,the edge object information is completed to provide a complete context for the object,so as to achieve more efficient and accurate remote sensing semantic segmentation.The main research contents of this paper are as follows:(1)A multi-scale feature extraction convolutional neural model based on the gating mechanism is proposed,which uses the object scale to train the gating weights,and weights and fuses the features of the objects under different convolution depths to obtain features that adapt to the object scale.(2)Perform weighted correction to the node graph aggregation process based on node similarity,which transmits and updates node selectively according to feature similarity,and correct the offset feature vector by local average pooling.(3)A cross-sample object perception and scene graph aggregation mode is designed.By constructing a sequence of adjacent samples,the receptive field of a single sample is expanded,so as to complement the missing feature information of edge objects in a sample to a certain extent.(4)Created a remote sensing object recognition dataset in complex geographic scenes,and designed related experiments to compare and analyze the geographic object recognition accuracy and effectiveness of the network model proposed in this paper.The research results show that,compared with traditional convolutional neural networks and graph neural networks,the scene graph-graph aggregation model based on gated multi-scale feature fusion and cross-sample perception has great advantages in complex geographic scenes,which can effectively extract the object features that adapt to the object scale and retain the clustering of the objects more completely,at the same time,it can effectively prevent the distorting effect of the object scale on the object features,and can effectively control the aggregation of nodes in the process of information aggregation to be conducive to node classification.Environmental features,compared to U-Net and basic GAT network classification accuracy can be improved by5-10 percentage points.In the future research,we will conduct in-depth research on the model in this paper from the perspectives of algorithm,data,and quantification,and expect it to be applied in the fields of resource and environment detection,social emergency rescue and so on.