Remote Sensing Scene Graph and Knowledge Graph Matching with Parallel Walking Algorithm

In deep neural network model training and prediction, due to the limitation of GPU memory computing resources, massive image data must be cropped into limited-sized samples. Moreover, in order improve generalization ability model, samples need randomly distributed experimental area. Thus, background information is often incomplete or even missing. On this condition, a knowledge graph applied semantic segmentation remote sensing. However, although single sample contains only limited number geographic categories, combinations objects are diverse complex different Additionally, involved categories span classification system branches. Therefore, existing studies directly regard all as candidates for specific segmentation, which leads high computation cost low efficiency. To address above problems, parallel walking algorithm based on cross modality proposed scene graph—knowledge matching (PWGM). The uses map visual features space through anchors designs that takes account scenes. Based algorithm, we propose experiments demonstrate our improves overall accuracy by 3.7% compared with KGGAT (which using attention (GAT)), 5.1% GAT 13.3% U-Net. Our study not effectively recognition efficiency sensing objects, but also offers useful exploration development learning from data-driven data-knowledge dual drive.