Dual Network Structure With Interweaved Global-Local Feature Hierarchy for Transformer-Based Object Detection in Remote Sensing Image

Frequent and accurate object detection based on remote sensing images is an encouraging approach for monitoring dynamic of the interest earth surface. Transformer-based was recently developed to cope with trade-off dilemma between large computation load accuracy sacrifice confronted by region-proposal-based regression-based detection, its self-attention mechanism can provide a global understanding that has potential ability reasoning location relationship within sparsely heterogeneously distributed geospatial objects. However, essentially weak at modeling local feature hierarchy compensate scale variation object, it extremely difficult train due lack inductive bias, resulting in slow convergence. To overcome problem, this study proposed Dual network structure InterweAved Global-local TRansformer architecture (DIAG-TR), alleviate incompatibility form, hierarchically embed features into representations. Besides, learnable anchor box incorporated positional query decoder part spatial prior, which accelerate The DIAG-TR validated widely used optical image DIOR dataset, results demonstrate global-local contributes 3.4% mean average precision compared original method, convergence time shortened 2.5-fold. State-of-the-art methods are also participated as benchmark comparison, outperforms Faster-RCNN-FPN 8.9%, proves great observation community.