A Novel Transformer-Based Adaptive Object Detection Method

• DOI: 10.3390/electronics12030478
• Publication Date: 2023-01-17T09:28:47Z
• Authors: Shuzhi Su; Runbin Chen; Xianjin Fang; Tian Zhang
• Keywords: Feature (linguistics); Backbone network

To accurately detect multi-scale remote sensing objects in complex backgrounds, we propose a novel transformer-based adaptive object detection method. The backbone network of the method is dual attention vision transformer that utilizes spatial window and channel group to capture feature interactions between different scenes. We further design an path aggregation network. In designed network, CBAM (Convolutional Block Attention Module) utilized suppress background information fusion paths different-level maps, new are introduced fuse same-scale maps increase maps. can provide more effective improve representation capability. Experiments conducted on three datasets RSOD, NWPU VHR-10, DIOR show mAP our 96.9%, 96.6%, 81.7%, respectively, which outperforms compared methods. experimental results remote-sensing better.