Object detection in remote sensing images based on feature refinement

Directional target detection plays an important role in organized image target detection. We need to work out the problem of large scale and shapes in remote sensing images target detection, an independent feature thinning network is proposed. The whole network has two transformer branches, which handle classification and regression tasks respectively, and extract specific features. The object location and recognition of remote sensing image are realized. In order to extract the key information of Directional objects, we design a polarization attention module to extract context information at the feature level, especially the key information. In this paper, the rotation convolution is nested in the transformer to extract the rotation coding information in the regression task. Experimental results show that the accuracy of the proposed target detection model on UCAS-AOD and DOTA datasets reaches 97.75% and 76.01%, respectively, and the improved model is superior to most of the current target detection algorithms.