Attention-based Remote Sensing Object Detection Methods

Remote sensing object detection is an important research content in the field of remote sensing imaging.Because of large changes in object scale in remote sensing images,many small objects,and complex scenes,it has brought great challenges to remote sensing object detection.This paper mainly studies the remote sensing object detection algorithm based on the attention mechanism,the main work includes:(1)A remote sensing object detection method based on multi-level attention mechanism was proposed.This method improved the common remote sensing object detection method based on deep convolution network by performing attention mechanism feature weighting operation on the features output by the receptive field module.Small remote sensing objects are susceptible to complex background interference,which affected the discriminative performance of object features,resulting in the phenomenon of missed detection of small remote sensing objects.In order to solve this problem,this paper weights the shallow features in the deep network by the attention mechanism feature,and directly predicts the position of the small object on the shallow feature layer,which can enhance the feature learning ability of the small object.Experimental results show that this method has better detection performance.(2)An anchorless frame remote sensing object detection method combining graph structure attention network and variable convolutional network is proposed.In the task of remote sensing object detection,the remote sensing object of interest has the characteristics of large change in appearance scale and complex background,which makes it difficult for the conventional detection algorithm based on fixed anchor frame to accurately characterize the object.To solve this problem,this paper proposes an object detection framework without anchor frame and combines graph structure attention network and variable convolutional network to improve the performance of object detection.Experimental results prove the effectiveness of this method.