A Study On Semantic-enhanced Remote Semsing Image Object Detection Method

With the rapid development of remote sensing technology,it has become easier to obtain a large amount of remote sensing data.Remote sensing images have become an important source for the development of various fields,including military reconnaissance,security monitoring,and traffic control.Object detection plays a critical role in remote sensing segmentation and has important research value in practical applications.This paper focuses on optical remote sensing images and proposes two object detection models that enhance the semantic understanding of feature maps and decompose and fuse features to enhance their semantics.These models are applied to solve practical problems in optical remote sensing images.The specific work is as follows:(1)A context-aware semantic-enhanced object detection method for remote sensing is proposed in this paper.In this method,feature pyramid is employed to address the issue of objects with different scales in optical remote sensing images.The feature pyramid has a strong hierarchical structure,where the deep feature maps have a deeper understanding of semantics,while the shallow feature maps retain more detailed information.By guiding the attention of shallow features with deep features,the shallow features can maintain detailed features and deepen their understanding of semantics,which results in effective detection of small and closely arranged objects.(2)This paper proposes a feature relation fusion enhanced remote sensing object detection method.In this method,to address the issues of false detection and unbalanced objects caused by the similarity between targets in remote sensing images,we analyze the relationship between features on the feature map.However,the feature map is too large to build relationships due to limited memory.Therefore,we obtain a more compact feature relation tensor and three transformation matrices through Tucker decomposition.After the fusion of bilinear vector interactions,the transformed output vector is obtained.The output vector is then fed back into the feature map to obtain a new feature map.Finally,the generalized model of a global sampling self-attention enhancement is applied.(3)The two proposed methods are combined and analyzed how to make them work better in object detection methods.Finally,three public remote sensing datasets,DOTA,DIOR,and NWPU VHR-10,will be compared and evaluated,achieving m AP scores of 71.5%,73.3%,and 90.1%,respectively.Compared with other advanced object detection methods,this experimental method is able to achieve the best results.