Research On Object Detection Method In Remote Sensing Images Based On Deep Semantic Feature

With the continuous development of remote sensing technology,the application of remote sensing images has become increasingly widespread.Efficiently extracting valuable information from massive remote sensing images is the key to the interpretation of these images.Remote sensing object detection aims to obtain the position and category of complex objects within remote sensing images,providing basic information for subsequent analysis and tracking of specific objects.In recent years,artificial intelligence algorithms have been constantly updated,and deep learning-based remote sensing image object detection methods have been widely applied.However,due to the complexity of the background in remote sensing images,the variability of object scale,and the limited utilization of prior information,deep convolutional neural networks are not sufficiently capable of extracting deep semantic features of objects within remote sensing images.Additionally,the issue of discriminative representation of objects severely affects the performance of remote sensing object detection.This thesis focuses on the research of deep semantic feature extraction for complex objects within remote sensing images,with the main work as follows:1.This thesis proposes a Transformer-based deep semantic feature extraction network,which exploits the correlation between feature map channels and spatial positions,as well as the positional information of feature units,to enhance the model’s representational ability in both channel and spatial dimensions.This method has been proven to improve the representational capability of convolutional neural networks in natural image object detection tasks and demonstrates superior performance in extracting features of objects within remote sensing images.2.This thesis presents a multi-scale feature learning network based on feature alignment and anti-aliasing effect,which addresses the information loss problem in the current feature pyramid.By preserving channel information in the spatial dimension through pixel shuffling,the method strengthens the spatial consistency of features at different resolutions through template matching and offset learning.A semantic encoder is then employed to unify the feature space.This approach has been validated for its effectiveness in multi-scale object detection on remote sensing image datasets.3.This thesis proposes a remote sensing image object detection algorithm based on correlation learning,which utilizes the top-level features of the feature pyramid to guide the learning of other level features,thereby enhancing the deep semantic information of feature maps.Transformers are introduced in the head network of the detector to learn the correlation information and positional information between candidate objects,thus strengthening the detector’s recognition ability of object categories.This thesis designs various solutions for the problems of object feature extraction and utilization of prior information in the field of remote sensing image object detection,which have high theoretical significance and practical value for the development of remote sensing image object detection.