Research And Implementation Of Object Detection Algorithm For Remote Sensing Images By Domain Classification

With the continuous development of space and computer technology,the application of satellite remote sensing technology has been gradually expanded from the earlier military field to the civilian field and has been widely used in the civilian field.Satellite remote sensing technology can break the geographical restrictions and provide abundant information resources in time.In these rich information resources,they are mainly presented in the form of image,I.E.satellite remote sensing image.There is a lot of useless information in the large amount of information,and with the further development of satellite remote sensing technology,the breadth and clarity of the image will be further improved,resulting in more complicated image information,so it seems impossible to continue to select useful information manually.In view of the above problems,based on the idea of remote sensing target detection,this thesis contains the following work:(1)A new satellite remote sensing image dataset,YNUFIST-RSOD(Yun Nan University Future Intelligent System Technology Lab-Remote Sensing Object Detection),which includes 100,000 images of aircraft and ships,collected from different airports,lakes and rivers,was constructed in this thesis.A total of 223 areas are collected to provide a base data for subsequent research work.(2)The YUNFIST-RSOD dataset has been evaluated from three aspects:the performance standard,the contribution of the dataset sample to the depth model and the research on the large-scale satellite remote sensing images.The evaluation results show that:1.The dataset can provide a new evaluation criterion for the related target detection algorithm.2.When the dataset is used for target detection depth model training,the luminance component of the dataset sample contributes a lot to the feature learning of the model,and the chrominance information component contributes less.3.The dataset can be used for the research of large-scale satellite remote sensing image target detection.(3)Due to the influence of satellite orbit fixed and weather interference,the number of satellite remote sensing images in different domain(Domain)is not balanced,and the trained depth model is poorly generalized.Knowledge can be transferred from the source domain to the target domain and the new depth model can be trained by migration learning.The experimental results show that the satellite remote sensing target detection algorithm combined with the migration learning method can improve the generalization performance of depth model.