Object Detection In Optical Remote Sensing Images Based On Deep Reinforcement Learning And Meta Learning

Optical remote sensing image is an important component of remote sensing imaging.It can be used to observe various objects and targets on the earth’s surface,such as airports,aircraft,buildings,etc.Therefore,it has a wide range of application value in the fields of national defense security,urban construction planning,disaster monitoring and so on.This undoubtedly brings higher requirements for automatic interpretation of optical remote sensing images.With the rapid development of optical remote sensing technology in China,the amount of data acquired by optical remote sensing image is more and more rich,the scale of optical remote sensing image is larger and larger,and the resolution of optical remote sensing image is higher and higher.In addition,although optical remote sensing images are abundant,the annotated images are still very scarce.Many researchers have proposed various solutions for object detection in optical remote sensing images,but the traditional methods still have many shortcomings in the face of challenges such as large scale,very high resolution and lack of annotated data.In the past ten years,deep learning methods,especially convolutional neural network(CNN),have shown great success in the field of computer vision such as object detection and object tracking.With the development of high performance computing technology,more and more scholars began to introduce deep learning method into optical remote sensing image interpretation,and achieved good results.In this paper,the characteristics of optical remote sensing images are thoroughly studied and utilized.Combined with the deep learning method,some research is done on object detection for optical remote sensing images.The specific work is as follows:1)In order to solve the problem casued by large scale optical remote sensing image,this paper puts forward a method based on wavelet transform combined with deep reinforcement learning for optical remote sensing image object detection.Firstly,the wavelet transform is used to transform the original RGB image into multiple wavelet subbands.Then,we use the deep reinforcement learning to select the limited number of subbands.Despite reducing the input data size and the number of subbands,the proposed object detection method is still able to achieve 2% performance improvement.In addition,in order to make the traditional object detection network concentrate on the scale features and direction features of the input wavelet subbands,the discrete wavelet multi-scale attention mechanism is introduced,which can significantly improve the ability of feature extraction and generalization of the detector,further improving the performance of the object detection network.2)To solve the problems of weak feature extraction ability and scarcity of annotated data in the existing background subtraction method,we designed a new background subtraction network using CGAN and DANN.Using the generated background and the current video frame,the method is trained on the natural video,and the background subtraction results of VHR optical remote sensing video are obtained by using the domain adaptation method.The proposed method uses the idea of deep network and GAN confrontation training to solve the problem that the traditional background subtraction method is not strong in feature extraction.The domain adaptation method is used to overcome the problem of lack of annotated data in current remote sensing datasets and the problem of similar but different feature distribution between training data and test data.In addition,we propose the parallel processing method of cutting and splicing the original video,which not only solves the difficulty that the traditional method cannot process the large-scale remote sensing video,but also greatly improves the computing speed and efficiency.3)In order to solve the problem of lack of annotated data in optical remote sensing image,we proposed a few shot object detection method of optical remote sensing image based on meta-learning,which used the feature similarity between the support images and the query images to carry out feature matching and similarity learning.On the basis of Faster R-CNN,the backbone is firstly used to extract the features of the input support images and query images,and then the feature matching module is used to match the support branch features and query branch features and get the heat map,which will be sent to RPN to generate regional proposals.Finally,the similarity measurement module is used to measure the global similarity measurement,local similarity measurement and pixel-level similarity measurement,and the similarity fusion is carried out to obtain the final similarity result and the bounding box regression offset.