Retrieving High Resolution Remote Sensing Images With Learned Deep Features

Searching for the required high-resolution remote sensing image efficiently from a massive image dataset is the basis for follow-up interpretation and application.It is a practical and challenging problem in the remote sensing "big data era".Content-based remote sensing image retrieval is a method based on the visual content of the remote sensing images,which means,users only need to input a query remote sensing image to search for the required images.The system compares similarity between query images features and image features of dataset,and returns the most similar remote sensing images as results.Retrieving remote sensing images efficiently is an important foundation for the subsequent application of remote sensing technology.However,there are still some problems in the existing solutions for high-resolution remote sensing image retrieval tasks.1)Traditional hand-crafted features can describe features such as color,texture,and structure through elaborate design,but they have very limited description capabilities for the high-resolution remote sensing images containing complex spectral information and spatial geometric relations,especially for the high-level semantic information,which greatly constrains the final retrieval accuracy.2)Existing remote sensing image content retrieval systems require the user to input an existing query image.In the absence of query images,existing search methods are losing efficiency.To solve the above problems,this paper propose a model that can extract high-level semantic information from high-resolution remote sensing images.At the same time,it introduces the sketch retrieval method to solve the problem of content-based image retrieval when there is no query image.In this regard,this paper mainly focuses on deep feature and deep cross-domain feature extraction algorithms for remote sensing image retrieval problems.Among them,the main research content and contributions are as follows:Firstly,we propose the deep feature extraction algorithm for the content-based remote sensing image retrieval.The performance of different network structures and different layer deep features on remote sensing image retrieval tasks are compared.Meanwhile,the deep features are transferred to the domain of remote sensing images through fine-tuning networks,which makes deep network extraction more suitable features for the visual characteristics of remote sensing images.To solve the multi-scale problems of remote sensing images,multi-spatial-scale deep features are proposed for remote sensing image retrieval to improve the accuracy of remote sensing image retrieval tasks.Secondly,for the problem of retrieving remote sensing image when there is no query image,we propose to use free-hand sketches to retrieve remote sensing images for the first time.The existing problems of content-based remote sensing image retrieval methods in this cross-domain problem are analyzed,and a deep cross-domain model is proposed to solve this problern.To implement this method,a remote sensing sketch-image corresponding dataset is collected.The dataset is used to train deep cross-domain models,which provided a basis for the training of deep cross-domain models.Finally,considering the different visual contents caused by the different levels of sketches,a multi-scaled cross-domain model is proposed.This model is trained by a mixing dataset of sketches,remote sensing images,and edge maps of remote sensing images,and each type of images in the dataset contains different levels of details.The multi-scale depth model can extract free-hand sketches and image features with different detail scales.It can overcome the cross-domain difficulties,and better describe the visual characteristics of common features between sketches and remote sensing images.The use of free-hand sketches can efficiently and accurately search for the required remote sensing images from massive images and solve the existing problems.To verify the performance of the proposed algorithm,we conducted experiments on two high-resolution remote sensing image dataset and compared the proposed method with existing methods.The results verify the effectiveness and superiority of the proposed deep features and deep cross-domain features.