SAR Image Retrieval Based On Unsupervised Domain Adaptive And Fuzzy Rules

Synthetic Aperture Radar(SAR)is a remote sensing technology that can be observed at any time and under any weather conditions,it has extremely high economic value and military value.In the field of remote sensing imagery,it is an important and challenging task of how to obtain data that users are interested in by effectively retrieving SAR images.Convolutional neural networks(CNN)are widely used in remote sensing image retrieval tasks for their excellent performance in image classification tasks,but a large number of images with known-labels are needed to train or fine-tune CNN models.At the same time,most of the current remote sensing image retrieval methods search the whole image database,but only a few images on the retrieval dataset are relevant or similar to the query image,so it takes much time and is unnecessary.It is an urgent task to unsupervised learn retrieval features of SAR images by utilizing the existing database of remote sensing images with known-labels and improve retrieval speed.In this paper,a SAR image retrieval method based on unsupervised domain adaptation and fuzzy rules is proposed for SAR remote sensing image retrieval.Firstly,the method combines the idea of adversarial domain adaptation with fuzzy image clustering to learn the domain invariant features between SAR images and optical aerial images with the help of existing optical aerial image classification labeling information to extract the retrieval features applied to SAR images.Secondly,the method reduces the search space using a method to determine the search space based on fuzzy rules.The method uses an unsupervised adversarial domain adaptation model and improved fuzzy clustering to determine the classification confidence level of the SAR images;divides the images into three fuzzy classifications of "high confidence","medium confidence" and "low confidence" according to the classification confidence level;establishes corresponding fuzzy rules for determining the retrieval space for each fuzzy category;and determines the retrieval space for the images by fuzzy rules to improve the retrieval speed.Finally,the distance measure between the features is also improved by adding the classification confidence level of the images to the image similarity measurement process to further improve the retrieval accuracy.Therefore,the main contributions of this paper are summarized as follows:1.A model for unsupervised extraction of SAR image retrieval features is proposed that combines CNN domain adaptation and fuzzy clustering to provide a new feature extraction method for SAR image retrieval;2.An unsupervised method for determining search space based on fuzzy rules is proposed,which reduces the search space and improves the speed of retrieval.