Synthetic-Aperture Radar image based positioning in GPS-denied environments using Deep Cosine Similarity Neural Networks

Abstract

Navigating unmanned aerial vehicles in precarious environments is of great importance. It is necessary to rely on alternative information processing techniques to attain spatial information that is required for navigation in such settings. This paper introduces a novel deep learning-based approach for navigating that exclusively relies on synthetic aperture radar (SAR) images. The proposed method utilizes deep neural networks (DNNs) for image matching, retrieval, and registration. To this end, we introduce Deep Cosine Similarity Neural Networks (DCSNNs) for mapping SAR images to a global descriptive feature vector. We also introduce a fine-tuning algorithm for DC-SNNs, and DCSNNs are used to generate a database of feature vectors for SAR images that span a geographic area of interest, which, in turn, are compared against a feature vector of an inquiry image. Images similar to the inquiry are retrieved from the database by using a scalable distance measure between the feature vector outputs of DCSNN. Methods for reranking the retrieved SAR images that are used to update position coordinates of an inquiry SAR image by estimating from the best retrieved SAR image are also introduced. Numerical experiments comparing with baselines on the Polarimetric SAR (PolSAR) images are presented.

Homography augumented momentum constrastive learning for SAR image retrieval

“…Contrastive learning can be also utilized for pretraining in a self-supervised manner in order to enhance the performance before applying it for downstream tasks [20]. The primitive attempt, called a memory bank scheme, stores the whole vectors obtained at the previous iterations and uses a subset of them at the current iteration [21,10]. Since the encoder is gradually updated via backpropagation and a stochastic gradient descent algorithm, the output vectors stored in the memory bank are on occasion incompatible with those produced at the current iteration, which, in turn, leads to slow learning.…”

“…Park et al [10] proposed deep cosine similarity neural network to generate a 𝑙2 normalized feature vector of a SAR image with the primary purpose of comparing SAR image pairs scalably. The proposed idea of normalizing the vector during training is also equipped in the training process of the encoders used in the present work so as to maintain its norm consistently.…”

“…For contrastive learning parameters, the queue size 𝐾 = 1024 and the momentum parameter 𝑚 = 0.999. To prevent feature vector d from having an excessively high 𝑙2 norm, prior to normalization we added the 𝑙2 regularization term derived in [10] into the loss function. Specifically, this term is simply denoted as (||d|| 2 − 1) 2 .…”

“…One of the applications is the SAR image retrieval task [7,8,9], which aims to retrieve images from a large database that are similar to a query image. This application is further utilized for complementing navigation systems when global positioning system (GPS) is not available [10]. For the SAR image retrieval tasks, it is necessary to extract a compressed feature vector from a SAR image while maintaining the semantic information.…”

“…Even though this enhances the performance of retrieving images, comparing image pairs is not scalable, thus it cannot be efficiently applied over the large-scale database. Thus, recent efforts usually focus on developing a CNN-based apporoach combining both global descriptor and local feature approaches; first retrieving images roughly using global descriptor and reranking by utilizing the local feature approach [7,10].…”