Recent years have witnessed rapid development and remarkable achievements on deep learning object detection in remote sensing (RS) images. The growing improvement of the accuracy is inseparable from the increasingly complex deep convolutional neural network and the huge amount of sample data. However, the under-fitting neural network will damage the detection performance facing the difficulty of sample acquisition. Thus, it evolves into few-shot object detection (FSOD). In this article, we first briefly introduce the object detection task and its algorithms, to better understand the basic detection frameworks followed by FSOD. Then, FSOD design methods in RS images for three important aspects, such as sample, model, and learning strategy, are respectively discussed. In addition, some valuable research results of FSOD in computer vision field are also included. We advocate a wide research technique route, and some advice about feature enhancement and multi-modal fusion, semantics extraction and cross-domain mapping, fine-tune and meta-learning strategies, and so on, are provided. Based on our stated research route, a novel few-shot detector that focuses on contextual information is proposed. At the end of the paper, we summarize accuracy performance on experimental datasets to illustrate the achievements and shortcomings of the stated algorithms, and highlight the future opportunities and challenges of FSOD in RS image interpretation, in the hope of providing insights into future research.

Accurate prow detection (i.e., ship heading prediction) is important in many applications that rely on optical remote sensing imagery, such as track forecasting and maritime navigation. In recent years, many advanced methods based on deep convolution neural networks (DCNNs) have succeeded in detecting multidirectional ships. However, these methods are not effective at determining the prow orientation, primarily due to three limitations: weak adaptability to geometric transformations of ship targets, confusing the semantic information between the prow and other parts of ships, and the boundary discontinuity problem. To address these problems, we propose an omnidirectional prow detection network (OPD-Net) based on feature enhancement and an improved regression model. OPD-Net consists of a feature refinement network (FRN), a prow attention network (PAN), and a complex plane coordinates regression model (CPCRM). First, the FRN balances the low-level location information and high-level semantic information from multiscale feature maps and then fits various geometric transformations regarding ship targets through deformable blocks. Next, the PAN, which is based on supervised learning, is used to enhance the ship prow feature as well as suppress background noise, which improves the accuracy of ship heading predictions. Finally, the CPCRM is designed to effectively solve the boundary discontinuity problem and correctly achieve prow detection in arbitrary orientations. Experiments on optical remote sensing image data sets demonstrate the robustness and superiority of our method for prow detection. Moreover, our approach is also competitive when used only for ship detection.

Few-shot object detection (FSOD) in remote sensing images, which aims to detect never-seen objects with few training samples, has attracted wide attention. Some recent works leverage meta-learning to tackle this challenging task and achieve promising performance. However, information attenuation during feature extraction and simple prototype representation hamper further improvement in detecting novel classes. In this article, we propose a novel meta-learning-based FSOD approach named DPL-Net. Specially, within the meta-learning-based framework, performing role-specific feature extraction in query and support branches, DPL-Net adopts a Fine-grained Information Fusion (FIF) module to capture scale-aware information within query regions of interest (RoIs) and a Multi-frequency Information Enhancement (MIE) module to retain the spectral information of support samples, respectively. Moreover, considering the variability of remote sensing objects, a Discriminative Prototype Learning (DPL) strategy is developed to rectify the ambiguous distribution of support samples for more representative class-aware prototypes. Experiments on two benchmark datasets (NWPU VHR-10 and DIOR) demonstrate that our method effectively improves the performance of meta-learning in detecting remote sensing images with limited training data.

High-resolution optical remote sensing data can be utilized to investigate the human behavior and the activities of artificial targets, for example ship detection on the sea. Recently, the deep convolutional neural network (DCNN) in the field of deep learning is widely used in image processing, especially in target detection tasks. Therefore, a complete processing system called the broad area target search (BATS) is proposed based on DCNN in this paper, which contains data import, processing and storage steps. In this system, aiming at the problem of onshore false alarms, a method named as Mask-Faster R-CNN is proposed to differentiate the target and non-target areas by introducing a semantic segmentation sub network into the Faster R-CNN. In addition, we propose a DCNN framework named as Saliency-Faster R-CNN to deal with the problem of multi-scale ships detection, which solves the problem of missing detection caused by the inconsistency between large-scale targets and training samples. Based on these DCNN-based methods, the BATS system is tested to verify that our system can integrate different ship detection methods to effectively solve the problems that existed in the ship detection task. Furthermore, our system provides an interface for users, as a data-driven learning, to optimize the DCNN-based methods.

Ship detection is of considerable significance in both military and civilian application domains. Deep Convolutional Neural Network (DCNN) with region proposal mechanism, e.g., Faster R-CNN, performs outstandingly in ship detection with high-resolution images. However, the accuracy limitation is induced by the region proposal restricted by the training set for multi-scale target detection. Therefore, the method of multi-scale ship object detection is proposed based on saliency estimation in our work. Saliency Estimation Algorithms (SEAs) are often used to extract saliency features in images. In ship detection of high-resolution remote sensing images, these algorithms can extract information such as the scale and position of the targets, and then help the DCNN-based ship detection method to obtain better performance. To verify the effectiveness of the saliency estimation algorithm in multi-scale ship detection of high-resolution remote sensing images, and analyze the advantages of different SEAs. This paper introduces 13 classic saliency estimation algorithms and 2 DCNN-based ones to compare them by using evaluation indicators. At last, in order to demonstrate the performance of different SEAs, the extracted saliency feature maps are used to assist the DCNN-based target detection under multi-scale ships condition. In general, this framework can improve the detection accuracy of large-scale ships under scenarios of training only with small scale ships or without enough datasets.