CubeSats provide a low-cost, convenient, and effective way of acquiring remote sensing data, and have great potential for remote sensing object detection. Although deep learning-based models have achieved excellent performance in object detection, they suffer from the problem of numerous parameters, making them difficult to deploy on CubeSats with limited memory and computational power. Existing approaches attempt to prune redundant parameters, but this inevitably causes a degradation in detection accuracy. In this paper, the novel Context-aware Dense Feature Distillation (CDFD) is proposed, guiding a small student network to integrate features extracted from multi-teacher networks to train a lightweight and superior detector for onboard remote sensing object detection. Specifically, a Contextual Feature Generation Module (CFGM) is designed to rebuild the non-local relationships between different pixels and transfer them from teacher to student, thus guiding students to extract rich contextual features to assist in remote sensing object detection. In addition, an Adaptive Dense Multi-teacher Distillation (ADMD) strategy is proposed, which performs adaptive weighted loss fusion of students with multiple well-trained teachers, guiding students to integrate the learning of helpful knowledge from multiple teachers. Extensive experiments were conducted on two large-scale remote sensing object detection datasets with various network structures; the results demonstrate that the trained lightweight network achieves auspicious performance. Our approach also shows good generality for existing state-of-the-art remote sensing object detectors. Furthermore, by experimenting on large general object datasets, we demonstrate that our approach is equally practical for general object detection distillation.

In the space station, visual gesture recognition is an important component for gesture based human-robot interaction to control the robot, which assists astronauts to complete simple, repetitive and cooperative work. However, public gesture datasets are captured in the daily environment and existing approaches trained on these datasets achieve imprecise performance for astronaut gesture recognition. In this paper, we introduce a new astronaut gesture dataset (DSSL-Astronaut gesture dataset) and a novel hierarchical attention single-shot detector network (HA-SSD) for astronaut gesture recognition. Specifically, this dataset consists of the real and augmented images. The real images captured in the simulated space station are closed to the real images in the space station environment. Meanwhile, we utilize the Mask-RCNN model to segment the foreground image including an astronaut from real data and capture the background images of the simulated space station at different views and illuminations. Then we combine them to synthesize the augmented images. A novel HA-SSD model consists of a lightweight backbone named MobileNet and a hierarchical attention mechanism. The MobileNet is used as the feature extractor, which contains several depth-wise separable convolutions to make a trade-off between latency and accuracy. Meanwhile, the hierarchical channel-wise attention module is adopted to exploit the fine semantic information to enrich the features for improving the performance of gesture recognition. We conduct experiments to demonstrate that our dataset is suitable for the space station and our approach is able to effectively localize and recognize the gestures with strong generality.

While general object detection with deep learning techniques has garnered immense progress, the performance of detecting small objects in remote sensing is far from satisfactory due to the lack of sufficient details. To address this problem, this paper designs a spectral network combining Fast Fourier Convolution (FFC) and detection network. By extending the receptive region of the network, features around the object are introduced as additional information to help detect small objects. Specifically, the FFC in the proposed spectral network extracts global features through its unique Fourier Unit: the spatial feature is first transformed into the frequency domain by the Fast Fourier transform (FFT), then a convolution block is performed to extract the frequency features, obviously this frequency-aware convolution block has a global receptive field in the spatial domain covering the whole image, and finally the features are recovered to the spatial domain using the inverse FFT. To demonstrate the effectiveness of our approach, we conduct experiments on the large remote sensing dataset DIOR, which shows that our approach has excellent performance compared to other detectors. It achieves an average accuracy (mAP) of 73.5% without any tricks.

Visual surface inspection is a challenging task owing to the highly diverse appearance of target surfaces and defective regions. Previous attempts heavily rely on vast quantities of training examples with manual annotation. However, in some practical cases, it is difficult to obtain a large number of samples for inspection. To combat it, we propose a hierarchical texture-perceiving generative adversarial network (HTPGAN) that is learned from the one-shot normal image in an unsupervised scheme. Specifically, the HTP-GAN contains a pyramid of convolutional GANs that can capture the global structure and fine-grained representation of an image simultaneously. This innovation helps distinguishing defective surface regions from normal ones. In addition, in the discriminator, a texture-perceiving module is devised to capture the spatially invariant representation of normal image via directional convolutions, making it more sensitive to defective areas. Experiments on a variety of datasets consistently demonstrate the effectiveness of our method.

Remote sensing object detection is a critical task in Earth observation. Despite the remarkable progress made in general object detection, existing detectors struggle with remote sensing scenarios due to the prevalence of numerous small objects with limited discriminative cues. Cutting-edge studies have shown that incorporating contextual information effectively enhances the detection performance for small objects. Meanwhile, recent research has revealed that convolution in the frequency domain is capable of capturing long-range spatial dependencies with high efficiency. Inspired by this, we propose a Frequency-aware Feature Pyramid Framework (FFPF) for remote sensing object detection, which consists of a novel Frequency-aware ResNet (F-ResNet) and a Bilateral Spectral-aware Feature Pyramid Network (BS-FPN). Specifically, the F-ResNet is proposed to extract the spectral context information by plugging the frequency domain convolution into each stage of the backbone, thereby enriching features of small objects. In addition, the BS-FPN employs a bilateral sampling strategy and skipping connection to model the association of object features at different scales, enabling the contextual information extracted by the F-ResNet to be fully leveraged. Extensive experiments are conducted for object detection in the public remote sensing image dataset and natural image dataset. The experimental results demonstrate the excellent performance of the FFPF, achieving 73.8% mAP on the DIOR dataset without using any additional training tricks.

Optical (OPT) and Synthetic Aperture Radar (SAR) cross-modal remote sensing image retrieval is a key technique for all-weather Earth observation. The enhancement of SAR image resolution has driven the development from conventional scene-level retrieval to object-level retrieval. Inter-class similarity, intra-class variability, and cross-modal discrepancies are general challenges in cross-modal image retrieval. Although existing methods have attempted to address them at the scene level, these challenges persist in object-level tasks due to inherent limitations of the methods, such as detail loss caused by global pooling and insufficient local discrimination. To address this gap, a novel FGHN is proposed to enhance the retrieval accuracy. First, a focus-aware residual network is designed to highlight discriminative local features, facilitating the extraction of semantic details. Second, an intra-class feature aggregation loss is introduced to improve intra-class compactness, reinforcing the consistency of fine-grained representations. Finally, a distance enhanced weighting strategy is incorporated into the hash learning loss to promote similarity among hash codes of the same class. To support fine-grained OPT-SAR retrieval, we construct a High-Resolution OPT-SAR dataset, termed HROS. It is worth noting that the SAR images in HROS were collected by our team. Extensive experiments are conducted to evaluate both FGHN and HROS. The experimental results demonstrate that HROS exhibits high resolution, strong robustness, and superior differentiation, establishing it as a promising benchmark for OPT-SAR retrieval. FGHN achieves state-of-the-art performance on HROS, obtaining 89.80% mAP and 93.95% P@10 for OPT image retrieval with SAR queries at 32 bits hash code. Its strong generalization capability is further validated on the public CBRSIR VS dataset. Ablation studies confirm the effectiveness of the three innovative modules. Moreover, FGHN proves potential for large-scale remote sensing image retrieval applications.