Bin Lin

Accurate flood mapping is essential for disaster response, yet optical–SAR fusion remains difficult when clouds, speckle noise, and terrain interference cause the two sensors to provide uneven or conflicting evidence. Multimodal models may suffer from “modality collapse”, while fixed scan patterns make water boundaries difficult to delineate. To address these issues, we propose CISRMamba, a multimodal state space model (SSM) based on a lightweight MobileMamba backbone. A Deformable Scan Router (DSR) adapts the scan trajectory to irregular flood contours while preserving SSM efficiency. Instead of forcing all regions into a rigid “align-then-fuse” process, the Feature Complementary Module (FCM) suppresses discrepant responses, supplies complementary information where one modality is incomplete, or enhances consistent evidence according to the local optical–SAR relationship. The Refined Feature Integration (RFI) module reduces distributional shift and modality collapse through a difference-guided residual, while the Spectral Refinement Block (SRB) recovers high-frequency spatial details weakened by downsampling. Experiments on CAU-Flood and Wuhan show that CISRMamba achieves mIoU scores of 91.53% and 58.67%, respectively, outperforming baselines based on CNNs, Transformers, and Mamba. These results indicate that CISRMamba combines context modeling with boundary-level refinement for efficient flood monitoring under heterogeneous weather and scene conditions.

Source extraction is a fundamental task in astronomical data analysis, yet existing methods are often specialized to specific instruments, spectral bands, or imaging conditions, which limits their generalizability. In this work, we propose Extract Anything for Astronomy (EAA), a unified and perception-driven framework that adapts the Segment Anything Model (SAM) into a general-purpose astronomical source-extraction system. By introducing domain-aware input reprojection and full fine-tuning of the feature encoder, we systematically bridge the substantial domain gap between natural images and astronomical data. EAA integrates two complementary extraction paradigms, i.e., interactive and automatic extraction, enabling flexible usage across a broad range of spectral regimes and representative imaging conditions. Extensive experiments on CANDELS (approaching 80% completeness for sources with H < 27 AB) demonstrate improved deblending capability, faint source detection, and extraction robustness. More importantly, we validate the generalizability of EAA across a wide range of spectral regimes, including ultraviolet (GALEX, >95% completeness for sources with near-UV < 21), ground-based visible-light observations (>90% completeness for sources with Gaia-G < 12), radio (FIRST, >95% completeness for sources with FPEAK > 5), and millimeter-wave (Atacama Large Millimeter/submillimeter Array) data. Compared with SExtractor, Morpheus, and SAM, EAA provides the more comprehensive functionality and stronger deblending and perceptual segmentation capabilities, while its main limitation is lower faint-source detection sensitivity than SExtractor. These results show that perception-driven segmentation based on foundation models can serve as a powerful and universal paradigm for astronomical source extraction, providing a scalable infrastructure for next-generation sky surveys, space domain awareness, and cross-band scientific analysis.

In charge-coupled device digital imaging, smears occur when light from a scene affects the detector during noninstantaneous charge transfer, creating bright lines across the image that hinder astronomical observations. Existing image processing methods struggle to remove these smears without influencing follow-up process. The lack of labeled data further complicates the development of deep learning-based approaches in this area. To address these challenges, we propose an imaging-model-based method to represent smears and guide supervised networks in learning smear suppression. Leveraging the imaging mechanism, the proposed network (IMGNet) identifies the charge transfer direction and smeared regions to produce clean images. Trained on simulated data, the method outperforms traditional, supervised, and unsupervised baselines. By explicitly representing smears, the network effectively mitigates the impact of domain gaps with authentic data. Experiments on downstream tasks, such as detection and registration, further demonstrate the superior performance of the proposed method.

星座识别用于在天文图像中识别出已知星座，是天文观测的重要基础。现有的星座识别算法聚焦于星历表的使用，缺乏仅利用图像信息的针对性研究。提出一种基于三角剖分以及相似性约束的星座识别方法。该方法使用Delaunay三角剖分构造三角形，以内角角度相似性进行匹配识别，并依据匹配星点得分情况实现初筛；引入确定点与被测点组合的方式，实现目标星座的选取。实验结果表明，本文方法在相对较短的识别时间内，平均识别准确率与匹配准确率分别为87.9%和76.1%，能有效提高目标星座的识别成功率。

UAVs of various shapes and sizes have rapidly emerged, creating an urgent need for a highly transferable, fast, and robust 6D pose estimation method for visual tasks such as UAV swarming and monitoring. To address this, this paper presents a robust method for estimating the 6D pose of a UAV relative to an airborne camera based on hybrid features. Firstly, we employ KAPAO3D to simultaneously obtain pose objects using 3D losses with implicit keypoint associations and keypoint objects predicted based on explicit semantic information. Leveraging prior pose object knowledge enables the establishment of precise 2D-3D correspondences, facilitating the creation of fusion objects that alleviate missed detections and false positives in keypoint object predictions. Secondly, pose objects and fusion objects are respectively utilized to construct point and edge constraints to obtain an initial pose. We minimize the 3D errors of hybrid features with predicted confidence scores to further improve accuracy. Experimental results demonstrate that our method adapts to various types of UAVs and achieves superior performance, with higher ADD and Te metrics compared to other methods and a running speed of 30fps. This method is expected to have potential practical applications in high-precision real-time tracking tasks.

During astronomical observation, the centroiding of space targets is a crucial technology for improving trajectory association in images, celestial positioning, and orbit determination. However, when the exposure time is short, the distribution of targets and stars in different regions of an image can vary significantly, seriously affecting centroiding accuracy. Motivated by the relationship between target distribution and the corresponding optical system, we propose a centroiding method based on local point spread function (PSF) estimation and iterative deconvolution. First, the imaging process of an ideal target is analyzed to identify key factors affecting the distribution of various targets. Next, the local PSF around a target is estimated using the distribution of nearby stars. Then, iterative deconvolution is used to reconstruct the region of interest (ROI) around the target with the estimated local PSF. Finally, an improved weighted centroid is calculated using a restricted region growing method. The proposed method was tested on simulation experiments and multiple sequences of collected real optical images with stars and targets in different regions. The results showed outstanding performance in terms of mean absolute errors, standard deviation of star location, and movement fluctuations compared to baseline methods. The mean absolute errors for centroiding in Gaussian noise with a standard deviation of less than 0.02 and displacement within 1 pixel are 0.16 and 0.06 pixels, respectively.

Astronomical image registration plays an important role in space surveillance tasks. However, long-term observation introduces challenges such as smearing, which significantly impact the effectiveness of current image registration techniques. In this letter, a novel method for registering smeared astronomical images without additional preprocessing is developed for ground-based space surveillance. Initially, star points are extracted and classified based on an analysis of the imaging mechanisms behind smearing. Subsequently, several of the most affected stars are selected to form stable triangles, facilitating the search for initial homogenous points. Finally, based on the initial outcomes, we straightforwardly proceed to create and match normal triangles. By combining all the homogenous points, homography transformation can be solved and optimized. The proposed method takes less than 30 milliseconds to process two affected images, each with a resolution of 3056×3056 pixels. It achieves superior levels of accuracy and precision with less time than three triangle-based approaches. The results show the strong noise rejection and stability of the method, indicating it can markedly enhance the efficacy of space surveillance for long-duration observations.

The temporal and spatial relationships among various onboard sensors are crucial for ensuring the accuracy of visual observations carried out by the multiple unmanned aerial vehicles (multi-UAV) system. To leverage the integration within the multi-UAV system, we propose a buddy calibration method focused on calibrating the relative time offset and rotation parameters between the airborne camera and flight controller. We designate another UAV as the target and perform 6D pose estimation to calculate the relative pose between the camera mounted on the host UAV and the target UAV. Subsequently, we use the pose estimation results and the parameters recorded by the sensors on both UAVs to establish a comprehensive temporal and spatial calibration model. Finally, a specific method is developed to effectively solve the proposed model. Our approach was validated using two-UAV flight experiments. The results demonstrate a significant improvement, with leader attitude estimation errors decreasing to approximately less than 1 $^{\circ }$ post-calibration compared to pre-calibration. Furthermore, positioning errors for dynamic and static targets were substantially reduced over 42.59% and 53.36%, respectively.

In the testing of chips, defect diagnostics in X-ray images of packaging chips is mainly performed by humans, which is time-consuming and inefficient. To overcome the abovementioned problems, a novel intelligent defect diagnostics system based on hybrid deep learning for chip X-ray images was proposed. The system consists of four successive stages: image segmentation and normalization, image reconstruction and defect detection, contour matching, and qualification diagnosis. The first stage is used to localize the external contours of the target chip and remove extraneous backgrounds through the improved UNet. Then, considering the variety of defects and the complexity of labeling, an unsupervised learning model is designed to reconstruct defect-free images to detect defects, which requires only normal samples for training. Third, the multicomponent template matching based on structural prior is used to localize the internal contours of the chip. In the final stage, the qualification is diagnosed based on the previous results through the Floyd–Warshall algorithm. The effectiveness and robustness of the proposed methods are verified by experiments on real-world inspection lines. The experimental results demonstrate that the developed system can successfully perform fault diagnostics tasks, achieving a judgment accuracy of 92.5%.