Zhihua Shen

Multi-frame infrared small target detection (IRSTD) plays a crucial role in low-altitude and maritime surveillance. The hybrid architecture combining CNNs and Transformers shows great promise for enhancing multi-frame IRSTD performance. In the paper, we propose LVNet, a simple yet powerful hybrid architecture that emphasizes low-level feature learning in hybrid frameworks for multi-frame IRSTD. Our key insight is that the standard linear patch embeddings in Vision Transformers are insufficient for capturing the scale-sensitive local features critical to infrared small targets. To overcome this limitation, we design a multi-scale CNN frontend that explicitly models local features by leveraging the local spatial bias of convolution. Additionally, we introduce a U-shaped video Transformer for multi-frame spatiotemporal context modeling, effectively capturing the motion characteristics of targets. To the best of our knowledge, we are the first to introduce the concept of “low-level matters” within the CNN-Transformer hybrid architecture for multi-frame IRSTD. Despite its simplicity, with only 1.77 M parameters and 17.88 G FLOPs, LVNet achieves state-of-the-art (SOTA) performance at a speed of 208 FPS. Experiments on multiple public datasets, comparing LVNet with 18 SOTA methods, demonstrate its superior performance and efficiency. Our code and pre-trained models are accessible at https://github.com/ZhihuaShen/LVNet.

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.

Infrared small target detection (IRSTD) is the pivotal technology for remote surveillance and search missions under infrared imagery. Due to the target’s nature of being small, dim, and lacking texture information, detection algorithms for relatively large visible targets are not suitable for infrared small targets. In this letter, a real-time, robust IRSTD network with feature enhancement and augmentation, FEA-Net, is proposed, which includes a hybrid multiscale enhancement (HME) module and a saliency detection network with a small target feature augmentation (STFA) module. The HME introduces a hybrid multiscale spatial attention mechanism to obtain a rough estimation of targets and then apply it to the original image. The STFA is directed against the insufficiency of target-related features and the consequent feature losses, which screens out the crucial information and represents it into a larger tensor for further feature fusion, increasing the proportion of target-related features of the output. The results show that the proposed method can achieve the same level of detection effect as the comparable methods and still has considerable real-time performance with the normalized intersection over union (nIoU) of 76.12% on the SIRST dataset and 92.28% on the NUDT-SIRST dataset.

With the increasing amount of space debris in near-earth space, the surveillance of space has become more crucial to prevent potential space collisions. In an optical surveillance system, a multitude of debris with varying sizes, speeds, and intensities exists within images, which brings great difficulties for the extraction of targets. Current methodologies mainly emphasize local saliency enhancement and source extraction based on global statistical attributes, yet often overlook the distribution similarity of individuals and how to map them in a standardized space. To address this limitation, this article introduces an innovative transformation from the grayscale space to the correlation space. Initially, a fourfold-structure model is designed to measure the local correlation for each element in the grayscale space. Subsequently, the standardized correlation space is constructed by correlation measurement. Then, combining an adaptive threshold choosing method based on statistics and a threshold limitation strategy based on correlation, the individuals can be separated from the feature space. Finally, a deblending method is applied to disentangle merged individuals based on the local gradient correlation. Thoroughly assessed using ten subimage sequences featuring targets with typical traits against complex backgrounds, the results confirm its superior extraction capabilities and effectiveness when compared to ten common baseline methods, which achieves an average 90.9% true positive rate in just 2.8 ms. Furthermore, the extraction outcomes serve as a basis for detecting true target trajectories in continuous sequences. This integration holds significant implications for practical applications in space surveillance engineering.