基于信息表征增强的空间弱小目标检测方法

In recent years, deep learning based object detection technologies have achieved significant advancements and have been widely utilized in various fields. However, there remains a lack of deep learning detection methods specifically designed for space targets. Compared to object detection in natural image, space target detection encounters unique challenges: One primary challenge arises from the extreme shooting distances involved, which results in space targets appearing as mere faint spots in space images, comprising only a few pixels and lacking distinct shape and color features. Additionally, the complex space environment and device-related factors contribute to various types of noise within the images, such as bright backgrounds caused by stray light and hot pixel noise arising from camera issues. Moreover, some regions in space contain a large number of stars, the densely packed stars within the field of view frequently lead to the partial overlapping of light spots in space images, further complicating the detection task. These difficulties undeniably exacerbate the challenges of space target detection. In response to these challenges, this paper introduces an anchor-free detection method for small space objects, grounded in enhanced information representation. The proposed method employs a specifically designed end-to-end convolutional neural network (CNN) model, capable of concurrently detecting small space targets and precisely locating their centroids. To tackle the issues of weak object signals and background noise interference, we have developed a cross-space-channel attention module and a squeeze-excitation multi-scale feature fusion module. These components are designed to enhance the model's ability to focus on the pertinent information within noisy images, thereby significantly improving its capability to detect targets in the challenging noise-laden backgrounds. Building upon this, to further address the problem of overlapping light spots, we incorporate a density map-based loss function. This approach enables the model to more effectively learn the Overall spatial distribution and quantity information of the targets within an image. As a result, the model accurately discerns the number of individual targets within overlapping Clusters, facilitating more precise differentiation and identification of each distinct target. To comprehensively validate the effectiveness of the proposed method, we simulated space images containing point targets and streak targets with various noisy backgrounds. These images were meticulously annotated with detailed information regarding the centroid locations, bounding boxes, and pixel coordinates of the targets. Experimental validations were conducted using both the simulated dataset and the publicly available real image sequences dataset SpotGEO. In testing conducted on the simulated image data, our method achieved a Fl score of 95. 34% and a sub-pixel level average centroid localization accuracy of 0. 4478. Additionally, we performed further tests to examine the impact of varying noise levels on centroid localization accuracy, as well as an analysis of processing efficiency under different hardware conditions to provide a comprehensive assessment of the method's capabilities. Furthermore, within the publicly available SpotGEO dataset, after integrating a sequential post processing method, our method yielded an Fl score of 93. 08%. Experimental results demonstrate the superior Performance of our approach in accurately detecting and precisely localizing these small but crucial space objects.