Collaborative Static-Dynamic Teaching: A Semi-Supervised Framework for Stripe-Like Space Target Detection

TL;DR

This work tackles stripe-like space target detection (SSTD) under low SNR and stray-light noise by introducing a collaborative static-dynamic teaching (CSDT) SSL framework with dual teachers and adaptive pseudo-labeling (APL). The approach integrates MSSA-Net, featuring multi-scale stripe-aware MDPC blocks and feature map weighted attention (FMWA), to robustly extract stripe patterns. Through EMA-based dynamic teaching and an adaptive pseudo-labeling strategy, CSDT delivers state-of-the-art performance on AstroStripeSet and demonstrates strong zero-shot generalization on real-world datasets. The combination of dual teachers, EMA feedback, and MSSA-Net components significantly enhances pseudo-label quality and SSTD robustness, reducing labeling requirements for high-performance space-target detection.

Abstract

Stripe-like space target detection (SSTD) is crucial for space situational awareness. Traditional unsupervised methods often fail in low signal-to-noise ratio and variable stripe-like space targets scenarios, leading to weak generalization. Although fully supervised learning methods improve model generalization, they require extensive pixel-level labels for training. In the SSTD task, manually creating these labels is often inaccurate and labor-intensive. Semi-supervised learning (SSL) methods reduce the need for these labels and enhance model generalizability, but their performance is limited by pseudo-label quality. To address this, we introduce an innovative Collaborative Static-Dynamic Teacher (CSDT) SSL framework, which includes static and dynamic teacher models as well as a student model. This framework employs a customized adaptive pseudo-labeling (APL) strategy, transitioning from initial static teaching to adaptive collaborative teaching, guiding the student model's training. The exponential moving average (EMA) mechanism further enhances this process by feeding new stripe-like knowledge back to the dynamic teacher model through the student model, creating a positive feedback loop that continuously enhances the quality of pseudo-labels. Moreover, we present MSSA-Net, a novel SSTD network featuring a multi-scale dual-path convolution (MDPC) block and a feature map weighted attention (FMWA) block, designed to extract diverse stripe-like features within the CSDT SSL training framework. Extensive experiments verify the state-of-the-art performance of our framework on the AstroStripeSet and various ground-based and space-based real-world datasets.

Figures (6)

• Figure 1: Representative space images with low signal-to-noise ratio stripe-like targets and their 3D pixel-level views under different stray light noise. (a) Space images affected by stray light and their 3D pixel-level views. (b) Detailed local images of stripe-like targets and their 3D pixel-level views. (c) Detailed local images of stars and stray light, along with their 3D pixel-level views. The stripe-like targets are highlighted with the dotted red circle.
• Figure 2: Comparison of SSL architectures with MSSA-Net as the teacher-student network: (a) Single-teacher architectures based on exponential moving average (EMA) strategies, including Mean Teacher (MT) tarvainen2017mean, Unbiased Teacher (UT) liu2021unbiased, Interactive Self-Training Mean Teacher (ISMT) yang2021interactive, and Pseudo-Label Mean Teacher (PLMT) mao2023semi. (b) Self-Training (ST) yang2022st, a single-teacher architecture without EMA. (c) Our proposed Collaborative Static-Dynamic Teaching (CSDT) architecture utilizing EMA. Panels (d), (e), (f) show the performances of these SSL architectures in terms of Dice score, mIoU, and detection rate $P_d$ across various labeling rates.
• Figure 3: Overview of our proposed CSDT SSL framework. (a) CSDT uses both labeled images $\mathcal{D}_l$ and unlabeled images $\mathcal{D}_u$ for SSL training, where the ST model is a pre-trained teacher whose weights remain fixed during training. The DT and S model are trainable, and the parameters of the DT model are updated through the EMA strategy using the parameters from the S model. (b) Details of the APL strategy implementation, which involves point set extraction and centralization, followed by covariance and linearity calculations. (c) ST uses a small subset of labeled images $\mathcal{D}_l$ for pre-training. (d) During inference, only the well-trained DT model is utilized.
• Figure 4: Overall configuration of MSSA-Net comprises two main parts: the encoder and the decoder. The encoder is used for multi-scale stripe feature extraction, while the decoder engages in multi-level feature fusion. (a) The internal architecture of the proposed MDPC block, designed to expand the receptive field and extract multi-scale stripe features. (b) The internal architecture of the ConvBN block. (c) The internal architecture of the proposed FMWA block, which enhances the stripe-like target regions in the feature map through feature reconstruction and suppresses noise.
• Figure 5: Visualization results of different SSL methods on AstroStripeSet using three network configurations: UNet, UCTransNet, and MSSA-Net as segmentation networks. Evaluations are conducted across these networks at a labeling rate of 1/4. (a) Visualization of detection results under sunlight interference. (b) Visualization of detection results under earth light interference. (c) Visualization of detection results under moonlight interference. (d) Visualization of detection results under mixed light interference. (Detected real stripe-like targets are highlighted in the dotted red circle, missed detections are marked in blue, and false alarms are presented in yellow.)