SAANet: Detecting Dense and Crossed Stripe-like Space Objects Under Complex Stray Light Interference

Highlights

What are the main findings?

• A novel Shape-Aware Attention Network (SAANet) is proposed, establishing a unified paradigm to effectively resolve the challenges of complex stray light interference and dense overlapping trajectories.
• SAANet achieves state-of-the-art performance in challenging scenarios with intense stray light interference and dense trajectories, demonstrating superior precision with APs of 0.864 and 0.815 on benchmark datasets.

What is the implication of the main finding?

• This study provides a high-precision and robust technical solution for Space Situational Awareness (SSA), enhancing the ability to monitor and manage the increasingly crowded space environment.

Abstract

With the deployment of mega-constellations, the proliferation of on-orbit Resident Space Objects (RSOs) poses a severe challenge to Space Situational Awareness (SSA). RSOs produce elongated and stripe-like signatures in long-exposure imagery as a result of their relative orbital motion. The accurate detection of these signatures is essential for critical applications like satellite navigation and space debris monitoring. However, on-orbit detection faces two challenges: the obscuration of dim RSOs by complex stray light interference, and their dense overlapping trajectories. To address these challenges, we propose the Shape-Aware Attention Network (SAANet), establishing a unified Shape-Aware Paradigm. The network features a streamlined Shape-Aware Feature Pyramid Network (SA-FPN) with structurally integrated Two-way Orthogonal Attention (TTOA) to explicitly model linear topologies, preserving dim signals under intense stray light conditions. Concurrently, we propose an Adaptive Linear Oriented Bounding Box (AL-OBB) detection head that leverages a Joint Geometric Constraint Mechanism to resolve the ambiguity of regressing targets amid dense, overlapping trajectories. Experiments on the AstroStripeSet and StarTrails datasets demonstrate that SAANet achieves state-of-the-art (SOTA) performance, achieving Recalls of 0.930 and 0.850, and Average Precisions (APs) of 0.864 and 0.815, respectively.

1. Introduction

With the rapid advancement of space technology, the technology for high-function-density small satellites has matured. Consequently, mega-constellations composed of thousands or even tens of thousands of small satellites, such as SpaceX’s Starlink, are being rapidly deployed [1,2,3]. The establishment of these constellations has vastly increased the quantity of on-orbit spacecraft and, simultaneously, the amount of space debris. The scale of this increase is one to two orders of magnitude greater than the total number of spacecraft launched since humanity entered the space age [4,5]. This rapid growth has led to a sharp increase in the number of RSOs in orbital space, posing a significant threat to the existing space environment [6].

The detection of space objects using optical imaging systems has long been a critical area of research. For long-exposure observations, these optical systems primarily operate in two modes: target-tracking and star-gazing. In target-tracking mode, as illustrated in Figure 1b, the camera tracks a specific RSO, rendering it as a point while background stars form streaks [7].

We focus on the star-gazing mode, illustrated in Figure 1a, where the observation platform maintains a fixed orientation relative to the celestial sphere. Due to their immense distance, stars exhibit negligible angular motion and thus appear as points. In contrast, RSOs orbiting the Earth have significant relative motion, causing their trajectories to be integrated over the exposure time and appear as distinct stripes [9]. The primary focus of this research is the detection of these stripe-like space resident targets under the star-gazing mode, a task we refer to as Stripe-like RSO Target Detection (SRTD).

The precise SRTD is critical for SSA and Space Traffic Management (STM) [10]. By interpreting a stripe’s geometric properties, such as length and orientation, one can extract kinematic data like velocity and motion vectors. This information is foundational for proactive STM tasks like collision warning and orbit prediction, elevating space surveillance from simple discovery to dynamic management. However, this detection task is complicated by two main challenges: environmental interference and the increasing number of RSOs.

1.1. Challenge 1: Complex Stray Light Interference

In the space environment, the detection of dim RSOs is severely hampered by complex stray light interference [4]. Long-exposure imaging, while necessary to capture dim targets, inevitably amplifies various types of environmental noise.

Specifically, intense background clutter arises from a diverse array of space stray lights and sensor disturbances, such as reflections from the Earth, Moon, and Sun, as well as dense star background interference shown in Figure 2. These interference sources manifest as non-uniform, high-intensity noise patterns that can overwhelm the faint signatures of RSOs [11]. Consequently, the targets are often submerged in the background with extremely weak contrast, making it difficult for standard detectors to distinguish valid stripe features from environmental disturbances.

1.2. Challenge 2: Dense and Overlapping Trajectories

The dramatic increase in the number of RSOs has led to the presence of many complex scenarios within the captured images [8]. In these scenarios, the trajectories of multiple RSOs overlap, making the precise detection of individual stripe-like targets difficult, as illustrated in Figure 3.

As can be seen in Figure 3, the dense accumulation of streaks, combined with their varying SNRs, makes it impossible to accurately identify each target. This constitutes another core challenge in the detection of stripe-like RSOs.

1.3. Limitations of Existing Methods

To address these challenges, existing research methods can be broadly categorized into three types [4,9,12]. However, traditional unsupervised methods [12,13] and semi-supervised learning strategies [14,15] often struggle in complex space environments. Relying on hand-crafted features or unstable pseudo-labels, these approaches lack the robustness to handle complex stray light interference and fail to adapt to variable target morphologies, resulting in limited generalization ability.

Consequently, deep networks based on fully supervised learning have become the mainstream approach due to their superior precision [16,17]. Nevertheless, current SOTA networks are predominantly optimized for generic or point-like infrared targets. They inherently lack the geometric modeling capabilities required to capture the dim features of stripe-like RSOs or to resolve the ambiguity of dense, overlapping trajectories. This fundamental mismatch between existing tools and the unique properties of RSOs highlights the urgent need for a specialized shape-aware architecture.

To effectively address the aforementioned challenges, we propose the Shape-Aware Attention Network (SAANet). The primary contributions are summarized as follows:

• We construct a specialized framework by integrating a streamlined Dense Nested Attention Network (DNANet) encoder with a Shape-Aware Feature Pyramid Network (SA-FPN). This design structurally embeds Two-way Orthogonal Attention (TTOA) to explicitly model linear topologies, efficiently preserving dim signals and overcoming complex interference bottlenecks.
• We propose an Adaptive Linear Oriented Bounding Box (AL-OBB) head. By formulating a joint geometric constraint mechanism via collinearity and aspect ratio, this strategy achieves precise target separation and localization, effectively resolving ambiguities in dense and intersecting scenarios.
• We perform extensive validation on the AstroStripeSet and StarTrails datasets. Results demonstrate that SAANet achieves SOTA performance, exhibiting superior accuracy and robustness compared to existing methods.

2. Materials and Methods

This section aims to review the related technologies, with a focus on discussing the traditional unsupervised learning paradigm, fully supervised learning strategies, and semi-supervised learning methods. By analyzing the limitations of these approaches in complex space environments, we identify the critical research gaps that motivate our proposed shape-aware framework.

2.1. Traditional Unsupervised Methods for SRTD

Traditional unsupervised methods for stripe-like target detection primarily solve the problem by designing specific features and filtering strategies. One such approach is the customized stripe pattern filtering method [18,19]. This method designs specific filter templates to match the expected stripe patterns in an image, thereby enabling the recognition and localization of stripe targets. It exhibits high efficiency and accuracy when dealing with specific targets that have a fixed morphology and well-defined features. However, due to the customized nature of the templates, this method struggles to adapt to the diversity of stripe targets. Its performance degrades sharply when the target’s shape or angle changes, and it is extremely sensitive to parameter settings.

Alternatively, methods based on morphological principles can be used [20]. For example, by applying morphological operations with a linear structuring element oriented parallel to the stripe’s direction, a stripe target can be enhanced while point-like stars and noise in the background are suppressed [21]. This approach demonstrates a certain degree of robustness when processing images with intense stray light interference. Nevertheless, its performance is heavily dependent on the design of the structuring element and it has difficulty adapting to the complex shapes and varying widths of stripe targets.

In summary, traditional unsupervised methods for SRTD are fundamentally limited by their reliance on rigid, hand-crafted geometric priors. This inflexibility prevents them from adapting to complex scenarios, as they struggle to disentangle overlapping targets and require separately designed filters for diverse backgrounds, ultimately leading to poor generalization and robustness.

2.2. Semi-Supervised Methods for Object Detection

In recent years, to address the challenge of large-scale data annotation, semi-supervised learning methods have been widely adopted in numerous fields, including object detection and image segmentation [1,2]. This paradigm significantly enhances a model’s generalization ability by combining a limited amount of labeled data with a large volume of unlabeled data.

Current mainstream semi-supervised learning approaches can be divided into two main categories: methods based on consistency regularization [22] and those based on pseudo-labeling [23,24]. Sajjadi et al. [22] first proposed the consistency regularization method, which compels the model to learn more robust feature representations by minimizing the prediction discrepancies between different perturbations of the same unlabeled data. The Mean Teacher framework is a classic example of this approach. However, such methods tend to be overconfident about low-quality or noisy labels, which affects the capture of the true structure of the data.

Lee et al. [23] first proposed the pseudo-labeling method, which utilizes a model trained on a small amount of labeled data to generate “pseudo-labels” for unlabeled data, subsequently incorporating this newly labeled data into the training process. The effectiveness of this approach is highly dependent on the quality of the pseudo-labels.

Early self-training methods were prone to errors due to incorrect pseudo-labels. To solve this problem, Liu et al. [25] and Xu et al. [26] introduced an end-to-end strategy based on the Mean Teacher (MT) framework [27], proposing the use of an Exponential Moving Average (EMA) mechanism to smooth the outputs of the teacher model. More recently, Zhu et al. introduced the MRSA-Net [28] and the CSDT semi-supervised framework to address SSTD under conditions of complex light pollution and data scarcity.

In summary, while semi-supervised learning reduces data annotation requirements, its performance is highly dependent on the quality of the unlabeled data. Since these methods rely on the model’s own predictions for self-supervision, they are prone to introducing and amplifying errors when data is noisy, leading to performance instability and limiting their reliability in complex SRTD scenarios.

2.3. Fully Supervised Methods for Object Detection

With the rise of Convolutional Neural Networks (CNNs), a number of deep learning-based methods for space object detection have been proposed. Unlike traditional methods that rely on hand-crafted filters, CNNs can automatically extract high-level abstract features from large-scale labeled data, allowing them to maintain excellent robustness when faced with diverse target morphologies and complex backgrounds.

Jia et al. [29] were the first to apply CNNs to space scenes, proposing a framework based on Faster R-CNN to classify point-like stars and stripe-like space objects. Wu et al. [30] proposed a simple yet effective “U-Net in U-Net” framework (UIU-Net), which embeds a smaller U-Net into a larger U-Net backbone. This design achieves multi-level [31], multi-scale representation learning for small infrared targets, addressing the issues of small target disappearance and limited feature differentiability in deep networks. To solve the problems of data scarcity and class imbalance in infrared small target detection, Sun et al. [32] constructed a large-scale dataset named IRDST and proposed a novel deep learning model called the Receptive Field and Direction-induced Attention Network (RDIAN). Li et al. [33] introduced a fully supervised method named the Dense Nested Attention Network (DNA) [33], which resolves the issue of small infrared target loss in deep networks by designing a Dense Nested Interactive Module (DNIM).

In essence, these fully supervised methods demonstrate the significant power of deep learning. By leveraging large-scale data, CNN-based approaches with innovative nested or attention-based architectures have proven highly effective for detecting small and dim targets. However, a critical review reveals that these state-of-the-art networks were predominantly designed for and evaluated on point-like small targets. Their architectural optimizations are therefore not tailored for targets with a distinct elongated morphology, such as the stripe-like RSOs central to the SRTD task. This fundamental mismatch between existing tools and the specific target shape constitutes a key research gap.

2.4. Summary of Related Works and Research Gap

In summary, existing methods for SRTD face significant limitations, revealing a clear research gap. Unsupervised methods lack robustness in complex backgrounds, while semi-supervised approaches are hampered by unreliable pseudo-labels. Even the most promising fully supervised methods are fundamentally mismatched for SRTD, as current architectures are designed for point-like targets and fail to address the unique elongated morphology and overlapping arrangements of RSOs. This highlights a critical need for a network architecture specifically tailored to the shape-aware detection of stripe-like targets.

Our research introduces and improves a network for infrared small target detection, directly addressing these methodological limitations and providing a more flexible, accurate, and computationally efficient approach for target detection. This allows for more precise detection of RSOs with a smaller number of parameters and floating-point operations.

3. Proposed Framework

To address the unique challenges of detecting stripe-like RSOs, we propose the SAANet. This architecture is established upon a unified Shape-Aware Paradigm, explicitly modeling the linear topology of targets. As illustrated in Figure 4, the framework comprises three theoretically aligned modules:

First, in the encoder stage, SAANet implements Streamlined Feature Maintenance to mitigate the loss of dim signals inherent in standard CNNs. By restructuring the DNANet encoder into a specialized feed-forward backbone without redundant upsampling, we ensure the high-fidelity preservation of faint features during the initial extraction phase.

Subsequently, in the decoder stage, to address the problem of traditional FPNs being shape-agnostic and diluting key morphological information, the multi-scale features are processed by our SA-FPN. Functioning as an Orthogonal Directional Modeler, this module structurally integrates TTOA to adaptively aggregate and enhance the target’s stripe-like morphology. This process transforms passive feature fusion into active topological refinement, offering a discriminative feature foundation for subsequent localization.

Finally, in the detection head stage, to resolve the inaccurate and unstable localization of elongated targets common with Horizontal Bounding Boxes (HBBs) [34] and standard Oriented Bounding Boxes (OBBs) [35], the feature pyramid is fed into the AL-OBB head. This module introduces a Joint Geometric Constraint Mechanism into the dynamic point set representation. By utilizing a linear-aware mechanism to impose strict collinearity and aspect ratio priors, it adaptively generates OBBs that are highly consistent with the physical morphology of RSOs, achieving precise separation and stable detection.

Figure 4. SAANet Framework Diagram.

Figure 4. SAANet Framework Diagram.

3.1. DNANet Encoder

Standard Convolutional Neural Networks (CNNs), such as ResNet, often cause the feature information of small, dim targets to be progressively lost in the deeper layers due to repeated pooling and strided convolutions. For stripe-like space objects that are only a few pixels wide, effectively preserving these dim signals is critically important. Therefore, we adopt the feature extraction philosophy of DNANet [31], which features a dense nested structure designed to prevent information loss. However, considering the specific characteristics of stripe-like targets and computational efficiency, we do not adopt the original architecture entirely. Instead, we introduce a streamlined version of the DNANet encoder, optimized to balance feature retention with inference speed.

3.1.1. Dense Nested Interactive Module

The Dense Nested Interactive Module (DNIM) is the core component of the encoder. Distinct from the original DNIM which employs tri-directional interactions (including bottom-up upsampling) between nodes, we implement a streamlined feed-forward structure, as illustrated in Figure 5. Specifically, we remove the computationally redundant upsampling pathways within the encoder to focus on pure feature extraction.

In our modified design, each node $L^{(i,j)}$ aggregates features mainly from shallower layers and the current layer. For the initial column ($j=0$), the node receives features only from the previous layer via max-pooling:

$$L^{(i,j)}={\mathcal{P}}_{max}\left(\mathcal{F}\left(L^{(i-1,j)}\right)\right),$$

(1)

where $\mathcal{F}(·)$ denotes a convolution operation, and ${\mathcal{P}}_{max}(·)$ denotes a max-pooling operation.

For nodes at deeper nested positions ($j>0$), we simplified the interaction to receive inputs from only two directions: the same layer and the shallower layer. The computation is redefined as:

$$L^{(i,j)}=\mathcal{F}\left(\left[L^{(i,j-1)},{\mathcal{P}}_{max}\left(\mathcal{F}\left(L^{(i-1,j)}\right)\right)\right]\right),$$

(2)

where $[·]$ denotes the concatenation operation. Note that the up-sampling term from the deeper layer $L^{(i+1,j-1)}$ present in the original DNANet is omitted here, and the input from the shallower layer $L^{(i-1,j)}$ is consistently processed via max-pooling to match the feature map resolution.

3.1.2. Cascaded Channel and Spatial Attention Module

To further enhance the representation of dim stripe-like targets, we integrate the Cascaded Channel and Spatial Attention Module (CSAM) into the convolutional blocks of our encoder. Based on the widely adopted Convolutional Block Attention Module (CBAM) design, this module sequentially infers attention maps along two separate dimensions: channel and spatial. By multiplying the input feature map with these attention maps, the network adaptively recalibrates features, emphasizing target-relevant information while suppressing background noise. This integration ensures that dim signals are amplified within the feature extraction stage, preventing them from being overwhelmed by background clutter.

3.2. Shape-Aware Feature Pyramid

Following the streamlined feature extraction, we propose the Shape-Aware Feature Pyramid Network (SA-FPN) to explicitly model the topological orientation of stripe-like targets. Standard Feature Pyramid Networks (FPN) rely on element-wise addition for feature fusion, which treats spatial information isotropically. However, this mechanism inherently dilutes the critical directional information required to distinguish stripe-like RSOs from noise. To overcome this theoretical limitation, we redesign the top-down pathway by constructing a shape-aware orthogonal fusion mechanism.

Instead of simple linear superposition, we structurally integrate the TTOA module [36] as the core fusion operator. Crucially, this integration establishes a global recursive directional refinement process. As illustrated in Figure 6, the SA-FPN accepts four hierarchical feature maps $\{C_2,C_3,C_4,C_5\}$ from the encoder as input. Through the top-down pathway, the high-level feature map is fused with the current stage features after upsampling to generate the pyramid levels $\{P_2,P_3,P_4,P_5\}$. Additionally, to distinctively expand the receptive field for ultra-long trajectories, we generate a fifth level, $P_6$, by applying max-pooling operation on $P_5$. Consequently, the final output consists of five scale levels $\{P_2,P_3,P_4,P_5,P_6\}$. This design ensures that the directional shape priors captured at coarser scales are explicitly propagated and preserved to guide the feature refinement at finer scales. By iteratively modeling row-wise and column-wise correlations across scales, SA-FPN transforms the fusion process from passive aggregation into active topological reconstruction.

The detailed mechanism of the TTOA fusion unit is shown in Figure 6. It contains two parallel branches: one branch applies a $1\times k$ deformable convolution to capture contextual information in the horizontal (row-wise) direction, while the other applies a $k\times 1$ deformable convolution to capture contextual information in the vertical (column-wise) direction. These attention maps then modulate the high-level features, effectively guiding the fusion process to focus on features consistent with the elongated structure of the target.

The output of the TTOA module, $a_{\mathrm{TTOA}}$, is the sum of the outputs from the two attention branches: row and column:

$$a_{\mathrm{TTOA}}=\mathrm{TTOA}(a_{\mathrm{low}},a_{\mathrm{high}})=a_{\mathrm{row}}+a_{\mathrm{column}}$$

(3)

The attention features for each direction are calculated as follows:

$$a_{\mathrm{row}}=S\left(F_c\left(F_h\left(a_{\mathrm{low}}\right)\right)\right)\otimes F_h\left(a_{\mathrm{high}}\right)+F_c\left(a_{\mathrm{low}}\right)a_{\mathrm{column}}=S\left(F_c\left(F_v\left(a_{\mathrm{low}}\right)\right)\right)\otimes F_v\left(a_{\mathrm{high}}\right)+F_c\left(a_{\mathrm{low}}\right)$$

(4)

where $S(·)$ is the Sigmoid function. $F_c(·)$ is a bottleneck structure containing a $1\times 1$ convolution for noise and channel reduction. $F_h$ is a $1\times k$ deformable convolution, $F_v$ is a $k\times 1$ deformable convolution, and ⊗ denotes element-wise multiplication. The kernel size k is a critical hyperparameter governing the receptive field of the TTOA module. A brief ablation study indicated that when $k<7$, the receptive field fails to cover the full context of elongated stripe targets. Conversely, when $k>15$, redundant background noise is introduced, interfering with feature aggregation. We found that setting $k=13$ effectively balances feature integrity and noise resistance, and thus adopted it as the default setting.

3.3. Adaptive Linear Oriented Bounding Box

While the dynamic point-set representation in DC-OBB offers flexibility for general object detection, it lacks specific geometric constraints for elongated objects, leading to unstable regression in dense RSO scenarios. To bridge this gap, we propose the AL-OBB detection head by incorporating a joint geometric constraint mechanism. This mechanism imposes a strong linear structure prior onto the adaptive point learning process. By forcing the representative points to align strictly with the target’s principal axis, our method resolves the ambiguity issues inherent in standard OBB approaches when dealing with high-aspect-ratio and overlapping trajectories.

The workflow of the AL-OBB detection head is divided into two closely coupled stages. The first is the generation and conversion of an adaptive point set, where a representative point set learned by the network is transformed into an OBB that precisely encloses the target. The second stage is a quality assessment and selection process based on enhanced linear awareness, which further ensures that the selected bounding box accurately reflects the elongated and directional characteristics of the stripe-like target. A detailed flowchart of this process is shown in Figure 7.

Through this approach, the AL-OBB detection head provides more accurate and stable detection results for elongated targets, making it particularly suitable for application scenarios where traditional HBB and OBB methods perform poorly.

3.3.1. Adaptive Point Learning

First, for each potential target in the image, a set of representative points is generated that can flexibly describe its morphology. Specifically, the detection head employs a Dynamic [26] Convolutional Network (DCN) to learn a position-dependent offset for each location on the feature map. At each location, we predefine an initial regular grid of N = 9 points as

$${\tilde{P}}_i={\left\{(x_i^n,y_i^n)\right\}}_{n=1}^N$$

(5)

where $(x_i^n,y_i^n)$ represents the 2D coordinates of the k-th initial point for the i-th potential target.

Simultaneously, the network learns a set of offsets, $\Delta p_i$, which are applied to this initial grid to adjust it, resulting in the final adaptive representative point set, $P_i$. This process can be expressed by the following formula:

$$P_i={\{{\tilde{P}}_i+\Delta p_i\}}_{n=1}^N$$

(6)

This adaptive approach allows the point set to distribute itself freely according to the target’s actual shape and orientation, greatly enhancing the fitting capability for stripe-like targets with variable morphologies. After obtaining the adaptive point set $P_i$, which accurately outlines the target’s contour, we convert it into a final OBB using standard computational geometry algorithms.

First, the minimum convex hull $C_i$ of the point set $P_i$ is constructed using an algorithm such as the Jarvis March, ensuring that all representative points are contained within it. Subsequently, the Rotating Calipers algorithm is applied to the convex hull $C_i$ to efficiently find the enclosing rectangle with the minimum area that envelops the hull. This minimum-area enclosing rectangle is the desired OBB, denoted as $O_i=(x_c,y_c,w,h,\theta )$.

3.3.2. Joint Geometric Constraint Mechanism

The Joint Geometric Constraint Mechanism is the theoretical core of the AL-OBB head, distinguishing it from general-purpose object detectors. We posit that for a point set to accurately represent a stripe-like RSO, merely satisfying collinearity is insufficient, as it fails to distinguish between a compact noise cluster and an elongated streak. Therefore, we formulate a novel joint quality metric, $Q_{line}^{\prime }$, which integrates two orthogonal geometric dimensions: collinearity and aspect ratio. This metric functions as a shape-aware regularizer during the learning process, comprehensively evaluating the topological consistency of the predicted point set.

The first step is to quantify how closely the spatial distribution of $P_i$ approximates a straight line, evaluating the collinearity of the point set as the basis for an initial metric. First, the centroid $c_i$ of the point set $P_i$ is calculated, and all points are centralized to obtain ${\overline{P}}_i$, as computed by the following formulas:

$$c_i={\displaystyle \frac{1}{N}}\sum _{n=1}^N{P}_i^n$$

(7)

$${\overline{P}}_i=P_i-c_i={\left\{({\overline{x}}_i^n,{\overline{y}}_i^n)\right\}}_{n=1}^N$$

(8)

where ${\overline{x}}_n=x_n-\overline{x},{\overline{y}}_n=y_n-\overline{y}$.

Subsequently, we calculate the second-order moments of the centralized point set, $x{x}_i$, $y{y}_i$, $x{y}_i$, and construct the covariance matrix C of the point set:

$$x{x}_i=\sum _{n=1}^N{\left({\overline{x}}_n\right)}^2,y{y}_i=\sum _{n=1}^N{\left({\overline{y}}_n\right)}^2,x{y}_i=\sum _{n=1}^N{\overline{x}}_n{\overline{y}}_n$$

(9)

$$C={\displaystyle \frac{1}{N-1}}\left(\begin{array}{cc}\sum {\left({\overline{x}}_n\right)}^2& \sum {\overline{x}}_n{\overline{y}}_n\\ \sum {\overline{x}}_n{\overline{y}}_n& \sum {\left({\overline{y}}_n\right)}^2\end{array}\right)={\displaystyle \frac{1}{N-1}}\left(\begin{array}{cc}x{x}_i& x{y}_i\\ x{y}_i& y{y}_i\end{array}\right)$$

(10)

These values are then used to construct an initial metric, ${\tilde{Q}}_{\mathrm{line}}$, which is highly sensitive to collinearity. Its calculation formula is as follows:

$${\tilde{Q}}_{\mathrm{line}}={\displaystyle \frac{x{x}_i+y{y}_i}{x{x}_i·y{y}_i-{\left(x{y}_i\right)}^2+ϵ}}$$

(11)

where ${\tilde{Q}}_{\mathrm{line}}$ [35] represents an initial, unnormalized collinearity score whose value increases sharply as the point set approaches collinearity, since its denominator is related to the determinant of the covariance matrix which approaches zero. The terms $x{x}_i,y{y}_i,x{y}_i$ are the second-order central moments of the centralized point set, respectively, used to describe its degree of dispersion. Finally, $ϵ$ (epsilon) is a very small positive number (e.g., $1\times {10}^{-8}$) used to prevent division by zero and ensure the numerical stability of the computation.

To facilitate network optimization, we use an exponential function to smoothly map the initial score into the interval $[0,1)$, yielding the final collinearity quality metric, $Q_{\mathrm{line}}$:

$$Q_{\mathrm{line}}=1-e^{(-\alpha ·{\tilde{Q}}_{\mathrm{line}})}$$

(12)

where $\alpha$ is a tunable positive hyperparameter that controls the steepness of the normalization function.

However, collinearity alone is insufficient to describe a stripe. A point-like noise artifact or a compact cluster of points could also achieve a high collinearity score. Therefore, in the second step of this mechanism, we introduce an aspect ratio metric to explicitly encourage elongated morphologies. This metric is inspired by the core idea of Principal Component Analysis (PCA).

We construct the $2\times 2$ covariance matrix $C_i$ from the centralized point set ${\overline{P}}_i$ and solve for its two eigenvalues, ${\lambda }_1$ and ${\lambda }_2$ (ensuring that ${\lambda }_1\ge {\lambda }_2\ge 0$). Based on these, we define the aspect ratio quality metric, $Q_{\mathrm{aspect}}$:

$$Q_{\mathrm{aspect}}=1-\sqrt{{\displaystyle \frac{{\lambda }_2}{{\lambda }_1+ϵ}}}$$

(13)

where ${\lambda }_1$ and ${\lambda }_2$ are the largest and smallest eigenvalues of the covariance matrix, respectively. ${\lambda }_1$ precisely reflects the variance or degree of spread of the point set along its principal axis (i.e., the direction of the stripe’s length), while ${\lambda }_2$ reflects the variance along the minor axis (i.e., the direction of the stripe’s width). For an ideal stripe, ${\lambda }_1$ is significantly larger than ${\lambda }_2$. The term $ϵ$ (epsilon) is, likewise, a very small positive number included to ensure numerical stability.

Finally, we multiply the collinearity metric and the aspect ratio metric to fuse them into a more robust, comprehensive, and accurate joint quality metric, $Q_{\mathrm{line}}^{\prime }$:

$$Q_{\mathrm{line}}^{\prime }=Q_{\mathrm{line}}\times Q_{\mathrm{aspect}}$$

(14)

The enhanced metric, $Q_{line}^{\prime }$, provides a highly precise geometric prior for the network’s learning process. During training, only those point sets that are both highly collinear and have an elongated morphology can achieve a final high score approaching 1, thereby occupying a dominant role in the loss function calculation. This mechanism effectively compels the model to generate representative points that precisely delineate the complete morphology of RSOs, laying a solid theoretical foundation for achieving superior localization and regression precision in complex space environments.

3.4. Loss Function

The training process of SAANet is guided by a joint loss function, $\mathcal{L}$, which is designed to simultaneously optimize target classification, localization, and the spatial distribution of the representative point set. The total loss function is composed of a weighted sum of three components: the classification loss, ${\mathcal{L}}_{\mathrm{cls}}$ [37]; the localization loss, ${\mathcal{L}}_{\mathrm{loc}}$ [38]; and the spatial constraint loss, ${\mathcal{L}}_{\mathrm{line}}$ [35]:

$$\mathcal{L}={\lambda }_1{\mathcal{L}}_{\mathrm{cls}}+{\lambda }_2{\mathcal{L}}_{\mathrm{loc}}+{\lambda }_3{\mathcal{L}}_{\mathrm{line}}$$

(15)

The weighting hyperparameters ${\lambda }_1,{\lambda }_2$, and ${\lambda }_3$ were determined through an empirical grid search on the validation set to balance the gradient magnitudes of different loss terms. Based on the experimental results, we set ${\lambda }_1=1.0, {\lambda }_2=1.0, \mathrm{and}{\lambda }_3=0.1$ to ensure stable model convergence.

3.4.1. Classification Loss

The classification loss term is used to distinguish between targets (stripes) and the background. Due to the severe class imbalance problem, where background pixels far outnumber target pixels in the images, we adopt the Focal Loss. The Focal Loss reduces the weight of well-classified, easy samples, forcing the model to focus more on learning hard-to-distinguish samples during training.

$${\mathcal{L}}_{cls}={\displaystyle \frac{1}{N_{cls}}}\sum _i-\alpha {(1-p_t)}^{\gamma }log\left(p_t\right)$$

(16)

where $N_{cls}$ represents the total number of points in a batch, and $p_t$ is the model’s predicted probability for the ground-truth class. The parameter $\alpha$ is a balancing factor used to adjust the contribution weight of positive versus negative samples. The parameter $\gamma$ is a modulating factor; when $\gamma >0$, it reduces the contribution of well-classified samples with high prediction probabilities ($p_t$) to the total loss.

3.4.2. Localization Loss

The localization loss term measures the geometric discrepancy between the predicted bounding box, which is converted from the representative point set, and the ground-truth OBB. We employ a loss function based on the Generalized Intersection over Union (GIOU) because it not only considers the overlapping area but also accounts for the non-overlapping regions, making it more effective at handling cases where the boxes do not overlap.

The calculation of GIOU relies on the standard Intersection over Union (IoU), which is defined as follows:

$$\mathrm{IoU}={\displaystyle \frac{\left|A\cap B\right|}{\left|A\cup B\right|}}\mathrm{GIOU}=\mathrm{IoU}-{\displaystyle \frac{\left|C\backslash (A\cup B)\right|}{\left|C\right|}}$$

(17)

where A and B represent the predicted bounding box and the ground-truth bounding box, respectively, and C is the smallest convex set that encloses both A and B.

The localization loss, ${\mathcal{L}}_{loc}$, is thus given by:

$${\mathcal{L}}_{loc}={\displaystyle \frac{1}{N_{loc}}}\sum _i(1-{\mathrm{GIOU}}_i)$$

(18)

where $N_{loc}$ represents the number of point sets in a batch that are assigned as positive samples.

3.4.3. Spatial Constraint Loss

The spatial constraint loss term is a key regularization term that ensures the adaptively learned representative points fall within the target region instead of being scattered outside of it. It enforces this constraint by penalizing any representative points that lie outside the ground-truth bounding box. The penalty function, ${\rho }_{ik}$, is defined as follows:

$${\rho }_{in}=\left\{\begin{array}{cc}d(p_{in},p_c),\hfill & \mathrm{if}\mathrm{point}{p}_{in}\mathrm{is}\mathrm{outside}\mathrm{the}\mathrm{ground-truth}\mathrm{bounding}\mathrm{box}\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(19)

where $d(p_{in},p_c)$ represents the distance between the representative point $p_{in}$ and the geometric center of the ground-truth bounding box, $p_c$.

With the penalty function ${\rho }_{in}$ defined, we can now formulate the full spatial constraint loss, ${\mathcal{L}}_{\mathrm{line}}$:

$${\mathcal{L}}_{\mathrm{line}}={\displaystyle \frac{1}{N_{\mathrm{pos}}}}\sum _{i\in \mathrm{pos}}{\displaystyle \frac{1}{N}}\sum _{n=1}^N{\rho }_{in}$$

(20)

where $N_{\mathrm{pos}}$ represents the total number of positive samples in a batch.

4. Experiment

This chapter aims to comprehensively evaluate the performance of our proposed SAANet on the task of detecting stripe-like RSOs. We begin by detailing the experimental setup, including the datasets, evaluation metrics, and implementation details. Subsequently, we demonstrate the superiority of SAANet through quantitative and qualitative comparisons with several state-of-the-art infrared target detection models and the traditional ResNet50. Finally, we conduct a series of thorough ablation studies to validate the effectiveness of each innovative component within our model.

4.1. Experimental Setup

4.1.1. Dataset

To address the challenges posed by the complex spatial situations involving stripe-like RSOs, we conducted extensive experiments on two datasets, each presenting distinct characteristics and challenges.

AstroStripeSet [28]: This dataset was specifically constructed for the task of stripe-like space object detection. It comprises images affected by four typical types of space stray light: Earthlight, Sunlight, Moonlight, and mixed light. The complete dataset contains 1500 images of size 256 × 256, divided into a fixed set of 1000 training images, 100 validation images, and 400 test images. The dataset simulates varying exposure durations through different streak lengths. This variation in signal intensity, combined with complex stray light backgrounds, allows for a comprehensive verification of the model’s robustness under diverse exposure and noise interference conditions.

StarTrails Dataset [8]: This dataset was built for space object detection in a scanning imaging mode, generated by simulating the motion and imaging processes of stars and space objects combined with time delay integration technology. The entire dataset consists of 2000 images of size 1024 × 1024, containing over 70,000 valid RSO annotations. To simulate specific exposure levels, the dataset is explicitly parameterized with 32, 64, 96, and 128 integration stages [8], resulting in targets with inconsistent signal intensities and streak lengths. This configuration, combined with a high density of overlapping and intersecting trajectories, serves as a rigorous benchmark for assessing a detection model’s ability to precisely discriminate and localize dense, arbitrarily oriented targets in complex environments.

The AstroStripeSet and StarTrails Dataset are derived from benchmarks established in [8,28]. We strictly adhered to their original data partitioning and annotation protocols for fair comparison. The datasets are available at https://github.com/Lyy-7/SAANet (accessed on 5 January 2026).

4.1.2. Evaluation Metrics

To quantitatively evaluate the model’s performance, we adopted widely recognized evaluation metrics from the field of object detection, including Recall and Average Precision (AP).

First, we introduce Recall, which is a metric that measures the model’s ability to find all existing targets. It is defined as the ratio of the number of correctly detected positive samples to the total number of actual positive samples:

$$\mathrm{Recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(21)

where TP is the number of correctly detected targets, and FN is the number of actual targets that the model failed to detect.

Additionally, we introduce the AP metric to provide a comprehensive evaluation of the model’s Precision and Recall for a single class. It measures the model’s overall performance by calculating the weighted average of precision at various recall levels.

4.1.3. Implementation Details

Our proposed SAANet model and all comparative experiments were implemented based on the MMRotate open-source toolbox and the PyTorch (version 1.11.0) deep learning framework. To ensure a consistent experimental environment, all models were trained and tested on a single server equipped with an NVIDIA RTX 4090 GPU.

We employed the Stochastic Gradient Descent (SGD) optimizer with an initial learning rate set to 0.008. To accommodate the characteristics of the different datasets, input images were uniformly resized to a fixed dimension. For the AstroStripeSet, the input size was $256\times 256$ pixels with a batch size of 16. For the StarTrails dataset, owing to its higher original resolution and sparser targets, the input size was set to $1024\times 1024$ pixels with a batch size of 2. Furthermore, to enhance the model’s generalization ability, we applied data augmentation techniques during the training phase, primarily consisting of random horizontal, vertical, and diagonal flips. We intentionally excluded geometric rotations to prevent interpolation artifacts from degrading the structural integrity of faint, pixel-level stripe targets. Furthermore, since the datasets already contain complex, real-world stray light interference, we avoided adding synthetic noise to preserve the realistic noise distribution characteristics. All models were trained for 100 epochs.

4.2. Comparative Experiments

To comprehensively evaluate the performance of SAANet, we conducted extensive comparative experiments against a diverse set of baselines. In the category of traditional methods, we selected representative algorithms including TopHat [39], MaxMedian [40], LCM [41], the Frangi Filter [42], and the recent tensor-based DDCF [43]. For deep learning comparisons, we employed the classic ResNet50 [44] as a baseline, alongside several state-of-the-art networks optimized for dim object detection, such as ISNet [36], MSHNet [45], ISTDUNet [46], LW-IRSTNet [47], and SARATRNet [48]. This evaluation encompasses quantitative, qualitative, and ablation analyses on both the AstroStripeSet and StarTrails datasets to verify the effectiveness and robustness of the proposed framework.

4.2.1. Analysis on AstroStripeSet

As presented in

Table 1. Performance comparison on the AstroStripeSet. The notation “Red” indicates the best performance in each category (closest to Gts for Dets).

Type	Network	Earth Light				Sun Light				Moon Light				Mix Light				Total
		Gts	Dets	Rec	AP	Gts	Dets	Rec	AP	Gts	Dets	Rec	AP	Gts	Dets	Rec	AP	Gts	Dets	Rec	AP
Traditional	TopHat [39]	100	38	0.330	0.330	100	24	0.200	0.254	100	41	0.360	0.349	100	36	0.310	0.351	400	139	0.303	0.321
	MaxMedian [40]	100	31	0.280	0.255	100	22	0.200	0.258	100	40	0.340	0.341	100	33	0.240	0.229	400	126	0.265	0.271
	Frangi [42]	100	117	0.650	0.242	100	157	0.630	0.266	100	275	0.550	0.205	100	225	0.510	0.143	400	774	0.585	0.214
	LCM [41]	100	114	0.330	0.197	100	20	0.100	0.106	100	61	0.140	0.112	100	112	0.200	0.078	400	307	0.193	0.123
	DDCF [43]	100	106	0.600	0.400	100	98	0.620	0.450	100	93	0.710	0.583	100	122	0.520	0.261	400	419	0.613	0.478
Deep Learning	Resnet50 [44]	100	115	0.830	0.764	100	112	0.880	0.783	100	109	0.870	0.779	100	114	0.930	0.778	400	450	0.877	0.778
	ISNet [36]	100	106	0.830	0.753	100	113	0.910	0.828	100	108	0.880	0.797	100	108	0.860	0.641	400	435	0.877	0.754
	MSHNet [45]	100	108	0.740	0.642	100	109	0.780	0.665	100	105	0.740	0.687	100	88	0.570	0.483	400	410	0.707	0.646
	ISTDUNet [46]	100	119	0.870	0.767	100	115	0.920	0.821	100	112	0.930	0.865	100	124	0.870	0.720	400	470	0.897	0.771
	LW-IRSTNet [47]	100	131	0.690	0.551	100	113	0.720	0.678	100	118	0.730	0.664	100	130	0.530	0.464	400	492	0.668	0.566
	SARATRNet [48]	100	121	0.750	0.682	100	131	0.730	0.673	100	122	0.800	0.779	100	128	0.670	0.597	400	502	0.738	0.681
	SAANet (Ours)	100	102	0.920	0.881	100	101	0.910	0.856	100	99	0.960	0.893	100	104	0.930	0.835	400	406	0.930	0.864

, SAANet achieves optimal performance across most evaluation metrics. The quantitative results clearly demonstrate the limitations of traditional approaches for this task. Relying on hand-crafted features, these methods struggle in complex stray light environments. As shown in the table, methods like LCM and Frangi yield limited performance. Even the latest tensor-based method DDCF, while outperforming classical filters, only attains a Recall of 0.613 and an AP of 0.478. In contrast, SAANet achieves a Recall of 0.930 and an AP of 0.864, representing a substantial performance leap over traditional baselines. This validates that data-driven deep networks are essential for distinguishing dim targets from complex backgrounds.

Among deep learning approaches, SAANet consistently outperforms both the classic ResNet50 baseline and representative SOTA networks. While ResNet50 and ISTDUNet show competitive performance with Recalls of 0.877 and 0.897, they lack specific mechanisms for orientation regression, limiting their AP to around 0.77. Notably, compared to SOTA networks like MSHNet and SARATRNet which are optimized for infrared small targets, SAANet achieves a significant lead of 7.6% and 17.9% in AP respectively. This superiority stems from the Shape-Aware Paradigm which explicitly models the linear topology that generic small target detectors fail to capture.

Regarding detection quantity, SAANet generates 406 detection boxes, a figure remarkably close to the ground truth count of 400. This high level of consistency demonstrates the exceptional precision of the model and its capability to effectively suppress false alarms. In comparison, networks such as SARATRNet and ISTDUNet generate 502 and 470 candidates respectively, indicating a higher incidence of false positives in noisy backgrounds. Under detailed illumination conditions, SAANet exhibits exceptional robustness. It achieves the highest performance in Moon Light and Earth Light scenarios with Recalls of 0.960 and 0.920 respectively, and maintains the top AP of 0.856 even under intense Sun Light interference.

Figure 8 presents a qualitative comparison of the proposed SAANet against several competing networks. The visual results demonstrate the superior detection capability and robustness of our method.

The performance gap is most evident under complex scenarios such as cluttered Mix Light and low-illumination Moon Light. When strong interfering light causes a sharp increase in background interference, networks like MSHNet and LW-IRSTNet exhibit significant missed detections. Although ISNet and ISTDUNet show relatively high precision, they generate incomplete bounding boxes, failing to effectively preserve target integrity. In stark contrast, SAANet demonstrates exceptional robustness, precisely detecting dim targets missed by other methods while maintaining their structural continuity.

Even under relatively clearer conditions like Earth Light and Sun Light, competing networks still exhibit limitations. Specifically, in scenarios dominated by intense Sun Light, the strong background radiance creates severe interference. Comparative methods, such as MSHNet and SARATRNet, struggle to distinguish between this high-intensity stray light noise and valid targets, resulting in noticeable false alarms. While SAANet maintains robustness, the misclassification by other models highlights the severity of strong stray light influence in these environments. Meanwhile, ResNet50 and ISNet suffer from missed detections due to weak feature extraction. Furthermore, even when competing networks successfully detect targets, their detection precision is inferior to that of SAANet.

Collectively, the visual results across all conditions validate the advanced capabilities of the Shape-Aware Paradigm in handling dim and obscured targets and confirm its strong generalization performance in diverse space imaging environments.

4.2.2. Analysis on StarTrails Dataset

As detailed in

Table 2. Performance comparison on the StarTrails dataset. The notation “Red” indicates the best performance in each category.

Type	Network	Gts	Dets	Recall	AP
Traditional	Top-Hat [39]	7774	1376	0.317	0.337
	Max-Median [40]	7774	934	0.210	0.252
	Frangi [42]	7774	1934	0.407	0.391
	LCM [41]	7774	857	0.202	0.258
	DDCF [43]	7774	2947	0.333	0.332
Deep Learning	Resnet50 [44]	7774	7215	0.794	0.711
	ISNet [36]	7774	8012	0.821	0.752
	MSHNet [45]	7774	8080	0.817	0.739
	ISTDUNet [46]	7774	9150	0.838	0.805
	LW-IRSTNet [47]	7774	6956	0.847	0.773
	SARATRNet [48]	7774	7084	0.801	0.730
	SAANet (Ours)	7774	7965	0.850	0.815

, SAANet demonstrates outstanding performance on the StarTrails dataset. Traditional methods struggle severely in this scenario characterized by dense and overlapping trajectories. Methods like LCM and Max-Median achieve an Average Precision of less than 0.26. Even the latest method DDCF [43] and the tubular-specific Frangi Filter fail to yield satisfactory results, achieving APs of only 0.332 and 0.391, respectively. This highlights the inability of hand-crafted features to disentangle complex geometric intersections in high-density environments.

In the deep learning category, SAANet leads all competing networks with a Recall of 0.850 and an AP of 0.815. It is worth noting that while recent SOTA networks like MSHNet and SARATRNet perform well on small targets, they show limited effectiveness on this dataset with APs of 0.739 and 0.730 respectively. This suggests that architectures optimized for point-like targets may fail to capture the orientation features required for dense stripes. SAANet surpasses the runner-up ISTDUNet by 1.2% in AP, validating the effectiveness of the proposed joint geometric constraint mechanism in resolving overlapping ambiguities.

An analysis of detection counts further underscores the superior quality of candidates generated by SAANet. As detailed in the table, SAANet yields 7965 detections, a figure that most closely aligns with the ground truth count. In contrast, LW-IRSTNet and SARATRNet produce notably fewer detections, specifically 6956 and 7084. These reduced counts correlate directly with their lower recall rates, indicating that these models fail to detect a significant portion of valid targets. Conversely, ISTDUNet generates an excessive 9150 candidates, which implies a propensity for false alarms within cluttered regions. By maintaining a detection count that is highly consistent with the ground truth, SAANet achieves an optimal trade-off between precision and recall, demonstrating its robustness in accurately localizing valid targets while effectively suppressing background noise in dense environments.

The macroscopic qualitative results in Figure 9 corroborate the quantitative findings. Distinct from the multi-color error analysis in Figure 8, we adopt a prediction-only protocol for the StarTrails dataset to directly assess regression consistency. The visualizations reveal that classic networks like ResNet50 and ISNet, along with recent methods such as MSHNet and SARATRNet, struggle with the dataset’s complexity, yielding numerous missed detections and false alarms. Although ISTDUNet achieves a relatively high recall rate, its performance is compromised by a large number of cluttered and fragmented detection boxes. In contrast, SAANet demonstrates a clear advantage, striking an optimal balance between a high detection rate and clean, precise results. It is important to note that due to the extremely high target density in this dataset, explicitly marking all detection errors in the global view would result in severe visual occlusion. Therefore, we focus on the macroscopic distribution here. While SAANet exhibits superior global coverage, it still encounters challenges in regions with explosive trajectory density, where signal saturation can lead to occasional missed detections.

To provide a more in-depth analysis, Figure 10 presents a magnified view of several typical challenging scenarios. In the case of overlapping trajectories (a), competing networks often produce fragmented detections or fail to disentangle the intersection. SAANet, however, maintains the structural integrity of each target. Similarly, for densely packed RSOs (b), other models exhibit low precision or incorrectly merge adjacent targets, while SAANet accurately separates each object, demonstrating a superior understanding of spatial relationships. However, a specific limitation persists in these ultra-dense intersections. When multiple trajectories intersect at narrow angles, the significant spatial overlap may trigger the Non-Maximum Suppression mechanism to incorrectly suppress valid adjacent bounding boxes, resulting in minor discontinuities.

Finally, in the high-noise environment (c), while networks like SARATRNet generate false alarms and ResNet50 misses dim targets, our model achieves an optimal balance, detecting both strong and weak-signal targets while effectively suppressing noise. Collectively, these extensive qualitative comparisons validate the exceptional performance and robustness of the SAANet when handling dense, overlapping, and interference targets.

4.2.3. Ablation Study

In this section, we conduct a series of comprehensive ablation studies to systematically evaluate the effectiveness of the core modules within our proposed framework. To quantify the contribution of each component to the final performance, we trained different model configurations on the AstroStripeSet and StarTrails datasets. The detailed experimental results are summarized in

Table 3. Comprehensive ablation study of each component in SAANet.

Network	Module				Average Metrics
	SA-FPN	CSAM	AL-OBB		AstroStripe		StarTrails
			${\mathit{Q}}_{\mathit{line}}$	${\mathit{Q}}_{\mathit{aspect}}$	Recall	AP	Recall	AP
Baseline	-	-	-	-	0.785	0.650	0.827	0.709
	✓	✓	-	-	0.875	0.735	0.821	0.776
	✓	✓	-	✓	0.890	0.742	0.830	0.784
	✓	✓	✓	-	0.910	0.824	0.832	0.790
	✓	-	✓	✓	0.913	0.830	0.842	0.804
	-	✓	✓	✓	0.920	0.836	0.839	0.803
	✓	✓	✓	✓	0.930	0.864	0.850	0.815

.

Our analysis begins with the baseline model. Equipping the baseline model with the SA-FPN and CSAM modules significantly enhances performance. On the AstroStripeSet, it achieves substantial gains of 11.5% in Recall and 13.1% in AP, while on the StarTrails dataset, the AP is also boosted by approximately 9.5%. This series of improvements fully validates the designated role of SA-FPN as a powerful decoder and CSAM as a feature enhancer, proving their exceptional effectiveness in capturing the dim image features of stripe-like RSOs.

Furthermore, to verify the effectiveness of the proposed joint geometric constraints, we analyze the internal mechanisms of the AL-OBB head. Here, $Q_{line}$ and $Q_{aspect}$ denote the collinearity and aspect ratio constraints in the AL-OBB head, respectively. As shown in Table 3, applying either $Q_{line}$ or $Q_{aspect}$ individually yields performance improvements compared to the configuration without geometric constraints. However, the consistent and significant gains are maximized when both constraints are applied jointly. This highlights the core value of the AL-OBB module: it successfully imposes effective constraints on the linear features of the targets, thereby generating more precise detection boxes that are better suited to the unique geometry of stripe-like objects.

Finally, when all modules including SA-FPN, CSAM, and the complete AL-OBB are integrated simultaneously, our complete model achieves optimal performance across all metrics on both datasets. On AstroStripeSet, it reaches a Recall of 0.930 and an AP of 0.864, while on StarTrails, it achieves a Recall of 0.850 and an AP of 0.815. This result profoundly validates the effectiveness of our proposed Unified Shape-Aware Paradigm: the explicit modeling of linear topologies by SA-FPN and the joint geometric constraint mechanism of AL-OBB are tightly coupled to successfully resolve the core challenges of complex stray light interference and dense overlapping trajectories, thereby establishing the SOTA performance of SAANet in space stripe-like object detection tasks.

4.2.4. Model Efficiency Analysis

In practical space-based applications, the computational resources and storage capacity of satellite platforms are extremely limited. Therefore, in addition to detection precision, we evaluated the model’s parameters (Params), floating-point operations (FLOPs), and inference speed (FPS). The experiments were conducted on an NVIDIA RTX 4090 GPU with a unified input size of $256\times 256$.

The quantitative results in

Table 4. Comparison of model complexity and inference speed. The arrows ↓ and ↑ indicate that lower and higher values are better, respectively.

Network	Params (M) ↓	FLOPs (G) ↓	FPS ↑
ResNet50 [44]	36.60	22.14	27.3
ISNet [36]	8.23	26.09	26.9
MSHNet [45]	11.25	107.89	15.8
LW-IRSTNet [47]	7.31	103.06	17.2
ISTDUNet [46]	8.04	415.43	9.9
SARATRNet [48]	75.06	45.73	37.5
SAANet (Ours)	11.22	105.92	16.1

highlight the significant advantages of SAANet in terms of model lightweighting. Its parameter count is only 11.22 M, which is approximately one-third of the standard ResNet50. More critically, compared to the Transformer-based SARATRNet, SAANet’s parameter count is only about 15% of the former, drastically reducing the demand for on-orbit storage space. Regarding computational complexity, although SAANet introduces attention mechanisms to effectively capture faint stripe features, resulting in a higher computational load than some other SOTA networks, its 105.92 G FLOPs remains far below the computationally intensive ISTDUNet and is on par with MSHNet and LW-IRSTNet.

Considering the precision performance in Table 1 and Table 2, SAANet achieves optimal detection performance with a lower storage cost and reasonable computational cost. This balance between precision and efficiency makes it a highly competitive solution for space situational awareness tasks.

A comprehensive analysis synthesizing the efficiency metrics in Table 4 with the detection precision in Table 1 and Table 2 reveals that SAANet achieves the highest Average Precision of 0.864 on the AstroStripeSet and 0.815 on the StarTrails dataset, while maintaining a moderate footprint of 11.22 M parameters and 105.92 G FLOPs. In contrast, although SARATRNet exhibits a higher inference speed, its extensive parameter count of 75.06 M yields substantially lower accuracies of 0.681 and 0.730. Similarly, LW-IRSTNet, despite having the smallest parameter size, attains APs of only 0.566 and 0.773, indicating limited generalization capability. This comparison demonstrates that SAANet utilizes parameters more efficiently to capture stripe-like features. Consequently, the optimal balance between precision and efficiency positions SAANet as a highly competitive solution for space situational awareness tasks.

5. Conclusions

In this work, we propose the Shape-Aware Attention Network (SAANet) to address the challenges of detecting stripe-like Resident Space Objects (RSOs). To tackle the limitations of standard detectors in complex stray light and dense environments, we establish a unified shape-aware paradigm with three core contributions. First, we implement streamlined feature maintenance by restructuring the encoder into a specialized feed-forward backbone, ensuring the preservation of dim signals. Second, we construct a shape-aware orthogonal fusion mechanism within the Shape-Aware Feature Pyramid Network (SA-FPN), explicitly modeling linear topologies to enhance feature discriminability. Third, we propose a joint geometric constraint mechanism in the Adaptive Linear Oriented Bounding Box (AL-OBB) head. By enforcing strict collinearity and aspect ratio alignment, this strategy theoretically resolves regression ambiguities in overlapping scenarios.

Extensive experiments on the AstroStripeSet and StarTrails datasets demonstrate that SAANet achieves state-of-the-art performance. Specifically, on the AstroStripeSet, it achieves a Recall of 0.930 and an Average Precision (AP) of 0.864, while on the StarTrails dataset, it attains a Recall of 0.850 and an AP of 0.815. Qualitative analysis further confirms its robustness in maintaining target integrity and suppressing noise. This work provides an effective theoretical and practical solution for Space Situational Awareness (SSA), offering new avenues for the detection of morphologically specific targets.

Although the datasets used in this study rigorously simulate space environments, we acknowledge the necessity of validation on real-world observational imagery. Due to the extreme scarcity of public, accurately annotated real-world datasets for stripe-like RSOs, we will dedicate future work to acquiring such data to further verify the generalization capability of the proposed model.