Infrared dim target detection via spatial-temporal contrast and bidirectional pyramid network

In deep space environments, extremely low target-background contrast, complex starry backgrounds, and non-uniform background patches often result in the target being overwhelmed by the complex background. To address the challenge of infrared dim target recognition being significantly impacted by complex backgrounds in a deep space environment, this paper proposes a bidirectional pyramid dim target detection algorithm based on spatial-temporal local contrast. A background suppression model capable of adapting to interference and variations in background brightness is introduced alongside a robust infrared dim target detection model. By thoroughly considering the characteristics of both the target and the background, as well as their gray value relationships, the model utilizes an inner and outer window mechanism to mitigate the effects of background discontinuities and large fluctuations, and constructs a Spatial-Temporal Local Contrast model that fully leverages spatial gray-level contrast and temporal gray-level variations. The inner window employs a bidirectional orientation contrast mechanism to effectively suppress missed detections caused by low-contrast targets being overwhelmed by background clutter. Additionally, the model exploits variations in the ratio between the inner and outer windows across multiple frames, combined with edge-based comparisons of the four sides of the outer window, to assess background changes and suppress thermal edge effects. Subsequently, the model incorporates target motion information to predict the likely region of target presence, enabling accurate localization even when the target's gray value is nearly identical to that of the background in certain frames. Furthermore, the detection network integrates a guided attention mechanism that leverages the spatial cues derived from the contrast module to focus on potential target regions. Accurate spatial dim target detection is then achieved through the proposed bidirectional pyramid framework. Compared with traditional algorithms, the proposed model not only exhibits strong anti-interference capability under extremely low targetto-background contrast and highly complex background conditions, but also enhances detection accuracy by integrating the potential target regions with an attention mechanism into a bidirectional pyramid detection framework, enabling more precise identification of infrared dim targets. Background Filtering Score (BFS), Signal to Clutter Ratio (SCR), and Signal to Clutter Ratio Gain (SCRG) are adopted as evaluation metrics in the background suppression experiment, and the experimental results show that the proposed method achieves a BFS of 93.89 and improves the SCR from 3.00 to 27.23. And the target detection experiments are evaluated using Precision, Recall, and F1 -score, among which the precision achieves a notable value of 94.7%, demonstrating its effectiveness in enhancing background suppression and improving detection performance.