The inherent non-uniformity of Infrared Focal Plane Arrays (IRFPA) inevitably results in stripe noise, which severely degrades image quality and hinders downstream applications. Existing deep learning methods often struggle to strike a balance between effective denoising and the preservation of fine thermal textures. To address this issue, we propose a Bottleneck-free Channel Dual-path Aggregation Network (BCDA-Net) based on a “Perception-Reconstruction” design principle. In the perception stage, the network jointly employs the Dual-Path Channel Down-sampling (DCD) module and the Context-Guided Stripe Attention Block (CGSAB). The DCD module utilizes a channel split strategy to simultaneously extract semantic features and preserve high-frequency textures, while the CGSAB performs global context modeling on these features to precisely perceive and locate global stripe noise patterns. In the reconstruction stage, we integrate the Cascaded Dense Feature Aggregation (CDFA) module with a Bottleneck-Free Aggregation Strategy (BFAS). The CDFA utilizes the perceived information to densely aggregate features and progressively reconstruct clean image details, whereas the BFAS structurally blocks the propagation of low-resolution noise during decoding, effectively mitigating aliasing artifacts induced by deep feature upsampling. Together, these components form a complete closed loop from accurate noise perception to high-fidelity reconstruction. Extensive experiments on public and real-world datasets demonstrate that BCDA-Net maximally preserves image details while removing non-uniform stripe noise. Both objective metrics and subjective visual quality outperform existing state-of-the-art methods.

To ensure a high signal-to-noise ratio and high image volume, a geostationary orbiting ocean remote-sensing system needs to maintain high platform stability over a long integration time because it is affected by satellite attitude changes. When the observation target is the ocean, it is difficult to extract image features because of the lack of characteristic objects in the target area. In this paper, we attempt to avoid using image data for satellite attitude and image pixel offset estimation. We obtain the satellite attitude by using equipment such as gyroscopes and performing time registration between the satellite attitude and the image data to achieve pixel offset matching between images. According to the law of satellite attitude change, we designed a Kalman-like filter fitting (KLFF) algorithm based on the satellite attitude change model and the Nelder–Mead search principle. The discrete attitude data were time-marked by a synchronization system, and high-precision estimation of the satellite attitude was achieved after fitting with the KLFF algorithm. When the measurement accuracy of the equipment was 1.0 × 10⁻³°, the average prediction error of the algorithm was 1.09 × 10⁻³°, 21.58% better than the traditional interpolation prediction result of 1.39 × 10⁻³°. The peak value of the fitting angle error reached 2.5 × 10⁻³°. Compared with the interpolation prediction result of 6.2 × 10⁻³°, the estimated stability of the satellite attitude improved by about 59.68%. After using the linear interpolation method to compensate for the estimated pixel offset, its discrete range was 0.697 pixels. Compared with the 1.476 pixels of the interpolation algorithm, it was 52.8% lower, which improved the noise immunity of the algorithm. Finally, a KLFF algorithm was designed based on the satellite attitude change model by using external measurement data and the synchronous clock as a benchmark. The instantaneous attitude of the satellite was accurately estimated in real time, and the offset matching between the images was realized, which lays the foundation for in-orbit satellite data processing.

Transient-source detection without relying on difference images still faces challenges in achieving high accuracy, especially under practical space-based astronomical survey conditions where the data volume is enormous, on-orbit transmission bandwidth is limited, and real-time response is required for rapid follow-up observations. To address these issues, this paper proposes a lightweight detection network that integrates multi-scale feature fusion with contextual feature extraction, enabling efficient real-time processing on resource-constrained edge devices. The proposed model enhances robustness to point-spread-function variations across observation conditions and to complex background environments, while simultaneously improving detection accuracy. To evaluate performance comprehensively, lightweight VGG and lightweight ResNet architectures and other baseline models—commonly used as baselines for transient-source detection—are adopted for comparison. Experimental results show that under the condition that the models have approximately the same number of parameters, the proposed network achieves the best accuracy, obtaining nearly 1% improvement compared with the best-performing baseline model. Based on this design, an ultra-lightweight version with only 7k parameters is further developed by incorporating a compact multi-scale module, improving accuracy by 1% over the version without the multi-scale structure. Moreover, through heterogeneous knowledge distillation and adaptive iterative training, the accuracy of the ultra-lightweight model is further increased from 93.3% to 94.0%. Finally, the model is deployed and validated on an AI hardware acceleration platform. The results demonstrate that the proposed method substantially improves inference throughput while maintaining high accuracy, providing a practical solution for real-time, low-latency, on-device transient-source detection under large data volume and limited transmission conditions. Specifically, the proposed models are trained offline on a high-performance GPU and subsequently deployed on the Fudan Microelectronics 7100 AI board to evaluate their real-world inference efficiency on resource-constrained edge devices.