Physically Constrained Deep Prior for Multi-Type Stripe Noise Removal in Linear Array Whisk-Broom Infrared Imaging Systems

Under the dimensional constraints of satellite platforms, employing linear array detectors with whisk-broom imaging offers a practical approach for infrared remote sensing systems to achieve wide-swath coverage and high resolution. However, stripe noise within linear array whisk-broom systems typically manifests as mixed types and frequencies, with stripe directions corresponding to both the satellite flight direction and the whisk-broom scanning direction. This paper proposes an optimized destriping method based on physics-informed deep priors to decompose stripe noise with multiple types and overlapping frequencies effectively. By leveraging deep-space calibration data acquired during satellite side-swing maneuvers, we isolate the noise frequency components aligned with the flight direction. A wavelet decomposition-based neural network model is employed to predict flight-direction noise, serving as a prior for the optimization method. To address multiple noise types, different physical characteristics of the noise are leveraged as constraints, enabling effective noise decomposition while simultaneously preserving background information. We conducted comparative evaluations against mainstream methods using a large dataset comprising 5,200 12-bit images. Performance was assessed using no-reference quality metrics and the Modulation Transfer Function. The results demonstrate the superior performance of the proposed method. This study provides a novel solution for noise decomposition in infrared imaging systems.