A Multi Scale Strip-wise ConvNet for Infrared Image Stripe Removal

Abstract. Unmanned Aerial Vehicles (UAVs) often perform all-day all-weather missions such as nighttime surveillance and search-and-rescue, where visible-light sensors fail, motivating the use of long-wave infrared (LWIR) cameras. However, IR images suffer from horizontal or vertical stripe artifacts caused by sensor instability, which degrade image quality. Conventional methods struggle with non-periodic stripes due to their reliance on fixed frequency priors and stationary pattern assumptions, while learning-based methods often neglect structural features because of insufficient inductive bias. To further exploit the anisotropic geometry of stripe artifacts, we design a series of strip-wise convolution kernels with multiple kernel lengths. To address this, we propose multi-scale strip-wise convolution kernels that extend along stripe orientations, providing an expanded receptive field in the principal direction while limiting orthogonal interference. Multiple kernel sizes enable the network to capture both local distortions and long-range stripe structures, and dense connectivity promotes cross-scale feature interaction for effective stripe separation. This architecture, which we call Densely Connected Multi-Scale Strip-Conv (DCMS), constitutes the core of our proposed method. Furthermore, we conduct experiments on infrared image datasets and carry out ablation experiments to validate the efficacy of our innovative method and its modules. Experimental results demonstrate that our method achieves superior performance compared to state-of-the-art approaches in both quantitative metrics and visual quality.