Jitter Correction Network - A Deep Multi-Scale Optical Flow Framework for Robust Satellite Image Restoration

Jitter-Induced distortions in satellite imagery can significantly degrade geometric accuracy and reduce the reliability of downstream applications such as object detection, classification, and change detection. This paper presents the Jitter Correction Network, a deep learning-based framework designed to correct jitter artifacts by combining optical flow estimation, multi-scale feature extraction, and adaptive feature modulation within a convolutional neural network. Instead of relying on frequency-domain techniques or fixed motion models, the network learns motion patterns directly from image data using optical flow-guided feature extraction. Its multi-scale architecture captures both fine and coarse jitter effects, while adaptive modulation enhances focus on distortion-affected regions. To improve robustness against noise and outliers, the model utilizes the Charbonnier loss and is optimized using AdamW for stable training and enhanced generalization. Results show that the Jitter Correction Network delivers consistent improvements in correction accuracy and image quality, providing a reliable solution for enhancing satellite image fidelity.