Leveraging Permuted Image Restoration for Improved Interpretation of Remote Sensing Images

In this study, we introduce a novel self-supervised learning adapter based on permutated image restoration (PIR) for effectively transferring pretrained weights from natural images to remote sensing object detection tasks. The adapter's unique methodology encompasses a three-phase process: segmenting and permuting image blocks, estimating permutation matrices for sequence reconstruction, and applying specialized loss functions for accurate block positioning. The use of our approach results in the maintenance of fidelity in both absolute and relative block positions as demonstrated by the evaluation of block similarities. The empirical results indicate significant performance enhancements for diverse datasets spanning optical and synthetic aperture radar data types, including high resolution ship collections 2016 (HRSC2016), Small Object Detection dAtasets - Aerial (SODA-A), and rotated ship detection dataset (RSDD) while effectively avoiding overfitting.