IBODet: Information-Bottleneck-Inspired Lightweight Oriented Detection in Remote Sensing Images

Abstract

Lightweight oriented object detection in high-resolution remote sensing imagery is challenging, since detectors must handle substantial variations in object scale and complex object geometries under strict computational constraints. Prevailing lightweight methods often focus on backbone compression, leaving the neck and head ill-equipped for multiscale fusion and geometry-aware feature extraction. Inspired by the information-bottleneck (IB) principle, this letter interprets lightweight detection as a tradeoff between task-relevant feature preservation and redundancy compression, and uses this principle as a design lens rather than a directly optimized objective. Accordingly, we propose IBODet, a lightweight detector that addresses this challenge at both the neck and head levels. First, our IB-FPN employs learnable weights for dynamic cross-level feature integration and a shared refinement module to enhance multiscale representations with minimal overhead, consistent with the IB tradeoff. Second, our context-interacted geometry-oriented head (CIG-head) models the interaction between global context and horizontal and vertical strip features to capture long-range dependencies in moderately elongated objects, after which a zero-initialized distribution-calibrated (DistCalib) block conditions the fused descriptor on branch-level statistics. Built upon a compact StripNet-T backbone, this unified design yields a model with only 6.72 M parameters and 87.21 GFLOPs. Extensive experiments on multiple challenging benchmarks demonstrate the superiority and robustness of our method. IBODet achieves an mAP of 79.34 on DOTA-v1.0% and 72.27% on DOTA-v1.5 and 69.20% on DIOR-R, outperforming state-of-the-art lightweight detectors in mAP with the best parameter–accuracy tradeoff among parameter-constrained methods.