INFF-Net: Rotation-Aware Remote Sensing Object Detection Based on Implicit Neural Feature Field

Remote sensing objects are characterized by arbitrary orientations, complex backgrounds, and high aspect ratios (ARs). Most existing methods rely on explicit discrete convolutions to construct feature networks, which suffer from limitations such as insufficient rotation-invariant representation, background noise introduced by uniform sampling of regions of interest (RoIs), and difficulty in handling objects with extreme ARs. To address these issues, this article proposes implicit neural feature fields (INFF-Net), a rotation-aware detection network based on INFF-Net. It comprises three core components: the implicit rotation feature field (IRFF) constructs a spatial-angular continuous representation and introduces a rotation consistency regularization loss to impose geometric constraints, thereby mitigating the pose sensitivity and angular boundary discontinuity associated with discrete convolution. The importance-guided sparse sampling modulation (ISSM) module adaptively modulates RoI features through sparse coarse sampling and importance-score-based refined sampling, effectively suppressing background noise. The adaptive shape-aware feature enhancement (ASFE) module dynamically adjusts the number of sampling points based on objects’ ARs and employs elliptical sampling to conform to the contours, improving the detection accuracy of elongated objects. The method achieves the mean average precision (mAP) scores of 77.87%, 66.67%, and 97.90% on the DOTA-v1.0, DIOR-R, and HRSC2016 datasets, respectively, obtaining significant improvements of 3.92%, 4.48%, and 1.40% over the baseline, demonstrating the effectiveness and generalization capability.