AFDR-Det: Adaptive Feature Dual-Refinement Oriented Detector for Remote Sensing Object Detection

Arbitrary-oriented object detection is a significant and challenging remote sensing image detection task. However, due to two types of feature misalignment in one-stage oriented detectors, their performance has yet to meet expectations. In remote sensing images, objects are distributed in arbitrary orientations, while convolutional features are inherently axis-aligned. This inevitably leads to spatial misalignment between the heuristically defined anchors and the convolutional features. Additionally, most head networks use two branch networks to implement object localization and classification tasks separately, resulting in a lack of information interaction between the classification and localization tasks, leading to spatial feature misalignment issues. In this paper, we propose an Adaptive Feature Dual-Refinement Detector (AFDR-Det) to address the above challenges. First, we propose a novel neck network that integrates gated fusion modules and transformer blocks into the feature pyramid architecture, thereby enhancing the neck’s ability to capture both local details and long-range contextual information. Secondly, we propose a Feature Dual Refinement Head, which consists of two sub-modules: a Modulated Anchor Refinement Module (MARM) and a Task-aware Feature Alignment Module (TFAM). MARM aims to generate high-quality anchors and adaptively align features with the corresponding anchors. TFAM further enforces feature consistency between the two tasks from the spatial dimension by enhancing information interaction between the classification and regression branches, thereby achieving feature alignment between the tasks of classification and localization. We evaluated the proposed AFDR-Det on three challenging public remote sensing image datasets: DOTA, DIOR-R, and HRSC2016. The experimental results demonstrate that our method achieved mean average precision (mAP) scores of 80.54%, 66.90%, and 98.51% on the DOTA, DIOR-R, and HRSC2016 datasets, respectively, indicating significant improvements over the baseline methods.