From Weak Textures to Dense Arrangements: Leveraging Prior Knowledge for Small-Object Detection in Remote Sensing Images

Detecting small objects in remote sensing images (RSIs) is a significant challenge due to their weak texture, scale variations, and dense spatial arrangements. Existing approaches often overlook the importance of prior contextual information and the aggregation of features in densely packed small objects, both of which are crucial for improving the performance of remote sensing small-object detection (RSSOD). In this work, we propose the prior-guided context fusion network (PGCFNet), which enhances small-object detection by decoupling scene contextual information through three novel components: the prior-guided context fusion module (PGCFM), the depthwise aggregator (DWA), and the prior-guided small-object detector (PGSOD). This architecture facilitates a deeper exploration of the relationships between small objects and their surrounding environment. Specifically, PGCFM improves feature representation by integrating multiscale features and applying prior-guided dynamic channel weighting, addressing the challenge of weak textures. Additionally, DWA refines feature aggregation using dilated convolutions and dynamic feature adjustment, enabling precise multiscale detection in environments with dense small objects. Furthermore, PGSOD leverages prior knowledge to reduce background interference, enhancing small-object detection across varying scales and orientations. Collectively, these modules work synergistically to advance small-object detection in RSIs, overcoming key challenges in complex environments. Extensive experiments on three public datasets demonstrate that the performance of the proposed method outperforms several state-of-the-art (SOTA) detectors, especially for tiny object detection. Specifically, PGCFNet achieves 86.0% mean average precision (mAP) on the DIOR dataset, 95.59% mAP on the NWPU VHR-10 dataset, and 58.5% mAP on the AI-TOD dataset. Additionally, we conducted generalization experiments for PGCFM, DWA, and PGSOD, demonstrating its effectiveness across different datasets and detection networks with varying model sizes.