FRLI-Net: Feature Reconstruction and Learning Interaction Network for Tiny Object Detection in Remote Sensing Images

In the field of computer vision, detecting tiny objects in remote sensing images has long been a challenging task. This difficulty primarily stems from the poor matching of tiny objects with specific feature scales, compounded by susceptibility to interference from complex surroundings. To this end, we propose a Feature Reconstruction and Learning Interaction Network (FRLI-Net) for overcoming the aforementioned difficulties in tiny object detection. Specifically, to mitigate information loss of tiny objects in the feature space, we introduce a Feature Reconstruction Supervision Branch (FRSB) used only during training, which utilizes supervised loss to guide the attention of tiny objects within the reconstructed feature regions. Then, to tackle false detections caused by blurred boundaries of tiny objects, we design a Saliency Learning Interaction Module (SLIM) that employs learnable weights and proposed sub-functions to facilitate feature interaction, thereby reducing interference from complex backgrounds in the foreground. Extensive experiments conducted on the remote sensing image datasets AI-TOD-v2 and TinyPerson demonstrate that our FRLI-Net significantly enhances the detection of tiny objects, outperforming other state-of-the-art detection methods.