SRHF-Net: Super-Resolution-Assisted High-Frequency Learning for Remote Sensing Small Object Detection

Abstract

Small target detection in remote sensing images (RSIs) remains challenging due to sparse target representations and severe background clutter. To address these issues, this letter proposes a super-resolution-assisted high-frequency learning network (SRHF-Net), a dual-modal fusion framework for remote sensing small object detection. First, a sampling reconstruction framework is designed to alleviate weak feature representation caused by aggressive downsampling. After reconstruction, small object features retain high-frequency details and maintain representations rich in detail. Second, a dual-attention with large kernel (DALK) enhancement module is introduced in the intermediate layers to enhance detailed target features. Meanwhile, a multiscale large kernel attention SPPF (MSLA-SPPF) module is applied in the deep layers to expand the receptive field and integrate global context. Finally, a feature-level super-resolution (FL_SR) auxiliary branch is introduced during training to enhance high-frequency detail learning. The branch is removed during inference, and low-resolution (LR) inputs further reduce computational cost. Extensive experiments on the VEDAI dataset demonstrate that SRHF-Net achieves 80.47% detection accuracy with only 11.6-M parameters, outperforming YOLOv8s and YOLOv11s by 20.3% and 23.8%, respectively.