MWAE-YOLO: Frequency–Spatial Collaborative Enhancement for Small-Object Detection in Remote Sensing Images
Abstract
Small-objectdetection in remote sensing images remains challenging due to insufficient feature representation, weak texture information, complex backgrounds, and high sensitivity to localization errors. To address these issues, this article proposes a frequency–spatial collaborative enhancement detector, multilevel wavelet attention enhancement you only look once (MWAE-YOLO), for remote sensing small-object detection. The proposed method introduces three key designs. First, a Cascaded Attention Wavelet Context Module (CAWCM) is developed as a local frequency-domain enhancement module. It performs cascaded single-level wavelet decomposition to strengthen low-frequency structural cues and high-frequency detail cues, thereby enhancing edge and texture representations of small objects. Second, a Wavelet Block Attention Module is designed as a shallow-stage multiscale feature fusion module built upon CAWCM. It progressively integrates wavelet-enhanced features and performs channel–spatial recalibration to improve the discrimination between small targets and complex backgrounds. Third, an adaptive scale-intersection over union loss is proposed to improve bounding-box regression by explicitly modeling scale discrepancies with an adaptive scale-gating strategy, which enhances localization accuracy for small objects. Extensive experiments are conducted on three public remote sensing datasets, AI-TOD, NWPU VHR-10, and RS-STOD. The experimental results show that MWAE-YOLO achieves 0.579, 0.957, and 0.772 mAP50 on AI-TOD, NWPU VHR-10, and RS-STOD, respectively, showing competitive or superior performance compared with representative object detection models. Ablation studies further verify the effectiveness and complementarity of the proposed modules. The results demonstrate that MWAE-YOLO can effectively enhance frequency-aware small-object feature representation, suppress background interference, and improve scale-sensitive localization accuracy in complex remote sensing scenes.