Cross-resolution semantic-guided small object detection via feature distribution self-alignment

Small object detection remains a bottleneck in computer vision tasks, as small objects suffer from insufficient visual information yet exhibit complex feature representation characteristics. Existing multi-scale feature fusion pipelines commonly adopt rigid structural alignment strategies, which tend to induce the feature space shift issue and hinder the model’s ability to fully capture the native features of small objects. To address this limitation, this paper proposes CRSA-YOLO, an improved algorithm built upon YOLOv11 that integrates cross-resolution semantic alignment and automatic feature distribution coordination. First, a cross-resolution semantic guidance module is devised to enable high-resolution feature maps to supervise the feature learning of low-resolution counterparts. Second, a Spatial-frequency dual-domain collaborative constraint loss (SF-DCCL) is proposed to facilitate efficient fusion of spatial and frequency-domain information. Furthermore, Cross-Source Feature Distribution Collaborative Optimization (CF-DCO) is introduced to automatically harmonize feature distributions across diverse real-world datasets, strengthening the generalization capability of the detection model. Extensive experiments are conducted on benchmark datasets including VisDrone and DsVOC2012. On the VisDrone test set, the proposed method elevates the average precision for small objects AP_\textS from 9.1% to 12.9%, which verifies the significant improvement in detection accuracy and robustness for small targets.