Small object detection concentrates on detecting objects with small size, holding great theoretical and practical significance in various scenarios including surveillance, agriculture monitoring, unmanned aerial vehicle, etc. Although deep learning networks have achieved remarkable success in the field of object detection, their generic feature fusion strategy continues to hinder further progress in small object detection. The traditional feature fusion strategy adopted in feature pyramid networks exhibits spatial information loss due to repeated downsampling operations, while reconstruction based on distorted low-resolution information further exacerbates the secondary attenuation of texture and contour. Meanwhile, the simple fusion operations that neglect cross-layer feature discrepancies fail to fully utilize multi-scale information, leading to suboptimal fusion. These limitations critically impair small object detection, since small objects are especially sensitive to spatial detail information. To this end, a task-specific feature fusion strategy is proposed for small object detection. The proposed strategy consists of two modules, the Global Information Compensation Module (GICM) and the Dual Path Alignment Module (DPAM). GICM can integrate global spatial information and employs a High-frequency Enhancement Block to filter out noise within these features, thereby compensating the loss of fine details in high-level features. DPAM learns the correlations between features from adjacent layers in the pyramid to achieve alignment across multi-scale features. The proposed strategy is integrated into the YOLOv11 framework through neck component replacement. Experimental results on VisDrone and Tsinghua-Tencent 100K datasets demonstrate the superiority of our strategy.