In complex backgrounds, the precise detection of small targets in unmanned aerial vehicle (UAV) images remains a crucial yet highly challenging task, especially considering that objects in natural scenes often display significant scale variations. These variations can lead to incomplete feature extraction and consequently hinder accurate recognition. As a common solution, Feature Pyramid Networks (FPNs) have been introduced to enhance multi-scale feature representation. However, existing FPNs typically employ simple feature concatenation and addition operations that are insufficient for capturing rich semantic information, thus constraining their effectiveness in handling large scale variations within complex scenes. To address this issue, we propose a novel Hierarchical Cross-scale Feature Fusion Pyramid Network (HCFFPN) that captures semantic relationships across adjacent feature levels via a Multi-level Feature Extractor (MLFE) to enhance inter-level correlation, and integrates high- and low-frequency information through a Deep Feature Aggregation Module (DFAM), resulting in more robust feature representations. To verify the effectiveness of the proposed HCFFPN, we conducted a large number of experiments and comprehensive evaluations on the VisDrone and UAVDT datasets. The experimental results show that, compared with existing FPN-based architectures, the proposed HCFFPN demonstrates significant improvements in both detection accuracy and robustness for small targets in complex UAV scenes.