<p>The detection of small objects in complex scenes remains challenging due to low resolution, weak textures, and background clutter. To address this, we propose a new architecture for small object detection that integrates three complementary modules. First, we propose a large-kernel Dual-Path Spatial Attention (DPSA) module that explicitly enlarges the receptive field and boosts responses in salient regions while preserving fine details. Second, we design a Detail-Injected Feature Pyramid Network (DiFPN) module that adopts a dual-path architecture. In this design, deep semantic features flow top-down, while high-resolution shallow features are aligned and injected into middle and high levels to reinforce detail preservation during fusion. This process enables rich semantic-spatial fusion and effectively mitigates the loss of small object details due to repeated downsampling. Third, to improve gradient flow and multi-scale representation with manageable overhead, we design a Deep Point-Wise Residual Block (DPWR) module that replaces key backbone units of the original YOLOv8s framework. Experiments on the VisDrone2019 dataset demonstrate that our method surpasses the baseline by 6.7 percentage points in mAP50 and by 4.5 percentage points in mAP50:95. Across densely packed scenes, our detector not only minimizes object merges and misses by reliably separating adjacent instances in crowded scenes but also enhances robustness against occlusion and varying illumination. This results in fewer false positives, more precise box boundaries, and improved recall. It separates adjacent instances more reliably, reducing object merges and misses; under occlusion, strong light, and low-light conditions, it suppresses false positives, yields more stable box boundaries, and increases recall. These results confirm that our design provides consistent improvements in recall, false-positive control, and localization accuracy for small objects, offering an effective solution for complex real-world applications.</p>