To address the low detection accuracy caused by irregular target shapes, large scale variations, and dense distributions in UAV aerial images, this paper proposes a multi-scale feature fusion algorithm for small-target detection. First, to better represent irregular targets, a feature extraction structure combining deformable convolution with a multi-branch fusion module is designed. Then, to counter the feature loss and inadequate contextual information for small targets amidst large scale variations, a dynamic adaptive feature pyramid network is constructed by integrating high-resolution detection heads and dynamic weighting mechanisms. To resolve missed detections of overlapping targets in dense scenes, which is a common flaw of standard suppression methods, a strategy combining SoftNMS and DIoU is proposed. Finally, model pruning is applied for lightweight deployment. Experimental results on the VisDrone2019 dataset demonstrate that the lightweight model achieves 48.3% mAP \(_{0.5}\) and 32.0% mAP \(_{0.5:0.95}\) , representing substantial improvements over YOLOv8s, validating both the effectiveness and efficiency of the proposed method.