Aiming at the challenges encountered in aerial drone photography under complex backgrounds, such as small targets, severe occlusions, and low detection accuracy, this paper proposes an improved YOLOv5s-based algorithm for detecting small targets in drone-captured images. Firstly, deformable convolution C3_DCN was integrated into the Backbone of the YOLOv5s model to achieve enhanced receptive field adaptability, feature representation capabilities, and positional accuracy. Secondly, a P2 detection layer was added to the Neck of the model, introducing a high-resolution feature map that extracts additional feature information, thereby improving the detection capability for small targets. Furthermore, an EMA (exponential moving average) attention mechanism was incorporated to suppress parameter oscillations, enhance the model's generalization ability, and accelerate convergence. Finally, the Focal-EIoU loss function was employed to improve the accuracy of bounding box regression and boost the model's performance on challenging samples and class-imbalanced datasets. Experimental results on the VisDrone2019 dataset demonstrate that the improved algorithm has achieved increases of 4.6%, 3.1%, and 2.9% in mAP@0.5, mAP@0.95, and F₁, respectively, significantly reducing the rates of missed detections and false positives in aerial drone imagery.