Anotace:
Due to challenges such as the small size of targets, complex backgrounds, limited feature extraction capa-bilities, and frequent false positives and false negatives, traditional detection algorithms often perform poorly in small object detection tasks. To address these challenges, this pa¬per proposes an enhanced small object detection algorithm, SOD-YOLO, based on YOLOv8s. First, the S_C2f_CAFM module is integrated into the feature extraction network, enabling the effective capture of fine-grained local features and broad contextual information, while simultaneously reducing model parameters and computational complexity. Second, in the feature fusion stage, the redesigned bidirectional feature pyramid network employs a spatial context awareness module to extract key features, adding a top-down path to optimize feature fusion and enhance discriminative information. In the Neck section, the D_C2f_MSPA module is introduced, which, while being lightweight, accurately models channel dependencies in feature maps, effectively reducing both false positives and false negatives for small objects. Finally, the inclusion of Normalized Wasserstein Distance (NWD) further improves detection accuracy and reduces the model’s sensitivity to small positional deviations in small objects. Experimental results on the DOTAv1.0, VisDrone2019, and TT100K datasets confirm that SOD-YOLO achieves excellent performance, demonstrating the effectiveness of the modifications made to the original YOLOv8 model.