ESO-YOLO: Enhanced Small Object Detection Algorithm from Multiple Perspectives

Accurate recognition of low-resolution small objects represents a critical technical challenge across diverse fields. Such small objects—exemplified by traffic signs—are indispensable for behavioral decision-making in autonomous driving and target localization in disaster monitoring and rescue. However, prevailing object detection algorithms commonly suffer from small-object feature loss, insufficient detection accuracy, as well as complex network architectures and excessive parameter counts.To address the limitations in small-object detection, this study proposes ESO-YOLO, an enhanced small-object detection algorithm built upon the YOLOv11 framework. The algorithm achieves active feature fusion via the construction of an Efficient Feature Fusion Module (EFFM), mitigates information loss of small objects through the designed Lightweight Spatial Down-sampling (LSDown), and explicitly preserves shallow fine-grained features by proposing Learnable Shallow Bypass (LSBypass) integrated into the LSC3 module. These improvements enhance small-object detection performance at the levels of network architecture and feature processing, while maintaining the lightweight nature of the algorithm.Experimental validation is conducted on the TT100K traffic sign detection dataset and the VisDrone2019 UAV aerial dataset. The proposed model achieves significant improvements in both detection accuracy and recall rate, accompanied by a reduced parameter count, and demonstrates superior cross-scene detection capability in generalization experiments.Extensive comparative experiments and ablation studies verify that the model presented in this paper effectively alleviates feature loss and background interference in small-object detection. It enhances small-object detection performance while realizing network lightweighting, exhibiting favorable practical application value and strong generalization ability.