DPSH-YOLO: An Efficient Detector for Capturing Local Details and Global Semantics of Tiny Objects

With the rapid development of high-resolution remote sensing technology, detecting tiny objects in complex scenes still faces challenges such as weak feature representation and strong background interference. Current mainstream detectors have limitations in capturing fine-grained details of tiny objects while balancing efficiency and accuracy. To address these issues, we propose a new detector called DPSH-YOLO, which efficiently captures both local details and global semantics for tiny objects. First, we design a Dynamic Perception Aggregation (DPA) module that dynamically captures local detailed information along both horizontal and vertical directions, and effectively aggregates it with global information, thereby enhancing edge detail and semantic feature representation. Then, we introduce a dual-branch Snake Hybrid Attention (SHA) module, which fuses the local details from the high-resolution branch with the global context from the low-resolution branch to mitigate background interference. Experimental results on the VEDAI and AI-TOD remote sensing tiny object detection datasets show that DPSH-YOLO achieves accuracies of 78.18% and 65.43%, respectively (using mAP50 as the evaluation metric), significantly outperforming current state-of-the-art methods.