EVMNet: Eagle visual mechanism-inspired lightweight network for small object detection in UAV aerial images

In recent years, unmanned aerial vehicles (UAVs) have experienced rapid development. However, image recognition tasks from the UAV's viewpoint frequently encounter difficulties, mainly due to the small object sizes, the intricacy of the scenes, and the vast quantity of network parameters, which hinder deployment on edge devices. To tackle these problems, this study proposes a novel lightweight one-stage object detection network based on an improved version of YOLOv9, inspired by the visual mechanisms of eagles, known for their high acuity, wide field of view, and rapid localization abilities. The network emulates the eagle's primary visual pathway, the Tectofugal Pathway, and reconstructs the overall network architecture, significantly reducing the number of parameters. The proposed Shallow & Deep Fovea Vision Block, inspired by the eagle's dual-fovea structure, endows the network with efficient feature extraction capabilities, leveraging the advantage of a large receptive field. Finally, the design of the nucleus rotundus module further enhances the network's ability to match the receptive field with global context, thereby improving detection performance. Ablation and comparative experiments were conducted on three UAV small object datasets: VisDrone2019, UAVDT, and AI-TOD, using the proposed EVMNet. Compared to YOLOv9t, the proposed EVMNet significantly reduces the number of parameters to only 1.46M, while improving the mAP0.5 metric on the VisDrone2019 dataset by 12.5%, reaching an accuracy of 49.2%. Considering both network performance and the number of parameters, our network achieves state-ofthe-art (SOTA) results.