A Lightweight Tiny Object Detection Network Inspired by the Visual Neural Mechanism of Eagle
Xi Chen, Chuan Lin · 2024
In recent years, unmanned aerial vehicles (UAVs) have advanced significantly, but image recognition from their perspective still faces challenges. These issues mainly stem from tiny object sizes, complex scene structures, and the large number of network parameters, all of which complicate the effective deployment of models on edge devices. To tackle these problems, this study proposes a novel lightweight one-stage object detection network, named VENet. The design of VENet is influenced by the visual mechanisms of eagles, which are known for their exceptional visual acuity, wide field of view, and rapid localization capabilities. VENet leverages an enhanced version of YOLO11 to achieve efficient object detection while maintaining high performance, particularly in terms of speed and accuracy. 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 C3-Eagle and Dual-Fovea 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. Comparative experiments were conducted on three UAV tiny object datasets: VisDrone2019, AI-TOD and UAVDT, using the proposed VENet. Compared to YOLO11n, the proposed VENet significantly reduces the number of parameters to only 1.9M, while improving the mAP0.5 metric on the VisDrone2019 dataset by 13.9%, reaching an accuracy of 48.6%. Considering both network performance and the number of parameters, our network achieves state-of-the-art (SOTA) results.