Bio-inspired Saliency Computing Neural Network for UAV Small Object Detection

The detection of small objects in dynamic, high-altitude scenarios is essential for many biological and mechanical systems, such as unmanned aerial vehicles (UAVs). The midbrain network of birds, including the Optic Tectum (OT) and associated nuclei, provides a biological blueprint for efficient visual saliency detection, which is crucial for this task. This paper introduces a novel Midbrain Recurrent Neural Network (Mb-RNN) model that mimics this avian system to improve small object detection in UAVs. Based on this description, we propose a novel mathematical description of the midbrain network and develop the Mb-RNN. The proposed Mb-RNN is embedded within an RNN cell to capture the dynamic interplay of saliency detection, global inhibition, and response amplification of the avian visual system. We further build a bio-inspired UAV small object detection system, OTNet+, which integrates Mb-RNN and other deep neural networks to improve detection performance. We demonstrate the superiority of OTNet+ over state-of-the-art approaches using UAV datasets, providing a promising direction for bio-inspired computer vision systems in practical applications. Our approach provides a deeper understanding of avian neural mechanisms. It illustrates how these biological visual mechanisms can be abstracted to improve applications in computer vision, particularly in UAV-based tasks.