Bio-inspired small target motion detection with time-delay feedback of full Hassenstein-Reichardt correlator in large scene

In large-scale backgrounds, small moving objects occupy only a few pixels in an image, with large receptive fields and extremely limited visual features. Deep learning-based object detection methods face significant challenges in detecting small targets, as repeated convolution and down-sampling operations nearly eliminate target features within the receptive field. The highly evolved visual systems of insects enable them to efficiently purse tiny prey and mates, synchronously detecting the visual features of motion targets across wide fields from complex dynamic environments. Exceptional sensitivity to small moving targets relies on specialized neurons known as Small Target Motion Detectors (STMDs). Existing STMD-based neural models typically consist of four sequentially arranged neural layers connected by feedforward loops to extract small target motion information from raw visual images. However, these STMD methods often experience performance degradation in complex, wide-field background contexts. In this study, we propose a bioinspired visual system based on Time-delay feedback of full Hassenstein-Reichardt Correlator (HRC) STMD, designed to suppress pseudo-small moving targets while enhancing the system's response to actual small targets. Specifically, the proposed visual system primarily utilizes the Retina layer, Large monopolar cells, and the Medulla layer as input and preprocessing units for photoreceptors. These units are responsible for receiving high-resolution images over a large scene and initially extracting shadow motion features of moving targets. Subsequently, the motion information are input into the STMD module, which is centered around time-delay feedback of full HRC. This module can respond to small targets with varying luminance under low background illumination. The time-delay feedback enhances the response of STMD to small moving targets, facilitating the filtering of false positive responses. Experimental results on synthetic and real-world datasets demonstrate that the proposed small target motion detection visual system is more competitive than existing methods in recognizing tiny moving targets from complex natural environments in large scene.