Bio-inspired small target detecting visual neural network with motion direction decoding compensation in large scene

Small target moving detection in a wide field of view and complex background is an extremely challenging task due to the small number of target pixels and interference from background noise. Surprisingly the visual systems of Drosophila accurately detect mates and track prey during rapid flight, even though targets occupy a minuscule portion of the visual fields. The heightened sensitivity to small targets is supported by specialized neurons known as Small Target Motion Detectors (STMD). Existing STMD models based on Hassenstein-Reichardt Correlator rely heavily on visual contrast and the temporal delay strategy of this correlator results in detection outcomes lagging behind ground truth for current frames. In this paper, we develop a bio-inspired visual system with motion direction decoding compensation mechanism. Specifically, the proposed visual neural network comprises two complementary submodules and a compensation channel. The first submodule extracts spatial and temporal motion patterns of targets by neuronal direction decoding. The second submodule captures small target motion information where its output integrates with signals from the first submodule via the compensation channel, thereby improving the detection rate while reducing the interference of small-target-like background noise. Experimental results demonstrate that our proposed vision system based on motion direction decoding compensation exhibits superior competitiveness compared to existing methods in discriminating small moving targets from complex large scenes.