Core Concepts
A novel neural network called TSOM is proposed for detecting small object motion, which is inspired by the Retina-OT-Rt neural circuit in the avian visual system and exhibits high biological interpretability and superior performance in extracting small moving objects from complex backgrounds.
Abstract
The paper presents a brain-inspired neural network model called TSOM (Tectum Small Object Motion detector) for detecting small moving objects in complex backgrounds. The model is inspired by the Retina-OT-Rt neural circuit in the avian visual system, which is highly sensitive to small moving objects.
The key highlights and insights are:
The authors mathematically model the Retina-OT-Rt neural circuit, including the spatiotemporal dynamics of SGC neuron dendrites and the two-stage motion information integration in the Rt projection.
The TSOM neural network model is designed based on the mathematical description of the Retina-OT-Rt circuit, consisting of four subsystems: the retina layer, SGC dendritic layer, SGC soma layer, and Rt layer.
The retina layer maps the input image to photoreceptors, the SGC dendritic layer encodes spatiotemporal information, the SGC soma layer computes motion information and extracts small objects, and the Rt layer integrates motion information from multiple directions to determine the position of small objects.
Experiments on pigeon neurophysiological data and image sequence datasets (BEVS and RIST) demonstrate that the TSOM model has high biological interpretability and superior performance in detecting small moving objects compared to other advanced methods.
Stats
The background moves at a velocity of 150 pixel/s along a fixed direction.
The object moves at a velocity ranging from 10 to 400 pixel/s.
The object radius ranges from 0 to 20 pixels.
Quotes
"Detecting small moving objects in complex backgrounds from an overhead perspective is a highly challenging task for machine vision systems."
"Birds exhibit superior visual abilities in situations of high altitude, and their excellent small object motion detection ability can provide new insights for current artificial intelligence algorithms."