Study on retinal synapse and motion detecting unit for neuromorphic vision system implementation

Abstract

In order to implement a neuromorphic vision system, research to simulate visual information recognition of living things and data processing methods obtained through it is becoming important. In this study, we developed a retinomorphic synaptic device of a three-terminal transistor structure that responds to light and a motion detection unit that can recognize the direction and speed of object movement. Retinomorphic devices are next-generation sensor devices that have characteristics of existing devices whose current values vary depending on light intensity and illumination time, while also recognizing the direction of light. In addition to simply recognizing information about light, it also has memory characteristics. Therefore, near-sensor computing is possible through a device that can respond to light while maintaining its information. In the case of a $TiO_x$-based device included in the motion detecting unit, thin $TiO_x$ was deposited on a Pt substrate to produce oxides with different oxygen ion concentrations in the upper and lower layers. Through this, an interface-type RRAM with an Ag/$TiO_x$/Pt structure was manufactured, which is a volatile device with short-term memory characteristics that exhibit current integration and relaxation characteristics. This was integrated on top of the $NbO_x$-based threshold switching device, and HR model, widely known in motion detection models, was implemented as a memristor unit using the delay time to reach the specific resistance of the $TiO_x$-based memristor. The two detailed studies as above will help to expand the neuromorphic vision system.