A Reservoir Computing-Based Binocular Vision System for Stereoscopic Spatial Classification Using Organic Ferroelectric Phototransistors

Current research on artificial vision systems primarily focuses on monocular vision systems (MVSs), which have inherent limitations in spatial perception and recognition accuracy. Here, we present a reservoir computing-based binocular vision system (RC-BVS) using PDVT-10 organic ferroelectric phototransistors (OFePTs) with P(VDF-TrFE) gate dielectric. The devices leverage oxygen-mediated persistent photoconductivity (PPC) for reservoir dynamics and ferroelectric polarization for hardware-implemented synaptic weights with 50 distinct analog states. Our system integrates two binocular vision mechanisms: a cosine-based intensity attenuation model for horizontal positioning and a perspective transformation model with depth-dependent size scaling for distance perception. By processing asymmetric binocular inputs through intensity heterogeneity and geometric disparity, the OFePT arrays generate diverse reservoir states encoding position and depth information. The RC-BVS simultaneously achieves digit recognition (87.33%), horizontal position recognition (99.98%), and depth position recognition (99.51%), significantly outperforming monocular systems. Notably, our approach reduces computational cost to approximately 1/32 of the multiply–accumulate operations of a standard multilayer perceptron while enabling on-chip weight storage and updates through ferroelectric conductance modulation. This work demonstrates a design strategy for artificial vision systems with stereoscopic information processing.