Abstract
Visual adaptation of the retina has enabled to adapt and perceive a wide-range light stimuli, which however induces long time adaptation process leading to blindness, dizziness or other potential hazards1–3. In the realm of machine vision, when replacing the human retina for real-time image processing, the intricate circuits and algorithms are essential to ensure optimal performance and accurate recognition, as currently developed vision perception systems struggle to adapt images with varying brightness levels. Despite the attention garnered by adaptive devices, their potential application in machine vision systems is hampered by a sluggish adaptation process, making integration challenging.4–10 Here, we take advantage of avalanche tuning as feedforward inhibition in bionic two-dimensional (2D) transistor to realize active and rapid switchover of light perception mechanism for fast and high-frequency visual adaptation behavior, avoiding the usual long visual adaptation process and occurrence of potential harms. The adaptation speed is over 104 times faster than that of the human retina and the currently reported bionic sensors relying on feedback inhibition circuit. The sense-computing integrated junction-field-effect transistor (JFET) exhibits an extraordinary avalanche performance with low breakdown voltage (VEB) of approximately 10 V and high multiplication factor of 8.2×103, which can be tuned by gate voltage and light intensity and outperforms the state-of-the-art 2D avalanche transistors. By changing light stimulus from dim to bright, the sensory responsivity experiences great changes in both magnitude and sign (from 9.6×105 to -4×103 A/W), due to spontaneous transition of the photo-sensing mechanism between avalanche and photoconductive effect. Notably, this mechanism switches much faster than the chemical reaction between rod and cone cells, and the charge trapping/de-trapping mechanism in existing 2D machine vision systems. Thus, the device can emulate high-frequency visual behavior at 4 and 2.5 kHz under simulated scotopic and photopic adaptation conditions, possessing ultra-fast adaptation process of 142 and 427 µs, respectively, that is far beyond human retina function with long adaptation process up to few minutes. The − 3 dB bandwidth of our device reaches 8.3 kHz at weak light, also surpassing dynamic response of retina (500 Hz)11. More importantly, an ultra-fast adaptative machine vision has been achieved by integrating convolutional neural networks with bionic avalanche transistor, making a groundbreaking achievement with respect to its remarkable microsecond-level rapid adaptation capabilities and robust image recognition with over 97% precision in both dim and bright conditions. This work introduces an innovative bio-inspired vision device that relies on avalanche tuning operation as a faster and more predictive feedforward inhibitory mechanism, holding a huge application potential in next generation of high-frequency machine vision systems, promoting the technological innovation in field of autonomous driving and facial recognition, etc.