Dynamic machine vision with retinomorphic photomemristor-reservoir computing

Dynamic machine vision requires recognizing the past and predicting the future of a moving object based on present vision. Current machine vision systems accomplish this by processing numerous image frames or using complex algorithms. Here, we report motion recognition and prediction in recurrent photomemristor networks. In our system, a retinomorphic photomemristor array, working as dynamic vision reservoir, embeds past motion frames as hidden states into the present frame through inherent dynamic memory. The informative present frame facilitates accurate recognition of past and prediction of future motions with machine learning algorithms. This in-sensor motion processing capability eliminates redundant data flows and promotes real-time perception of moving objects for dynamic machine vision. Dynamic machine vision (DMV) technology has numerous significant applications in video analysis, robotic vision, self-driving technology, and intelligent transport. The ability to use present vision to recognize past motion and predict future trajectories is crucial in DMV. Current imaging systems utilize multiple modules, including sensors, signal converters, memory, and processors, to recognize and predict motion by analyzing massive frame-by-frame image sequences and using complex algorithms, engendering redundant data flows and high-energy consumption. Different from modern image sensing and processing systems, the biological architecture of human vision is highly capable of recognizing and predicting motion, for instance, aiding humans in the perception of danger in wildlife or traffic. In recent years, inspired by the biological vision system wherein visual short-term memory plays a key role, retinomorphic image sensors with memory capability, such as switchable photovoltaic sensors, non-volatile phototransistors and memristors, have shown adaptive and all-in-one sensing capability, facilitating in-sensor computing, self-adaptive imaging, and motion detection. Besides, in-sensor reservoir computing systems with spatiotemporal processing capabilities have been demonstrated for language learning and image classification. However, motion recognition and prediction within a compact dynamic sensing system, which is crucial for DMV technology, has not been realized yet. Here, we report recurrent photomemristor networks consisting of a retinomorphic photomemristor array (PMA) operating as a dynamic vision reservoir and readout networks for processing. In the retinomorphic photomemristor-reservoir computing (RP-RC) system, the inherent dynamic memory of the PMA stores spatiotemporal information of a frame-by-frame visual sequence as hidden states in the last frame. The dynamic PMA reservoir, containing all the past spatiotemporal visual information, is used for various dynamic processing tasks through the training of readout networks. To demonstrate the spatiotemporal processing capability of the RP-RC system, we implement the classification of videos playing English words ending with the same letter but with different spatiotemporal dynamics for language learning. Furthermore, we realize the most crucial DMV task—motion recognition and trajectory prediction—in the RP-RC system using classification and inherent memory association by the readout networks, providing a promising neuromorphic platform for in-sensor DMV.

Hongwei Tan, Sebastiaan van Dijken