In-sensor image memorization, low-level processing, and high-level computing by using above-bandgap photovoltages

In-sensor computing holds great promise for ultrafast and energy-efficient machine vision. However, the development of a versatile in-sensor computing system that can integrate image memorization, low-level processing, and high-level computing functions remains a challenge, primarily due to the limited availability of photosensors that can offer both dynamic photoresponse and programmable photoresponsivity. Here, we demonstrate the integration of these multi-functions into a ferroelectric photosensor (FE-PS)-based array. Thanks to its unique photovoltaic mechanism known as the bulk photovoltaic effect, the FE-PS exhibits above-bandgap, dynamically responding, and electrically switchable photovoltages. By using the dynamic photovoltage response, the FE-PS-based array is capable of memorizing and pre-processing images, with the ability to adjust the memory and pre-processing effects by ferroelectric polarization. On the other hand, the electrically switchable photovoltages featuring multi-level switchability and retrievability enable the FE-PS-based array to perform in-sensor high-level computing, achieving 100% accuracy in a 4-class image recognition task. Notably, the high precision and reliability of the photovoltage-based image memorization and processing greatly benefit from the high photovoltages produced by the FE-PS. This study represents a significant step towards developing versatile in-sensor computing systems that could be utilized across a wide range of machine vision scenarios.