Despite that in-sensor processing has been proposed to remove the latency and energy consumption during the inevitable data transfer between spatial-separated sensors, memories and processors in traditional computer vision, its hardware implementation for artificial neural networks (ANNs) with all-in-one device arrays remains a challenge, especially for organic-based ANNs. With the advantages of biocompatibility, low cost, easy fabrication and flexibility, here we implement a self-powered in-sensor ANN using molecular ferroelectric (MF)-based photomemristor arrays. Tunable ferroelectric depolarization was intentionally introduced into the ANN, which enables reconfigurable conductance and photoresponse. Treating photoresponsivity as synaptic weight, the MF-based in-sensor ANN can operate analog convolutional computation, and successfully conduct perception and recognition of white-light letter images in experiments, with low processing energy consumption. Handwritten Chinese digits are also recognized and regressed by a large-scale array, demonstrating its scalability and potential for low-power processing and the applications in MF-based in-situ artificial retina.

尽管传感器内处理已被提出来消除传统计算机视觉中空间分离的传感器、存储器和处理器之间不可避免的数据传输过程中的延迟和能耗，但其用于人工神经网络（ANN）的硬件实现一个设备阵列仍然是一个挑战，特别是对于基于有机的人工神经网络。凭借生物相容性、低成本、易于制造和灵活性等优点，我们使用基于分子铁电（MF）的光忆阻器阵列实现了自供电传感器内人工神经网络。人工神经网络特意引入了可调谐铁电去极化，从而实现了可重新配置的电导和光响应。将光响应性视为突触权重，基于 MF 的传感器内 ANN 可以进行模拟卷积计算，并在实验中成功实现了白光字母图像的感知和识别，且处理能耗低。手写中文数字也可以通过大规模阵列进行识别和回归，展示了其可扩展性和低功耗处理的潜力以及基于 MF 的原位人工视网膜的应用。