Machine vision systems are crucial in intelligent scenarios, but actual image acquisition is frequently compromised by the inadequate proficiency of photosensors in photoadaptation. Inspired by biological vision, neuromorphic synaptic phototransistors endowed with photoadaptive capabilities have emerged as a prospective strategy. However, most synaptic phototransistors only exhibit unidirectional positive photoresponses, whereas those capable of bidirectional photoresponses offer a greater possibility of accurately capturing images in complex lighting scenes. Herein, bidirectional photoadaptable organic heterojunction synapse phototransistors as sensing and processing units in systems are reported, which facilitate image contrast enhancement and improve image feature extraction under adverse lighting conditions. The bidirectional plasticity transformation of biomimetic neuromorphic synapses is mimicked. Specifically, n–n heterojunctions exhibit a unidirectional excitatory postsynaptic current, whereas n-p heterojunctions show a bidirectional response with a more prominent inhibitory postsynaptic current. Most interestingly, by integrating the device characteristics into convolutional neural networks and simultaneously optimizing algorithm architecture, the details and edges of low-contrast images are markedly enhanced, and the accuracy of image recognition is increased to 97.4% within ten cycles. This work serves as a novel idea for the development of high-performance neuromorphic visual systems, rendering them promising candidates for in-sensor computing applications.
Loading....