基于传感器内计算的动态视觉预测分析

Analysis of Dynamic Visual Prediction Based on Insensor Computing Network

阐述一种创新的基于多维信息的传感器内计算方法,实现图像的高效识别与预测。该方法融合储层计算(RC)网络的理论框架,并以此为基础,开发出一种基于二硫化钼(MoS_(2))的探测器阵列。充分利用探测器本身固有的光电导率(PPC)效应。通过这种方式,成功地将当前与过去的多帧信息整合至单一帧内,从而改变传统的逐帧计算模式。此外,在器件层面进行深入的预处理,细致比较不同衰减曲线对任务性能的影响。最终实现多帧字母变换预测任务,为机器视觉领域带来新的突破。

This paper describes a method of in-sensor computing with multidimensional information to achieve image recognition and prediction.In the work,incorporating the computational theory of reservoir computing(RC)networks,it developed a MoS_(2)-based detector array that functions as a dynamic photoelectronic reservoir,and by leveraging the inherent persistence of photoconductivity(PPC)effect of the detector itself,integrated present and past multi-frame information into a single frame,breaking traditional frame-by-frame computing paradigm.Furthermore,it performed preprocessing on the device to compare the impact of different decay curves on task performance,ultimately accomplishing multi-frame letter transformation classification and word prediction tasks.