神经形态工程学
计算机科学
人工智能
傅里叶变换
人类视觉系统模型
计算机视觉
特征提取
模式识别(心理学)
快速傅里叶变换
电子工程
人工神经网络
算法
物理
图像(数学)
工程类
量子力学
作者
Baocheng Peng,Qianlu Sun,Haotian Long,Ke Xu,Lesheng Qiao,Zehua Hu,Changjin Wan,Qing Wan
摘要
The hierarchical structure of the biological visual system enables multilevel features of sensory stimuli to be pre-extracted before being transmitted to the nerve center, rendering the remarkable ability to perceive, filter, categorize, and identify targets in complex environments. However, it is a challenge to resemble such extraction capability with respect to spatial features in a neuromorphic visual system. In this Letter, we propose an indium-gallium-zinc-oxide synaptic transistor-based Fourier neuromorphic visual system for image style classifying. The images are transformed into the frequency domain through an optic Fourier system, greatly reducing energy and time dissipation in comparison with numerical computation. Then, the transformed information is coded into spike trains, which are nonlinearly filtered by synaptic transistors. The energy consumption for this filtering process is estimated to be ∼1.28 nJ/pixel. The features of drawing style could be enhanced through the filtering process, which facilitates the followed pattern recognition. The recognition accuracy in classifying stylized images is significantly improved to 92% through such Fourier transform and filtering process. This work would be of profound implications for advancing neuromorphic visual system with Fourier optics enhanced feature extraction capabilities.
科研通智能强力驱动
Strongly Powered by AbleSci AI