摘要
细胞神经网络(cellular neural network, CNN)具有简单的局部互联结构和高速并行处理能力,是构造人工视网膜的基础模型,可被应用于机器视觉中图像处理时的图像增强等方面.然而,现有的此类图像增强方法尚存在一些不足,例如,在处理实际复杂图像时,采用固定模板难以取得理想效果;而且,未能模拟人类视觉系统的全局和局部自适应调节特性,缺乏仿生考虑.因此,本文融合自适应三高斯(tri-Gaussian)理论和纳米信息器件忆阻器,提出了一种用于图像增强的新型仿生自适应忆阻细胞神经网络.其中,基于忆阻器的可编程性、非易失性、突触可塑性等优点,构建忆阻细胞神经网络架构.基于神经元感受野三高斯模型,利用高斯核函数和细胞神经网络的图像处理特征,提出对应的仿生自适应图像增强模板设计算法.最后,分别以灰度和彩色图像为例进行了图像增强实验和对比分析,结果表明,提出的仿生自适应忆阻细胞神经网络能够显著提高图像的全局亮度、局部对比度和清晰度.本研究可为细胞神经网络提供自适应模板设计及实现方案,提升细胞神经网络的仿生特性和硬件实现的可行性,并为图像增强等智能图像处理提供新思路.
Cellular neural networks(CNN) have a simple local interconnect structure and high-speed parallel processing capability. As the basic model for a constructing artificial retina, CNNs can be applied to image enhancement in the machine vision field. However, existing image enhancement methods based on CNNs face several challenges. For example, when using common fixed templates, it is difficult to obtain ideal results when handling complex images in real-world applications. In addition, the lack of bionic considerations can cause failures when simulating the powerful global and local adaptive adjustment characteristics of human vision. Therefore,this paper proposes a biomimetic adaptive memristive CNN(BAM-CNN) that combines CNNs, human visual adaptive tri-Gaussian theory and memristor, and nano information devices. The proposed CNN can be used for image enhancement. Specifically, the memristive CNN is constructed based on emerging memristors that are programmable, non-volatile, and synapse-plastic. Merged with the tri-Gaussian model for the receptive field of neurons, an adaptive CNN template design algorithm for biomimetic image enhancement is proposed using the image processing features of the Gaussian kernel function and CNNs. In this paper, gray-scale and color images are taken as target examples in image enhancement experiments. The experimental results demonstrate that the proposed BAM-CNN significantly improves the global brightness, local contrast, and sharpness of the image. This paper provides a novel design and implementation scheme for adaptive templates of CNNs, which can improve CNNs’ biomimetic characteristics and hardware implementation feasibility. The proposed BAM-CNN can be used to develop innovative techniques for intelligent image processing besides image enhancement.
作者
郑雅文
胡小方
周跃
罗丽
段书凯
Yawen ZHENG;Xiaofang HU;Yue ZHOU;Li LUO;Shukai DUAN(College of Computer and Information Science,Southuest University,Chongqing 400715,China;College of Artificial Intelligence,Southwest Uniuersity,Chongqing 400715,China;College of Electronic and Information Engineering,Southuest University,Chongqing 400715,China)
出处
《中国科学：信息科学》
CSCD
北大核心
2020年第12期1850-1866,共17页
Scientia Sinica(Informationis)
基金
国家自然科学基金(批准号:61976246,61601376,61672436)
国家重点研发计划(批准号:2018YFB1306600,2018YFB1306604)
中国博士后科学基金特别资助(批准号:2018T110937)
重庆市留学人员创业创新支持计划(批准号:cx2019126)资助项目。
关键词
细胞神经网络
图像增强
自适应三高斯模型
仿生图像处理
忆阻器
cellular neural networks(CNN)
image enhancement
adaptive tri-Gaussian model
biomimetic imageprocessing
memristors
