新型忆阻神经网络及其在图像处理中的应用

发布时间：2018-04-10 15:21

本文选题：忆阻器 + 数字逻辑器件　；参考：《西南大学》2017年硕士论文

【摘要】：随着信息量的急剧增长和信息处理要求的不断提高,人们迫切需要更加智能化和微型化的信息处理系统,因此具有并行计算优势的神经形态系统受到极大的关注。由于现有的半导体晶体管的尺寸无法进一步缩小,这使得与电子技术密切相关的神经形态系统的研究受到严重限制。忆阻器具有类似于人类大脑的“记忆”功能,其纳米级尺寸和非易失性存储的特性,有望彻底改变现有的信息处理方式。本文将忆阻器应用到神经网络系统中,提出新一代的忆阻神经网络,该网络能有效改善传统神经网络电路复杂、不易集成的缺点,在降低能耗方面也表现出强大的潜力。被认为是天然电子突触的忆阻器,能够在仿生系统里得到完美应用,让忆阻神经网络变得更加的灵活。本文深入研究了忆阻器特性,并在此基础上来构建新型忆阻神经网络电路。讨论了纯忆阻逻辑电路,并构建了忆阻数字逻辑器件;将忆阻器和neuMOS晶体管相结合,提出了新型忆阻离散Hopfield神经网络,并研究了其在彩色数字图像恢复中的应用;构建了一种参数自适应的新型忆阻脉冲耦合神经网络,并提出了一种图像增强自适应算法。具体来说,本文内容主要分为四个部分,如下所示:首先,本文重点讨论了经典的惠普忆阻器模型和阈值自适应模型,探讨了忆阻值与电荷、磁通量三者之间的关系。利用SPICE仿真验证了该模型的忆阻特性,并重点研究了该模型的阈值特性和突触特性,为忆阻器后续应用研究提供良好的理论参考和实验依据。然后,本文基于惠普忆阻器的逻辑计算能力和信息存储特性,设计了纯忆阻逻辑电路。不同于传统的忆阻逻辑电路,本文提出的电路用电压来直接表示逻辑状态,更加直观方便。相比于晶体管逻辑电路,则在电路复杂程度上有明显的改善。在此基础上,本文构建了忆阻编码器和忆阻译码器,仿真验证了其逻辑的正确性。该方案推进忆阻器在数字电路中的应用,为优化逻辑器件提供了新的思路。其次,利用神经元晶体管的加权求和特性以及阈值可控功能,结合忆阻器的突触特性,提出了一种全新的忆阻Hopfield神经网络,并将其运用在联想记忆和彩色数字图像恢复中。该网络仅由neuMOS、忆阻器和普通电阻构成,能够完全模拟神经元信息传导过程,相比传统电路,省去了复杂的差分运算电路以及电流与电压信号的转换电路,电路结构简单,可用于大规模集成。同时,该网络还具有能耗低、阈值动态可控、权值可编程的优点。可见,该方案不仅极大地简化网络结构,还能加强网络性能,有助于促进人工神经形态系统的硬件实现。最后,本文将忆阻器和传统PCNN模型相结合,提出了一种基于阈值自适应忆阻器的M-PCNN神经元模型。模型中用忆阻器电路的输出来模拟神经元间的连接强度,实现实际情况中神经元间的连接强度随外部刺激自适应动态变化的过程。这种全新的神经元模型扩展了神经网络的动态特性,为参数自适应神经网络的发展提供了一个新的思路。进一步,本文提出了一种基于M-PCNN神经元模型的自适应图像增强算法,从人眼视觉主观特性和客观性能评价指标两方面证明了该算法的优越性。该算法能突出图像局部细节,明暗对比更显著,为进一步促进神经网络在图像处理中的应用和发展奠定了基础。
[Abstract]:With the explosion of information and information processing requirements continue to increase, there is an urgent need for information processing system more intelligent and miniaturization, so it has the computational advantage of parallel neural morphological systems has attracted much attention. Due to the size of the existing semiconductor transistor cannot be further reduced, which makes the research of neuromorphic system closely related to electronic technology the limits. The memristor is similar to the human brain memory function, the nanometer size and non easy characteristics of nonvolatile storage, is expected to completely change the existing mode of information processing. In this paper the memristor is applied to the neural network system, a new generation of memristive neural network. This network can effectively improve the traditional neural network circuit is complex and not easy to integrate the disadvantages in terms of reducing energy consumption shows great potential. Considered natural electronic process The memristor touch, can obtain perfect application in biomimetic system, let memristive neural network become more flexible. This paper studies the memristor characteristics, and on this foundation to construct the model of memristor neural network circuit is discussed. The pure memristor logic circuit, and the construction of the memristor digital logic device; the the memristor and neuMOS transistor combination, put forward the new memristor discrete Hopfield neural network, and discussed its application in color digital image restoration; constructed a model of memristor parameter adaptive pulse coupled neural network, and put forward a kind of adaptive image enhancement algorithm. Specifically, the main content of this article four parts as follows: firstly, this paper discusses the classical HP memristor model and adaptive threshold model, discusses the relationship between memory resistance and charge flux is three. By SPICE simulation test The characteristic of the memristor model, and focuses on the threshold characteristics and synaptic characteristics of the model, providing theoretical and experimental evidence for good memristor subsequent application research. Then, the computing power and information storage characteristics of HP memristor based on the logic design, the pure memristor logic circuit. In the memristor traditional logic circuit, this voltage to said logic state, more intuitive and convenient. Compared to the transistor logic circuit, it has obvious improvement in circuit complexity. On this basis, this paper constructs the memristor encoder and decoder of memristor simulation, verified the correctness of the logic. The scheme to promote the application of memristor in digital circuit, provides a new idea for the optimization of logic devices. Secondly, using the weighted sum of the neuron transistor characteristics and threshold controllable function, combined with the memristor Synaptic characteristics, proposed a new memristor Hopfield neural network and its application in the recovery of associative memory and color digital image. The network only by neuMOS, a memristor and ordinary resistance, can fully simulate the process of neuronal information transmission, compared with the traditional circuit, eliminating the need for complex difference operation circuit, current and voltage signal conversion circuit, the circuit structure is simple, can be used for large-scale integration. At the same time, the network also has the advantages of low energy consumption, dynamic threshold controllable, the advantages of programmable weights. Obviously, this scheme not only greatly simplify network structure, can enhance network performance, help to promote the form of artificial neural system the hardware implementation. Finally, the memristor and the combination of the traditional PCNN model, proposes a M-PCNN neuron model with adaptive threshold based on memristor. Output model with the memristor circuit to die The connection strength between neurons, the connection strength between neurons in the actual situation with the external stimulus. The dynamic changes of the adaptive neuron model extends the dynamic characteristics of the new neural network, provides a new idea for the development of adaptive neural network. Further, this paper proposes an adaptive image enhancement algorithm the M-PCNN neuron model based on human visual subjective and objective characteristics of two aspects of performance evaluation index to prove the superiority of the algorithm. The algorithm can enhance the image details, the contrast is more significant, which lays a foundation for the further promotion of the development and application of neural network in image processing.

【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】