【作者】 李小军；

【导师】 马义德；

【作者基本信息】 兰州大学 ， 无线电物理， 2012， 博士

【摘要】 脉冲耦合神经网络(Pulse Coupled Neural Network, PCNN)是基于哺乳动物视皮层神经元机理建立起来的人工神经网络模型。由于它更接近生物视皮层神经元对信息的处理方式,所以在智能信息处理领域得到了广泛的研究和应用。然而,针对具体的实际应用要求和特点尝试完善和改进PCNN模型,往往可以取得更好的处理结果。PCNN众多的有益特性使基于PCNN的纹理检索和最短路径求解等应用研究一直是热门的研究领域。传统的纹理检索方法大多都假设纹理图像具有相同的旋转、位移、尺度且未发生扭曲,而现实生活中很难保证所有纹理都是在相同的角度、位置、聚焦长度和不受污染变形的情况下捕获的。同时现有的不变纹理检索方法大都没有考虑到纹理同时发生位移、尺度和旋转变化或仿射等几何变化对系统的影响。PCNN的时间序列特征具有这些不变特性,但是针对纹理检索的实际应用进一步改进PCNN模型使其获得更为稳定和有效的特征向量具有一定的研究价值。另外,PCNN具有自动波特性和神经元并行处理特性,使得它成为解决最短路径问题的有效手段之一。结合最短路径问题的特点实现PCNN更多的有益改进将可以进一步缩短最短路径求解的迭代时间。本论文围绕基于PCNN的不变纹理检索系统和最短路径求解算法展开研究。论文的主要研究内容和创新点如下：1.针对不变纹理检索问题,提出一种DPCNN (dual-output PCNN)模型。该模型相比标准PCNN模型,具有一些新的特点：(1) DPCNN的每一个神经元具有两次机会使自己被激发；(2) DPCNN能够自动调整输入激励大小；(3)接收到的来自周围神经元的局部刺激会受到输入激励的调制影响；(4) DPCNN模型被认为符合人眼视觉系统。在不变纹理检索中DPCNN新的特性使其可以获得更为稳定的图像特征。实验结果表明,使用DPCNN的纹理检索方法具有很好的位移、尺度、旋转和仿射不变性。同时DPCNN模型也具有较好的抗噪声性能。2.在DPCNN模型中引入彩色补偿机制,使其可以应用于彩色纹理检索。由于单一的DPCNN模型只允许接收单通道的外部激励,所以不能直接应用于彩色纹理图像的检索。但是单纯地把彩色纹理图像当做灰度图像来处理,必然会导致系统对色彩相关信息利用的不足。本文将明度通道作为DPCNN模型的常规输入,同时考虑色调和色饱和度通道的像素值作为补偿输入。实验结果表明,考虑彩色补偿机制的纹理图像检索系统优于单纯考虑单通道灰度纹理的方法和传统的彩色纹理检索方法。3.尝试提出一种PCNN相关模型三通道彩色纹理检索框架。该方法分别对彩色纹理的各个通道采用PCNN相关模型处理,最后在不增加特征长度的情况下合并各个通道的特征向量。实验表明,该框架可以有效地解决PCNN相关模型对彩色纹理的检索问题。4.针对最短路径求解问题,提出一种SAPCNN (self-adaptive autowave PCNN)模型。该模型可以根据当前网络状态自适应地调整自动波传播速度,从而更快地找到最短路径的解。同时,论文将基于SAPCNN模型的最短路径求解方法应用到K则最短路径问题,实验表明该方法较传统方法能够减少K则最短路径问题求解的时间消耗。更多还原

【Abstract】 Pulse Coupled Neural Network (PCNN) is an Artificial Neural Network which is proposed based on the mechanism of mammal visual cortical neurons. It is close to the approach of information processes by visual cortical neurons, and therefore is widely studied and used in the field of intelligent information processing. However, if the PCNN model could be improved and modified based on the requirements of practical application, the better results will be obtained.Since PCNN has so many useful characteristics, the applications of texture retrieval and finding shortest path by using PCNN have become some of the hottest research points. Many existing texture retrieval methods assume that all texture images are identical with respect to rotation, translation, scale and without distortion. However, it is hard to ensure that all the images have been captured under the conditions of the same angle, position, focal length and without distortion in the real world. Moreover, most existing invariant texture retrieval methods did not consider the influences of rotation, scale and translation at the same time, or the affine transform of the input texture. Although time series of PCNN have the invariant characteristics, it is valuable to improve PCNN model for obtaining more stable and effective feature vectors based on the practical application of texture retrieval. In addition, PCNN has the autowave and parallel processing characteristics, which makes it suitable for dealing with the shortest path problem. However, it will reduce the iteration times in solving the shortest path problem by doing some useful modifications in PCNN model based on the characteristics of shortest path problem.The paper focuses on the PCNN based invariant texture retrieval systems and shortest path problem. The main contents and innovative points of this dissertation are as follows:1. For invariant texture retrieval, a new DPCNN model (dual-output PCNN) is proposed. Compared with standard PCNN model, the proposed DPCNN has some new characteristics:(1) each neuron has two opportunities to get itself stimulated;(2) DPCNN model can automatically change the external stimulus of the neuron;(3) the received local stimulus from neighbors of a neuron is controlled by the external stimulus;(4) DPCNN model is believed to be coincident with human visual characteristic. The new characteristics are applied to obtain a more stable texture features in invariant texture retrieval. Experimental results demonstrate that DPCNN based texture retrieval method has better translation, scale, rotation and affine invariance. Moreover, once the texture images are contaminated by noise, DPCNN shows good performance as well.2. Color compensation is introduced in DPCNN model for its application in color texture retrieval. Because a single DPCNN model can be only used for one channel input stimulus, it can not be directly applied to the color texture retrieval. However, if the color texture images are processed only as gray texture images, it will lead to the insufficient utilization of color information. The method regards value channel as the regular input of DPCNN, and considers hue and saturation channels as its compensation inputs. Experimental results show that the proposed method outperforms the methods which only consider the gray texture and also outperforms the classical color texture retrieval method.3. The paper tried to propose a three channels color texture retrieval frame based on PCNN related models. The frame processes each color channel by using PCNN related models, and then combines the feature vectors of all channels under the condition that overall feature length will not increase. Experimental results show that the frame can effectively solve the PCNN related models based color texture retrieval problem.4. For shortest path problem, a new SAPCNN model (self-adaptive autowave PCNN) is proposed. The model can adjust the propagation speed of autowave adaptively according to the current network state, which will result in faster resolving time. In addition, SAPCNN model based finding shortest path method is applied to K shortest paths problem. Experimental results show that the method can reduce the time consumption for finding K shortest paths.更多还原