PCNN在多尺度图像融合中的应用研究

发布时间：2018-07-01 15:00

本文选题：脉冲耦合神经网络 + 图像融合　；参考：《中国矿业大学》2017年硕士论文

【摘要】：脉冲耦合神经网络(Pulse Coupled Neural Network,PCNN),作为第三代人工神经网络新模型,被成功应用到图像处理各领域。脉冲耦合神经网络模拟了猫等哺乳动物视觉皮层视觉神经细胞活动,利用神经元的线性相加及非线性相乘调制耦合两种特性,并将生物学传输的时延特性和指数衰减特性、动物视觉神经系统相邻神经元同步激发产生震荡特性以及神经元处于抑制状态时内部活动的平衡态考虑在内,这样该模型就更符合真实的生物神经网络。脉冲耦合神经网络是单层模型神经网络,不需要训练过程即可进行特征提取、图像分割、图像融合、模式识别等,因此非常适合应用到数字图像处理中。本文主要针对脉冲耦合神经网络在图像融合领域应用的算法进行探讨与改进。研究了遗传算法优化的PCNN应用在非下采样Contourlet变换(Nonsubsampled Contourlet Transform,NSCT)的图像融合。脉冲耦合神经网络在图像融合中运用广泛,但是模型中仍存在较多参数需要凭借人工经验设置。模型参数的设置十分关键,它直接影响融合结果的好坏,因此本文对PCNN模型参数的设置方法进行研究,提出了基于遗传算法优化的PCNN改进模型结合非下采样Contourlet变换应用在图像融合中的新方法。该方法实现了对模型参数进行自适应设置,减少了需要凭借人工经验进行设定参数的个数,也避免了在参数选择过程中的盲目性。通过仿真实验,验证了本优化算法的可行性。研究了图像梯度激励的NSCT-PCNN图像融合。对PCNN模型进行改进操作,在PCNN中引入图像梯度的强度及相位相关性的加权乘积作为模型的反馈输入,克服了原始模型中仅利用图像像素灰度值作为输入而造成未考虑人眼主观视觉对图像局部因素敏感的不足,使脉冲耦合神经网络与非下采样Contourlet变换相结合的融合效果更佳。研究了PCNN应用在二维经验模态分解(Bidimensional Empirical Mode Decomposition,BEMD)的图像融合。鉴于PCNN在多尺度图像分解工具非下采样Contourlet变换中的成功应用,本文将该神经网络模型引入到二维经验模态分解中,提出一种将脉冲耦合神经网络与图像压缩感知相结合,运用在图像二维经验模态分解中的多尺度图像融合新方法。首先,BEMD将待处理原始图像分解成多个二维内蕴模式函数(Bidimensional Intrinsic Mode Function,BIMFs)和一个趋势图像。然后对BIMFs分别进行压缩测量,对所得的压缩测量系数进行PCNN图像融合后得到测量采样BIMFs,再对测量采样BIMFs进行重构得到最终的BIMFs系数以及对趋势图像基于图像熵权重的融合得到最终的趋势图像系数;最后通过BEMD逆变换可得融合结果图像。通过仿真实验验证,PCNN与二维经验模态分解结合的图像融合同样有较佳的效果。
[Abstract]:Pulse coupled neural network (PCNN), as a new model of the third generation artificial neural network, has been successfully applied to various fields of image processing. Impulsive coupled neural network (PNN) is used to simulate the visual nerve cell activity in the visual cortex of mammals such as cats. The linear addition and nonlinear multiplication modulation of neurons are used to couple the neural cells, and the delay characteristics and exponential attenuation characteristics of biological transmission are analyzed. The concussion characteristics of adjacent neurons in the animal visual nervous system and the equilibrium of the internal activities of the neurons in the inhibitory state are taken into account so that the model is more in line with the real biological neural network. Pulse coupled neural network is a single-layer model neural network, which can be used for feature extraction, image segmentation, image fusion and pattern recognition without training, so it is very suitable for digital image processing. In this paper, the algorithm of pulse coupled neural network applied in image fusion is discussed and improved. The application of PCNN optimized by genetic algorithm in image fusion of nonsubsampled Contourlet transform (NSCT) is studied. Pulse coupled neural network is widely used in image fusion, but there are still many parameters in the model that need to be set by artificial experience. The setting of model parameters is very important, which directly affects the quality of fusion results. Therefore, this paper studies the setting method of PCNN model parameters. A new method for image fusion based on genetic algorithm (GA) based on improved PCNN model and non-downsampling Contourlet transform is proposed. This method realizes the adaptive setting of model parameters, reduces the number of parameters that need to be set by human experience, and avoids blindness in the process of parameter selection. The feasibility of the algorithm is verified by simulation experiments. The image fusion of NSCT-PCNN with image gradient excitation is studied. The PCNN model is improved and the weighted product of the intensity and phase correlation of the image gradient is introduced into PCNN as the feedback input of the model. It overcomes the shortcoming that only the gray value of image pixels is used as input in the original model, which does not consider the sensitivity of human subjective vision to local factors of the image, and makes the fusion effect of pulse coupled neural network and non-downsampling Contourlet transform better. The image fusion of Bidimensional empirical Mode decomposition (BEMD) based on PCNN is studied. In view of the successful application of PCNN in the non-downsampling Contourlet transform of multi-scale image decomposition tool, this paper introduces the neural network model into two-dimensional empirical mode decomposition, and proposes a new method which combines pulse coupled neural network with image compression perception. A new multi-scale image fusion method in image two-dimensional empirical mode decomposition (EMD) is proposed. First, BEMD decomposes the original image into two dimensional intrinsic mode BIMFs and a trend image. Then the compression measurements of BIMFs are carried out, After the compression measurement coefficients are fused with PCNN images, the BIMFs are obtained, the final BIMFs coefficients are reconstructed from the measured samples, and the trend image coefficients are obtained by fusion based on the entropy weight of the trend images. Finally, the fused image can be obtained by BEMD inverse transform. The simulation results show that the image fusion based on PCNN and two-dimensional empirical mode decomposition is also effective.
【学位授予单位】：中国矿业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】