基于神经网络的图像复原方法研究

发布时间：2018-08-24 07:32

【摘要】： 图像复原是数字图像处理领域中最重要、最基本的研究课题之一,具有重要的理论价值和实际意义。其目的就是要尽量的恢复被退化图像的本来面目。传统的图像复原方法,如逆滤波法、维纳滤波法、卡尔曼滤波法、奇异值分解伪逆法、最大熵复原法等,或面临着高维方程的计算问题,或要求恢复过程满足广义平稳过程的假设,这些均是使得图像复原问题没有广泛应用的根本原因。神经网络由于其固有的自学习、自适应性、强鲁棒性及并行处理方面的潜能,因此被用于解决图像处理领域内多种问题,用神经网络进行退化图像的复原便是其中的应用之一。本文在研究了基于Hopfield神经网络的图像复原算法的基础上,为克服Hopfield神经网络易于陷入局部极小值的缺点,进一步提高复原图像的信噪比和视觉效果,研究了一种在Hopfield神经网络模型中引入暂态混沌机制和小波理论用于图像复原算法的新方法,实验证明了改进方法的有效性。本文的主要工作包括: 1.讨论了一种基于Hopfield神经网络的图像复原算法。在分析网络更新规则过程中,为降低时间和空间复杂度,将原神经网络复原算法中用每一个像元恢复电平值由对应的M个神经元状态之和表示改进为用神经元状态变量群加权方案来表示。使得在保证网络良好容错性的同时,减少神经网络的整体规模;为保证网络能够精确地收敛到全局最小,采用状态连续变化的状态变量替代原阶跃取值的状态变量。 2.讨论了一种基于混沌Hopfield神经网络的图像复原算法。为改善Hopfield神经网络易于陷入局部极小值的缺点,在Hopfield神经网络模型中引入混沌机制,可以获得比Hopfield神经网络更加丰富和更加灵活的动力学特性,从而具有更强的搜索全局最优解或近似全局最优解的能力,较大程度提高了收敛性能和初值鲁棒性。 3.讨论了一种基于小波混沌神经网络的图像复原算法。原网络模型的激励函数采用单调递增的Sigmoid函数,其逼近函数的能力没有基函数强,并且Sigmoid函数在逼近过程中会产生冗余。因此将小波理论引入混沌Hopfield神经网络中,构造由小波函数和Sigmoid函数组成的新的激励函数,使得网络具有更强的函数逼近能力。
[Abstract]:Image restoration is one of the most important and basic research topics in the field of digital image processing, which has important theoretical and practical significance. The aim is to restore the degraded image as much as possible. Traditional image restoration methods, such as inverse filtering, Wiener filtering, Kalman filtering, singular value decomposition pseudo-inverse, maximum entropy restoration, etc. The restoration process is required to satisfy the assumption of generalized stationary process, which is the fundamental reason that the image restoration problem is not widely used. Because of its inherent self-learning, self-adaptability, strong robustness and potential in parallel processing, neural networks are used to solve many problems in the field of image processing, and the restoration of degraded images using neural networks is one of the applications. In this paper, based on the research of image restoration algorithm based on Hopfield neural network, in order to overcome the shortcoming that Hopfield neural network is easy to fall into local minimum, and further improve the SNR and visual effect of reconstructed image. A new method for image restoration based on transient chaotic mechanism and wavelet theory in Hopfield neural network model is studied. The experimental results show that the improved method is effective. The main work of this paper includes: 1. An image restoration algorithm based on Hopfield neural network is discussed. In order to reduce the time and space complexity, The restoration level of each pixel in the original neural network restoration algorithm is improved from the sum representation of the corresponding M neuron states to the weighted scheme of neuron state variables group. In order to ensure that the network can converge to the global minimum accurately, the overall scale of the neural network can be reduced while ensuring the good fault-tolerance of the network. The state variable with continuous state change is used to replace the state variable of the original step value. 2. An image restoration algorithm based on chaotic Hopfield neural network is discussed. In order to improve the disadvantage that Hopfield neural network is prone to fall into local minima, chaotic mechanism can be introduced into the Hopfield neural network model to obtain more abundant and flexible dynamic characteristics than Hopfield neural network. Therefore, it has stronger ability to search global optimal solution or approximate global optimal solution, and improves convergence performance and initial value robustness to a great extent. An image restoration algorithm based on wavelet chaotic neural network is discussed. The excitation function of the original network model adopts the monotone increasing Sigmoid function, and its ability to approximate the function is not as strong as the basis function, and the Sigmoid function will produce redundancy in the process of approximation. Therefore, the wavelet theory is introduced into chaotic Hopfield neural network, and a new excitation function is constructed, which is composed of wavelet function and Sigmoid function, which makes the network have stronger function approximation ability.
【学位授予单位】：江苏科技大学
【学位级别】：硕士
【学位授予年份】：2010
【分类号】：TP391.41

【参考文献】