节点文献
基函数神经网络图像复原算法研究
【作者】 田茂新；
【导师】 田启川；
【作者基本信息】 太原科技大学 ， 模式识别与智能系统， 2012， 硕士
【摘要】 图像复原技术是图像处理领域重要的研究课题之一。在天文遥感军事、道路交通及侦破、医学影像等领域具有重要的现实意义和应用价值。图像的复原技术属于图像研究的预处理部分，也是最基础的环节，复原结果的好坏对数字图像的后续研究工作有着重要的地位。然而在现实中，获取图像时，会由于各种因素造成图像的退化。因此如何复原出退化的数字图像是近年来数字图像处理领域研究的热点问题。由于现阶段造成图像退化的因素复杂多样，图像的退化函数难以准确地确定，论文中提出了一种利用基函数神经网络的图像复原算法，它是以一组正交多项式函数作为隐层神经元的激励函数，根据误差传递算法（BP算法）对权值进行修正，达到收敛目标。由于神经网络的拓扑结构对网络的性能影响极大，文中利用了衍生算法确定神经网络在给定误差下得出处于最佳性能时的隐层神经元的数目，针对传统的神经网络学习需要进行反复的迭代，文中寻找出在基函数神经网络的迭代次数与网络权值的关系，提出了权值直接确定的图像复原算法，相对于传统算法，具有较好的效果。针对局部退化图像的问题，论文中提出了一种利用图像的频谱特征进行图像退化区域提取的方法，与传统方法相比，该方法不受背景区域的影响，能较好的完成图像完全退化区域的提取。
【Abstract】 Image restoration is one of the important research topics in the field ofimage processing. It has important practical significance and application valuein military remote sensing, Road traffic, detection and medical imaging and soon. The technology of image restoration technology belongs to imagepreprocessing, but the fundamental part. The results of the restoration of digitalimages play an important role in the further research. However, it would bedegraded when we obtain the image in reality. So how to recover the degradedimages is becoming a hot issue in recent years.Image degradation is caused by complex and diverse factors, and the pointspread function is difficult to determine. According to the fact that point spreadfunction of the degraded image can’t obtain accurately, basis function neuralnetwork for image restoration was constructed based on the Orthogonalpolynomial basis functions in this paper, The hidden-layer neurons are activatedby a series of orthogonal functions, update its weights by the errorBack-propagation training algorithm and finally reached convergence target. Asthe topology of the neural network take a great impact on its performance, thispaper determine the number of the hidden layer neurons which under the bestperformance in a given error using hidden-neuron growing algorithm. To avoidlengthy BP-training of the iterative weights-up dating, this paper found therelationship between the number of iterative training and the weights,weights-direct-determination algorithm for image restoration was proposed,Compared with traditional algorithms, it is has good results.Finally, A new method for region extraction of partial blurred image ispresented based on the Fourier spectrum amplitudes, this method can avoid theaffect of background area, realize the extract blur region well compare to thetraditional methods.