Texure Image Segmentation by Using Genetic Algorithms and Neural Networks

本文研究了基于遗传算法和神经网络的高精度纹理图像分割方法，遗传算法以其对大规模、高复杂性优化问题的高效性而备受关注。本文将非平稳纹理图像分割作为优化问题，提出了一种基于遗传算法的高精度纹理分割方法。针对遗传算法的收敛问题，提出了一种利用自组织神经网络提高遗传算法收敛速度的新方法。多分辨率图像处理将原始图像分解为一组不同分辨率的图像进行图像分析，目前已广泛应用于多种图像分析方法中。在纹理图像分割方面，提出了通过构建多分辨率图像金字塔和搜索图像上、下层像素之间的父子关系来分割纹理图像。但是，该方法只利用图像的灰度特征，不能适用于纹理图案复杂的图像。本文提出了一种新的纹理图像分割方法。该方法首先将原始图像分解成大小相同的小区域，然后利用提取的特征向量构造金字塔，并利用神经网络确定特征向量之间的父子关系。通过从一系列二值图像中提取拓扑纹理特征，提出一种新的特征提取方法。

In this study, highly accurate texture image segmentation methods by using genetic algorithms and neural networks have been inspected.Genetic algorithm has attracted much attention because of its efficiency to large-scale, high-complexity optimization problems. In this study, considering segmentation of an unstationary texture image an optimization problem, we proposed a highly accurate texture segmentation method by applying genetic algorithms. To deal with the problem of convergence in genetic algorithms, a new approach was proposed to improve the convergence by using self-organized neural networks. And a novel pre-processing method was also proposed to extract texture features from highly unstationary texture images by using wavelet transformation.Multi-resolution image processing decomposes an original image into a set of images with different resolution for image analysis and has been widely by many image analysis methods up to now. As to texture image segmentation, methods have been proposed by constructing a pyramid with multi-resolution images and searching parent-son relationship between image pixels at an upper level and a lower level. However, this method can not be applied to image with complex texture patterns because it only uses image intensity features. In this study, a new texture image segmentation method was proposed. In this method, the original image is decomposed into small regions with same sizes: a pyramid is then constructed with feature vectors extracted from each region, and the parent-son relationship between feature vectors are determined using neural networks. And a new feature extraction method was proposed by extracting topological texture feature from a series of binary images.