A study on Texture Segmentation Method by Using Neural Networks

由多种纹理区域组成的纹理图像的分割方法可以分为两个问题。它们分别是从随机纹理场中精确提取特征和将图像分割成均匀纹理区域。在纹理图像的分割中，许多方法都需要确定判别函数及其阈值来判断纹理的均匀性。但在理论上很难建立函数式和确定阈值。因此，本文以发展一种不需要判别函数和判别阈值的基于神经网络的纹理分割方法为目标，提出了将Kohonen的自组织神经网络、基于反向传播的神经网络和基于决策的神经网络（DBNN）应用于二维AR模型和分形维数提取的纹理特征的纹理图像分割方法。实验结果表明，Kohonen自组织神经网络和DBNN的分割效果最好，但由于实际纹理图像具有严格的非平稳性，因此将该方法应用于此类图像的分割效果并不理想。本文提出了一种利用小波变换对原始图像进行预处理的新方法。基于此，本文提出了一种新的纹理分割方法，并将其应用于彩色纹理图像的分割，通过对单色纹理图像的特征提取和多种颜色特征的综合，得到了一种通用性强、分割能力强的纹理图像分割方法，并将研究成果发表在期刊上。

A segmentation method of a texture image composed of several kinds of texture areas can be divided into two problems. They are to extract the features from random texture field exactly and to segment an image to homogeneous texture areas. In order to segment the texture image it was necessary to get a discriminant function and its threshold value for the decision of homogeneity of textures in many methods developed until now. But it is difficult to formulate the function and to decide the threshold value theoretically. So I aimed at the development of texture segmentation method which did not need a discriminant function and its threshold value by using neural networks.In this research I proposed the segmentation method of a texture image by applying Kohonen's self-organizing neural network, neural network based on backpropagation method, and decision based neural network (DBNN) to the texture features extracted by two-dimensional AR model and fractal dimension. It was obviously that the best segmentation result was obtained by using Kohonen's self-organizing neural network and DBNN.As the actual texture image has strictly non-stationarity, good segmentation result was not obtained by applying this method to such an image. So I propose a new method which pre-processes a original image by using wavelet transform.Until now it is difficult to obtain the optimum segmentation number contained in a texture image automatically. So I proposed a new method to solve the problem.Furthermore, it was obviously that the proposed method could apply to the segmentation problem of a color texture image by using the features for monochrome texture image and some kinds of color features.From the above-mentioned researches I think that a texture segmentation method with generality and good segmentation ability is completed.I will contribute the results of this research to jounal soon.