Texture is one of the most important properties of visual surface that helps in discriminating one object from another, or an object from a background. Any machine vision system that tries to compete with a human visual system must therefore be able to deal with textures. One of the mam tasks of texture analysis is the recognition of image regions using their textural properties. The work presented in this thesis aims to study a number of experimental investigations of neural networks (supervised versus unsupervised) used for the \ classification of natural-textured images (from Brodatz album), and suggest the best neural network architecture leading towards the goal of high accuracy texture-image classification. Therewithal, attempt to draw a conclusion about the best feature-set (recommended) that could be used to discriminate texture images and improve the overall performance. To do this, five different sets of features (statistical, spectral, direct features, or combination of statistical and spectral features) were used to train a number of texture-classification neural networks. Experimental results that show the classification accuracy of each texture-classifier were introduced. Then these results were analyzed and compared to arrive at the best neural network texture-classifier, and decide which of the feature-sets (among the five sets) leads to the minimum classification error rate (i.e. has the best texture-discrimination ability). From the experimental results, it was found that training a neural network texture classifier using statistical/spectral features improves the texture classification results. Further, the recommended feature-set was used in training the neural networks «£; presented in this work to classify two composite-texture-images (one at a time). Then, networks performances were tested on the entire composite-texture-image, and their classification abilities were compare visually. A Conventional (statistical) classification algorithm was trained with the same five feature-sets mentioned above to classify the natural-texture images (used in this work, comparing the classification results of neural network approaches with traditional statistical classifiers, it was found that neural network approaches allow one to perform the same.texture classification tasks more accurately. Weight initialization is one of the essential problems the designer faces when using neural net approaches to solve any problem. In most cases, weights initialized randomly. Random weight initialization may cause generated weights to be stuck in isolated regions and local minimas. In this work, a new weight initialization strategy (depending on the convex-combination method) for competitive neural nets was developed to improve the convergence time and overcome the isolation problem. The developed technique of weight initialization leads to better clustering operation. A software system called TICS (an acronym for Texture-Image Classification System) is developed to act as a texture-classifier using one of the neural network -approaches. The system is implemented on IBM PC compatible using Borland C version 2.0 with mouse toolbox support.