Classification is one of the most important topics in image processing and computer vision. Proper classification of an image requires selection of suitable features that discriminate different images. In addition, ideal features must be invariant to factors such as rotation, scaling, brightness and noise. The main features that are used in image processing are those that reflect color, shape and texture of an image. Texture is an important feature among many types of images. Different images ranging from remote sensing images to microscopy images, all contain textures and therefore texture classification is used in many applications including automated iection, image retrieval and medical image analysis. So far, many methods have been proposed to describe texture, including statistical methods, structural methods, methods based on filters, and methods based on models. In this thesis, we propose a texture descriptor based on more local features compared to the conventional methods. The proposed method is also rotation invariant. The proposed descriptor vector has a higher resistance against noise when compared to other methods based on local features. The proposed descriptor is named Local Patch Difference Vector and is based on the Euclidean distance of circular neighborhoods values around each pixel. The classification results of the proposed descriptor is comparable to other conventional methods and also when noise is applied to images, classification rate using the proposed method leads to significantly higher results when compared to conventional methods. In Outex databases when applying Gaussian noise with variance of 0.015, the classification results have been improved by an average 4.38% comparing to conventional methods based on local features. Keywords: Computer Vision, Image Processing, Classification, Texture, Descriptor.