Abstract
Texture classification problem (e.g. material surface images) receives great attention from the scientific community with many conducted studies. The four main approaches of these studies include local statistics, developing models based on Texton analysis (i.e. repeating structures in textures), geometric methods using features such as SIFT, HOG, and signal processing using filters. However, it is not easy to determine which algorithm or model is best suited for classifying material surface images. The reason is that texture recognition poses many challenges and is prone to errors because it depends on the physical factors of the material surface image acquisition process such as wide or narrow angle, low or high light intensity, large or small scale, etc. The above problems affect the accuracy of the recognition and classification process.
The thesis proposes a multi-feature combination between the features extracted by Weber's Law Descriptor and ones from Local Binary Patterns. This integration helps us to obtain complete information about the image texture, independent of the rotation, intensity and scale of the input image. Therefore, it helps to improve the accuracy of material surface image classification.
The comparison of experimental results of the proposed model with existing models is carried out on the practical problem of detecting fake login credentials. The experimental results show the feasibility of the proposed model on two datasets: KTH-TIPS 2a and Brodazt. The mentioned integrated method provides high classification accuracy and can be applied to various image texture datasets.