Colored multi-scale texture classification using morphological mathematic operations and CNN

Abstract

In recent years, digital image processing became a rapidly growing research area of computer science. Its multidiscipline uses make it a center of interest of many researchers in order to achieve better understand of the image's content. Texture is considered as one of the fundamental descriptors of the image besides color and shape. Different methods have been proposed to analyse and represent texture. Regardless their efficiency, lot of these methods are not able to extract all the features distributed in different scales of interest. In this work, we address this limitation by proposing an improved method based on morphological mathematic operations and CNN classifier to represent different scale of interest. First, we apply on the input color image a series of opening and closing morphology operations to generate different scales of interest presented as new images. The output will then be fit to different models of CNN for purpose of extracting features and classify them. Our method aims to extract both color and multi-scale information from texture images. The obtained results from KTH_TIPS dataset were about (84% in our proposal model, 85% on the VGG model, and 56.54% on the RES-net model). The results were satisfying and promising compared to methods that target only one scale of interest.