DeepLumina: A Method Based on Deep Features and Luminance Information for Color Texture Classification

Abstract

Color texture classification is a significant computer vision task to identify and categorize textures that we often observe in natural visual scenes in the real world. Without color and texture, it remains a tedious task to identify and recognize objects in nature. Deep architectures proved to be a better method for recognizing the challenging patterns from texture images. This paper proposes a method, DeepLumina, that uses features from the deep architectures and luminance information with RGB color space for efficient color texture classification. This technique captures convolutional neural network features from the ResNet101 pretrained models and uses luminance information from the luminance (Y) channel of the YIQ color model and performs classification with a support vector machine (SVM). This approach works in the RGB-luminance color domain, exploring the effectiveness of applying luminance information along with the RGB color space. Experimental investigation and analysis during the study show that the proposed method, DeepLumina, got an accuracy of 90.15% for the Flickr Material Dataset (FMD) and 73.63% for the Describable Textures dataset (DTD), which is highly promising. Comparative analysis with other color spaces and pretrained CNN-FC models are also conducted, which throws light into the significance of the work. The method also proved the computational simplicity and obtained results in lesser computation time.