Classification of Oromo cultural apparels using machine learning techniques

Abstract

Like the other nationalities and ethnic groups in Ethiopia, the Oromo have their own distinctive cultural attire. In Oromia, there are many different kinds of ethnic attire, each with its own style, that is worn for formal and ceremonial occasions. The identification and examination of Oromo cultural cloth are reliant on human expertise. Even for someone who was born and raised in Oromia, it might be challenging to tell from a piece of clothing's appearance where it came from. We propose a machine learning method for identifying and classifying the Oromo cultural clothes to which culture in Oromia they may belong. This is due to Oromia's diverse nature and culture. Various researchers have done the identification of clothing, whether it is a dress, t shirt, skirt, trousers, shoes, jacket, jeans, and the like. A dataset of 5083 Oromo Cultural cloth images is collected from different designers shop for our models which of 4,066(80%) images for training and 1,017(20%) images are for testing purpose. The proposed system has three main components: preprocessing, feature extraction and classification. In the preprocessing stage we applying augmentation, converting the color image into Grayscale, and removal of background image. in feature extraction, we apply GLCM and CNN on the image dataset of Oromo cultural Cloth fabrics. It is used to select the important features that account for the identification of Oromo Cultural Cloth Fabrics. Convolutional Neural Networks (CNN) and CNN+LSTM from deep learning, Support Vector Machines (SVM) from classical machine learning.We have also used VGG16 and VGG19 from pretrained deep learning for classification purpose. All experiments are carried out using Python 3.7, Anaconda Jupiter, and Google Collab. and sampled using samples from many Oromo traditional cloth designers’ shops located in Finfinnee and Ambo and also from different Facebook pages. Our proposed CNN and CNN+LSTM models had testing accuracy of 93.2% and 93.6%, respectively. Using the pretrained CNN model, VGG16 and VGG 19, the SVM model with each of the feature extractors (GLCM and CNN) obtained accuracy of 93.3% and 93.3%, 55%,85% respectively.