Habesha Women's Dress Embroidery Design Identification Using an Ensemble Feature Extraction Approach

Habesha women's dress is cultural clothing in most areas of Ethiopia, which is common in the Amhara region. Although the embroidery design of the Habesha women's dress appears to be the same, it has a different texture, shape, and color. This embroidery design belongs to the Awi, BahirDarZege, Sekota, SemenShewa, RayaKobo, Lalibela, Wollo, Gojjam, and Gondar Habesha women's dress cultures dressing styles. However, as fashion design on clothing grows, the cultural identity of the embroidery design is lost. There is a lot of research on clothing patterns and color detection, recognition, and classification. However, it is challenging to classify clothes with the same patterns, texture, and color into various classes due to the problem of uncontrolled environment, segmentation of ROI, color space, and feature extraction of the embroidery design of Habesha women's dress. The main aim of this study was to use an ensemble feature extraction approach to identify those patterns. We have collected 8100 Habesha women's dress embroidery images from different particular embroidery designer for different cultural clothe dress. We have classified those images into nine classes which contain 900 images per class. Then, we applied pre-processing to the labeled images. Also, we have done feature extraction with CNN, SGAN, and GLCM For classification. Each model uses SVM as a linear kernel. Finally, we have developed different ensembled models such as CNN + GLCM (accuracy: 95.80%), SGAN + GLCM (accuracy: 97.84%), CNN + SGAN (accuracy: 97.96%), and SGAN + GLCM + CNN (accuracy: 98.52%) with PCA. So, based on the model's accuracy, an ensemble of SGAN + GLCM + CNN with the SVM classifier has been selected.