A Robust Approach for Categorizing and Fine-Grain Classification of Indian Ethnic Wear

There is a growing popularity of Indian ethnic wear as more Indians embrace their culture and look inwards for inspiration. Most fashion forecasting agencies worldwide use deep learning techniques to analyse millions of images to understand various product categories and their attributes to identify changing trends. The existing work efficiently understands the attributes of Western wear, such as T-shirts, pants, trousers, skirts, and other accessories. However, they are seldom correct when it comes to Indian ethnic wear, as the categories have very different structures in regard to drapes, silhouettes, and form. The Indian fashion market is growing exponentially, and understanding the changing trends will be highly viable for the industries. Unlike most previous attempts, this work intends to streamline an in-depth classification of Indian ethnic categories and further performs a fine-grained classification for a category (kurta). We seek to predict the attributes like hem type, sleeve length, hem length, and colour. The proposed approach comprises of following steps: (a) Classification of Indian ethnic wear categories using EfficientNet, (b) Localization of attributes in kurta using instance segmentation technique, (c) Detection of human key points using a pre-trained R-CNN, and (d) Classification of the kurta attributes using a combined mathematical and deep neural network approaches. The experimental results demonstrate that the proposed approach is highly effective and classifies the categories with an accuracy of 92%, further predicting the kurta attributes with an average precision ( \(\text {AP}_{50}\) ) of 83%.