A Multitask Learning Approach for Habesha Fashion Cloth Recognition and Classification

This paper presents a multi task deep learning method for jointly recognizing Habesha fashion clothing and clas-sifying fabric types. This work represents a significant step for-ward in automating the identification of pure fabrics, particularly for potential use in e-commerce, in response to the increasing demand for Habesha fashion clothing during formal events. The proposed multitask model consists of two sub-networks that are trained to collaborate and work together to recognize fashion type within three Habesha clothing categories (kemis, T-shirt, and Trouser) and classify fabrics into five types (Abujede, Magg, Fasha, Mennen, and Shash). We conducted experiments using a dataset of 2,322 images from Habesha fashion cloth shops and designers in Bahir Dar, Gondar, and Addis Ababa to evaluate our models. In the jointly trained model, we achieved 92.4 % accuracy for fashion type recognition and 72.8 % for fabric type classification. However, when trained separately, the model's fabric type performance significantly dropped to 64.5 %, while fashion type recognition remained high at 94.6 %. These results show that the model performs well in distinguishing between different fashion types when trained separately. Joint learning substantially improves fabric type classification by approximately 8 %, though there is a slight 2 % decrease in fashion type recognition performance.