Multi-Task Learning with Soft Parameter Sharing Using MobileNetV2 for Body Type and Skin Tone Classification for Recommend Clothing

Dorai Raj, Nayana (2025) Multi-Task Learning with Soft Parameter Sharing Using MobileNetV2 for Body Type and Skin Tone Classification for Recommend Clothing. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (6MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

The proposed paper presents body-type and skin tone classification from a female full-body image with dual-phase computer vision system that combines unsupervised representation learning and supervised multi-task classification with soft sharing parameter to allow automated fashion recommendation that is tailored to each user. In the initial phase there is an unsupervised pipeline that carries out body-type discovery and skin-tone clustering based on the use of feature extraction, PCA dimensionality reduction and K-Means clustering with a refinement step made through K-Nearest Neighbour smoothing. Estimating skin-tone uses the luminance-invariant YCrCb chrominance statistics, region-of-interest sampling, and body-type clustering uses posenormalised geometric descriptors to cluster shapes to form a consistent phenotype group, with no labelled supervision. On the second stage the labels of unsupervised learning is passed to the next step, a multi-task learning (MTL) neural architecture simultaneously predicts a body type and dermatologically appropriate skin-tone classes through late fusion of body spatial features and facial skin patches, sharing soft-parameters to optimise cross-task generalisation. Results of learning streams are combined into a silhouette-colour recommendation system by which morphological and chromatic features are associated with proven styling guidelines to women fashion. Experimental findings suggest that involving unsupervised clustering will give accurate results by avoiding manual labelling to achieve representation coherence, and supervised MTL to achieve discriminative accuracy, results in stronger phenotype estimation, fairness across under-represented tones, and consistency in automated outfit generation than individually run single-task and supervised baseline methods do. The accuracy obtained for skin tone and body type are 62.96% and 81.48% respectively with Multi-task learning.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Fajemisin, Ade UNSPECIFIED
Uncontrolled Keywords:	Unsupervised Learning; Supervised Learning; KNN; K-Means; MobileNetV2; Multi-task Learning with Soft Parameter Sharing; PCA dimensionality reduction; YCrCb chrominance; RGB; Sampling; Balanced sampling; Joint-balanced sampling; Fitrizpatrick Classification; Unsplash API; Gradio interface; Deep Learning; CNN
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Computer vision Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Computer vision H Social Sciences > HD Industries. Land use. Labor > Specific Industries > Fashion Industry
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	02 Sep 2026 09:02
Last Modified:	02 Sep 2026 09:02
URI:	https://norma.ncirl.ie/id/eprint/9753

Actions (login required)

View Item