INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING, MANAGEMENT & APPLIED SCIENCE (IJLTEMAS) ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue VI, June 2025 www.ijltemas.in Page 1012 A Deep Learning Framework for Personalized Fashion Recommendations Based on Skin Tone Analysis H. P. Dissanayake 1 , J. M. Chan Sri Manukalpa 2 1 Department of Physical Science, Rajarata University of Sri Lanka 2 Department of Information Technology, SLIIT City Uni, Colombo 03, Sri Lanka DOI: https://doi.org/10.51583/IJLTEMAS.2025.1406000112 Received: 17 June 2025; Accepted: 21 June 2025; Published: 24 July 2025 Abstract—The personalization of virtual fashion recommendations remains hindered by limited integration of chromatic and anthropometric factors, especially skin tone compatibility. This study addresses a critical research gap by proposing a voice-enabled 3D fashion recommendation system that incorporates deep learning-based skin tone classification and adaptive garment suggestion. The primary objective is to enhance aesthetic compatibility and user satisfaction through real-time, personalized recommendations. Utilizing a custom-designed Deep Convolutional Neural Network (DCNN) and reinforcement learning algorithms, the system classifies user skin tones with 89.14% accuracy and adapts recommendations based on user feedback, reducing outfit resets by 54%. A curated dataset encompassing five skin tone categories and a multi-stage image preprocessing pipeline ensures inclusive and robust performance. The results demonstrate significant improvements in recommendation relevance and user engagement, with 88% satisfaction and 93.7% dominant tone detection accuracy. These findings underscore the system's potential to set new benchmarks in personalized fashion retail while promoting inclusivity and sustainable consumer practices. Keywords—Skin Tone Classification, Personalized Fashion Recommendation, Reinforcement Learning, 3D Virtual Try-on, Deep Convolutional Neural Network (DCNN) I. Introduction The digital transformation of fashion retail has precipitated a paradigm shift in consumer engagement strategies, with virtual try- on (VTO) systems emerging as critical interfaces between e-commerce platforms and end-users [4]. While contemporary recommendation systems have achieved notable success in garment size prediction and body shape accommodation [3,7], a persistent research gap exists in addressing chromatic compatibility between apparel and skin tone - a fundamental determinant of aesthetic harmony and personal style satisfaction [9]. This oversight persists despite empirical evidence demonstrating that color coordination accounts for approximately 38% of purchase decisions in online fashion retail [5]. Current implementations of virtual fitting technologies predominantly focus on geometric fidelity in garment simulation [1,6] or body measurement accuracy [2], while relegating skin tone analysis to secondary consideration. As Huang and Huang [1] demonstrate in their CLO3D-based down jacket simulation, even advanced 3D garment visualization systems often employ standardized skin textures, fundamentally limiting their personalization capabilities. Similarly, Sekine et al. [2] achieved breakthroughs in single-shot body estimation but acknowledged the absence of skin tone variables in their fitting algorithm as a key limitation. The academic literature reveals three critical shortcomings in existing approaches: First, an over-reliance on supervised learning paradigms that require extensive manual annotation of skin tone categories [10]. Second, inadequate representation of diverse demographic groups in training datasets, as noted by Wazarkar et al. [3] in their analysis of body type classification systems. Third, the absence of dynamic feedback mechanisms to iteratively refine color recommendations based on user preferences [8], a gap particularly evident in smart mirror technologies [10]. This research addresses these limitations through three principal innovations: (1) development of an automated skin tone classification system using hybrid computer vision and deep learning techniques, (2) implementation of a reinforcement learning framework for continuous recommendation refinement, and (3) creation of an ethically curated dataset encompassing the full spectrum of human skin tones, validated against the Fitzpatrick scale. Our methodology builds upon established work in deformable 3D shape matching [6] and neural body fitting [7], while introducing novel adaptations for chromatic analysis. The system architecture incorporates advancements from knowledge-based recommendation systems [5] and human-fashion interaction paradigms [8], synthesizing these elements into a unified framework for skin tone-aware fashion personalization. The practical implications of this research are substantial for both academic and commercial domains. As Werdayani and Widiaty [11] emphasize, next-generation virtual fitting rooms must transcend basic size prediction to address holistic style preferences. Similarly, Sridevi et al. [9] demonstrate that image-based neural networks achieve superior engagement metrics when incorporating chromatic compatibility factors. Our findings corroborate these observations while providing quantifiable performance benchmarks: an 89.14% classification accuracy across five skin tone categories and a 54% reduction in outfit reset instances.