This repository contains the Latex files and compiled PDF of my Bacholor manuscript.
Real-time hair and clothes segmentation on mobile devices
In recent years, convolutional neural networks have been developing rapidly and solving many challenging tasks in computer vision. In fact, there are numerous possibilities of integrating on beauty and fashion industries with deep learning. One of the thesis's core works, which plays a vital role in intelligent systems for these two industries, is hair and clothes semantic segmentation. The semantic information provided by the semantic segmentation is an understanding of the targeted objects in images, and this scene understanding is vital for many other tasks, such as recommendation systems, augmented reality.
Although there are many resolutions to the semantic image segmentation in theliterature, such as DeeplabV3+ [5] and PSPNet [6], they fail to offer a low latency inferenceas to their complex architectures in aim to acquire the best accuracy. As a part of thisthesis work, we provide an efficient architecture for hair and clothes segmentation using theU-Net architecture with the MobileNetv2 base network. It delivers real-time performanceon smartphones while maintaining a high intersection over union (IoU) of 86% on average.
Based on the success of the models, the thesis proceeds to propose and develop a beauty app on the Android platform that allows users to recolor hair and clothes directly on the camera preview. It has proven to be a fully functional solution capable of providing a real-time experience while keeping adequate functionality.