HAIR COLOR SIMULATION USING A HAIR COLOR CLASSIFICATION GUIDED NETWORK

Smart overview of the Invention

A novel approach is introduced for hair color simulation utilizing a generative model guided by a hair classifier model. This method is particularly useful in virtual try-on (VTO) applications, allowing users to virtually try on different hair colors. The system incorporates a color mapping network that processes an input image and a target hair color to generate an output image that simulates the desired hair color.

Technological Context

The invention addresses challenges in the realm of e-commerce, where virtual try-on technology enhances customer experience by allowing them to preview products. Specifically, in the context of hair color, the technology processes user-supplied images to classify and simulate hair colors. Traditional methods face difficulties due to biases in human-labeled data and the cost of obtaining such data, necessitating improved models for accurate classification.

Key Features

The system employs a generative neural network trained under the guidance of a hair classification network. This approach helps mitigate human bias by separating real label distribution from annotator biases using confusion matrices. The model leverages semi-supervised learning frameworks to enhance performance with limited labeled data, achieving classification accuracy significantly higher than human annotators.

Implementation Details

The VTO pipeline integrates several components, including a color refinement neural network and a color mapping network. These components work together to produce output images that blend input images with target hair colors. The hair classification network comprises shade and reflectance classifiers trained using cross-entropy losses derived from expert-labeled training images.

Training Methodology

Training involves using expert votes to create target labels for shade and reflectance, employing techniques like mean teacher frameworks and annotator confusion matrices. This ensures that the classifier outputs align closely with expert opinions. The system's architecture allows for robust training even with limited labeled data, ensuring the generative model's effectiveness in simulating realistic hair colors.