Abstract:
"Colour consulting, or personal colour analysis, is a vital tool for assisting people in choosing colours that enhance their natural skin, hair, and eye colours. Finding one's undertone and recommending a distinctive colour palette that improves one's appearance are the goals of personal colour analysis. This process ultimately increases self-assurance and makes choosing clothes easier. Unfortunately, users must rely on limited accuracy while navigating DIY resources because existing web tools frequently do not adequately accommodate users with varied skin tones and cultural backgrounds.
This project uses deep learning techniques to tackle this problem. It integrates an image classification model for individual colour diagnosis with a face segmentation model utilising FaRL (Facial Representation Learning) for accurate facial feature segmentation. The Image Classification Model classifies facial images into predefined personal colour categories, making it easier to recommend appropriate colour palettes based on individual features.
The face parsing model achieves an average F1 score of 0.94, and the face detection module has an average accuracy of 84% in preliminary testing on the WIDER FACE dataset, which shows encouraging results. Furthermore, the accuracy rate of the personal colour analysis module is about 70%. These first results highlight the effectiveness of the suggested methodology in correctly identifying individual colour kinds and segmenting facial features, providing a strong basis for further improvement and optimisation. These preliminary findings clearly show that the personal colour analysis technique described here is among the most useful and efficient options out there."