Paper Title
Image Captioning and Facial Attribute Analysis
Abstract
This research paper addresses the issue of racial bias in existing public facial image datasets, which are heavily
skewed toward Caucasian faces and underrepresented other races such as Latino. To address this issue, the researchers
created a new facial image dataset that adjusted for race, with images classified into seven racial categories. Evaluations
were conducted on existing and new datasets, with the model trained on the new dataset performing significantly better and
with consistent accuracy across race and gender groups. The researchers also developed a method for finding semantically
similar images using convolutional neural network activations and selecting captions based on unigram frequency.
Additionally, they created an application for mining semantic descriptions from facial attributes to understand beauty, using
a completely data-driven approach. Experimental results showed that beauty semantics are reasonable and beneficial for
modification. Overall, this paper provides important insights into addressing racial bias in facial image datasets and
developing new methods for analyzing facial attributes.
Keywords - Racial bias, facial image datasets, Latino underrepresentation, Race classification, Convolutional Neural
Network, Semantically similar images, unigram frequency, beauty understanding