An Analysis of Bias in Facial Image Processing: A Review of Datasets

Facial image processing is a major research area in digital signal processing. According to recent studies, most commercial facial image processing systems are prejudiced by bias towards specific races, ethnicities, cultures, ages, and genders. In some circumstances, bias may be traced back to the algorithms employed, while in others, bias can be elicited from the insufficient representations in datasets. This study tackles bias based on insufficient representations in datasets. To tackle this, the research undertakes an exploratory review in which the context of facial image dataset is analyzed to thoroughly examine the rate of bias. Facial image processing systems are developed using widely publicly available datasets since generating datasets are costly. However, these datasets are strongly biased towards Whites and Asians, and other geo-diversity such as indigenous Africans are underrepresented. In this study, 40 large publicly accessible facial image data sets were examined. The races of the datasets used for this study were visualized using the t-distributed Stochastic Neighbor Embedding (t-SNE) visualization method. Then, to measure the geo-diversity and rate of bias of the dataset, k-means clustering, principal component analysis (PCA) and the Oriented FAST and Rotated BRIEF (ORB) feature extraction techniques were used. The findings from this study indicate that these datasets seem to exhibit an obvious ethnicity representation bias, particularly for native African facial images; as a result, additional African indigenous datasets are required to reduce the bias currently present in the most publicly available facial image datasets.