Recent advancements in machine learning have been aimed at helping the differently abled communicate with the privileged who have their natural faculties functioning properly. As with all machine learning technologies that use supervised learning for training, any potential bias must be mitigated to ensure effectiveness. For example, the field of facial recognition has identified and documented bias against dark-skinned populations due to skewed training datasets. Similar skin tone bias has been identified in medical imaging, and subsequent research has made progress on mitigating it. Accessible technology must also work to identify and mitigate similar skin tone bias. As convolutional neural networks gain prevalence in technologies that perform human interaction or detection, care must be taken to ensure the technologies are not biased due to improper training data as that would restrict equitable access to solutions and further widen the gap that exists. This paper introduces a novel approach to debiasing an image dataset, particularly one for detecting American Sign Language fingerspelling, to improve performance across different skin tones by adding an augmentation step of HSV color shifts. Testing showed this method yielded a 143% increase in accuracy when tested against a dataset containing images of a user with a different skin tone than the skin tones represented in the training dataset. This method can be used to improve real-world accuracy and debias datasets in cost-effective ways for tasks involving hand pose classification, ensuring that developments in machine learning powered accessible technology are effective for all users.
