A carregar...
Publicação
Impact evaluation of skin color, gender, and hair on the performance of eigenface, ICA, and, CNN methods
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 1.88 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Although face recognition has made remarkable progress in the past decades, it is still a challenging area. In addition to traditional flaws (such as illumination, pose, occlusion in part of face image), the low performance of the system with dark skin images and female faces raises questions that challenge transparency and accountability of the system. Recent work has suggested that available datasets are causing this issue, but little work has been done with other face recognition methods. Also, little work has been done on facial features such as hair as a key face feature in the face recognition system.
To address the gaps this thesis examines the performance of three face recognition methods (eigenface, Independent Component Analysis (ICA) and Convolution Neuron Network (CNN)) with respect to skin color changes in two different face mode “only face” and “face with hair”.
The following work is reported in this study, 1st rebuild approximate PPB dataset based on work done by “Joy Adowaa Buolamwini” in her thesis entitled “Gender shades”. 2nd new classifier tools developed, and the approximate PPB dataset classified based on new methods in 12 classes. 3rd the three methods assessed with approximate PPB dataset in two face mode.
The evaluation of the three methods revealed an interesting result. In this work, the eigenface method performs better than ICA and CNN. Moreover, the result shows a strong positive correlation between the numbers of train sets and results that it can prove the previous finding about lack of image with dark skin. More interestingly, despite the claims, the models showed a proactive behavior in female’s face identification. Despite the female group shape 21% of the population in the top two skin type groups, the result shows 44% of the top 3 recall for female groups. Also, it confirms that adding hair to images in average boosts the results by up to 9%. The work concludes with a discussion of the results and recommends the impact of classes on each other for future study
Descrição
Project Work presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Knowledge Management and Business Intelligence
Palavras-chave
ICA Eigenfaces PCA CNN ResNet152 Dark Skin Face Recognition