Illumination-invariant face skin pigmentation prediction

Simple item page
dc.contributor.advisorTheart, Rensuen_ZA
dc.contributor.advisorBooysen, Thinusen_ZA
dc.contributor.authorMbatha, Success Katlegoen_ZA
dc.contributor.otherStellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.en_ZA
dc.date.accessioned2024-02-28T07:43:56Zen_ZA
dc.date.accessioned2024-04-26T16:40:10Zen_ZA
dc.date.available2024-02-28T07:43:56Zen_ZA
dc.date.available2024-04-26T16:40:10Zen_ZA
dc.date.issued2024-03en_ZA
dc.descriptionThesis (MEng)--Stellenbosch University, 2024.en_ZA
dc.description.abstractENGLISH ABSTRACT: Skin tone estimation is a critical task with a wide range of applications in fields such as cosmetic science, dermatology, image processing, and facial recognition. Accurate skin tone estimation plays a significant role in improving the inclusivity and fairness of these systems. As machine learning and artificial intelligence continue to advance and finds application in widely used systems, addressing the challenges of skin tone estimation has become increasingly important to ensure that these technologies perform consistently across diverse skin tones. This study focuses on contributing to this field of study by developing a CNN based skin tone classification/estimation model capable of delivering consistent accuracy for individuals with varying skin tones and under diverse lighting conditions. Early in 2022, Google introduced a new skin tone classification system known as the “Monk Skin Tone (MST)” which offers a broader range of skin tones compared to the commonly used “Fitzpatrick Skin Type (FST)” system. This more comprehensive scale is designed to be inclusive and representative of the diverse global population and has been adopted in this study. A data collection campaign was hosted at Stellenbosch University with an aim to develop a dataset with a diverse range of skin tones, classified into the MST, addressing the shortcomings of existing openly accessible datasets. This effort resulted in the acquisition of 21 375 images from 285 participants. Furthermore, the model development process involves exploring various CNN architectures, model configurations, and data pre-processing techniques to maximise the accuracy of skin tone estimation. Experimental findings on various model configurations are summarised as follows: The regression model trained on LAB images, which was selected as the best performing model, demonstrated the highest accuracy at 58.12%. In contrast, a pre-trained CNN model showed limited accuracy, achieving a modest 36.85%. When colour balancing techniques were applied to the dataset images, the resulting accuracy was 55.05%, falling short of the performance of the regression model using LAB images. Moreover, an attempt to increase both image and model dimensionality by converting images to RGB-LAB-HSV led to overfitting issues, resulting in an accuracy of 45.76%. Furthermore, the study explored the model’s performance under different lighting conditions, with the highest accuracy recorded under warmer lighting conditions such as “Halogen warm white” (62.07%), and “Florescent warm white” (58.37%). The study also discusses the impact of spectral characteristics on lighting conditions, particularly noting that “LED warm white” exhibited lower accuracy at 55.73%. This reported accuracy is of samples at distance ≤ 0.5 units from the actual targets. However, when the margin distance is increased to 1 and 2 units, the average accuracy across all light types becomes (85.45 }2.01) % and (97.16 }1.00) %. In the evaluation of skin tone estimation accuracy based on Monk skin tones in the LAB space, the study found that Group 1, representing individuals with lighter skin tones (Monk skin tone 1-3), exhibited strong accuracy. Group 2, encompassing individuals with middle-range skin tones (Monk skin tones 4-7), displayed comparatively lower accuracy in the skin tone estimation. Group 3, consisting of individuals with higher pigmentation skin tones, fell between the performance levels of Groups 1 and 2 in terms of accuracy. Despite the non-linear performance observed across all these skin tone groups, the overall skin tone estimation performance is satisfactory, with an average predicted-to-target error distance value of 16.40 } 20.62 in the LAB space for all samples. Overall, this research contributes to the advancement of skin tone estimation, with practical implications for enhancing the performance of facial analysis algorithms in real-world applications.en_ZA
dc.description.abstractAFRIKAANSE OPSOMMING: Geen opsomming beskikbaar.af_ZA
dc.description.versionMastersen_ZA
dc.format.extentxv, 125 pages : illustrations.en_ZA
dc.identifier.urihttps://scholar.sun.ac.za/handle/10019.1/130412en_ZA
dc.language.isoen_ZAen_ZA
dc.language.isoen_ZAen_ZA
dc.publisherStellenbosch : Stellenbosch Universityen_ZA
dc.rights.holderStellenbosch Universityen_ZA
dc.subject.lcshNeural networks (Computer science)en_ZA
dc.subject.lcshImage processing -- Digital techniquesen_ZA
dc.subject.lcshHuman face recognition (Computer science)en_ZA
dc.subject.lcshImage segmentationen_ZA
dc.subject.lcshHuman skin coloren_ZA
dc.subject.lcshUCTDen_ZA
dc.titleIllumination-invariant face skin pigmentation predictionen_ZA
dc.typeThesisen_ZA
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
mbatha_face_2024.pdf
Size:
48.58 MB
Format:
Adobe Portable Document Format
Description:
Download
Collections
Masters Degrees (Electrical and Electronic Engineering)