Design and development of skin detection model based deep learning on different skin tones
The aim of this study is to design and develop a systematic dataset for multiple skin tones and toanalyse the reasons behind misclassifications of skin and non-skin, using different deep learningmodels, colour spaces, and different optimisation parameters. Related academic literature havecited three...
Saved in:
| Main Author: | |
|---|---|
| Format: | thesis |
| Language: | eng |
| Published: |
2021
|
| Subjects: | |
| Online Access: | https://ir.upsi.edu.my/detailsg.php?det=5828 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The aim of this study is to design and develop a systematic dataset for multiple skin tones and toanalyse the reasons behind misclassifications of skin and non-skin, using different deep learningmodels, colour spaces, and different optimisation parameters. Related academic literature havecited three problems, namely data-related issues (e.g. skin-like), data volume (e.g. large volumerequires high computer source), and technical issues (e.g. optimising parameters). Twoarticles on Deep Learning (DL) for skin detection failed to address the issues extensively.DL foundation is a training dataset and the quality of training depends on the quality ofthe data-input. To address the issues, a systematic dataset consisting of 17 million patcheswas created for multiple skin tones with (skin-like) images. The dataset was then converted intodifferent colour spaces with multiple labels that characterise different scenarios, runningdifferent DL. Experimentally utilised YCbCr and CNN present high performance of binaryand multi-class classifications. Binary classification of skin and skin-like resulted in 98% andmulti-class classification of four classes 84% and 69% for five classes respectively. Furthermore,a binary classification between skin tone and skin-like (e.g. black skin tone and blackskin-like) resulted in 97%, 81%, 60%, and 51% for black, brown, medium, and fairconsequently. From empirical experiment, darker skin tone is a better classification accuracyfollowed by optimising parameters (Hidden-Layers, Neurons, Activations-Functions, Optimiser,Initialiser, Data-Input, and Data-Size). A hybrid CNN-RNN benchmark improves the accuracy by99% compared with 98%, and 97% for SAE compared to 91% as reported. By studying different skinscenarios, one can analyse the reasons behind overlapping between skin and skin-tones.This is a promising study for further research by developing and applying a generalised versionof skin detector with different applications. |
|---|