Pre-processing strategies for skin detection using MLP
Skin detection is an important preliminary step in a wide range of image processing applications such as face detection, person identification, gesture analysis and access control. Several techniques have been used for skin detection. In this thesis, the multilayer perceptron (MLP) neural network an...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2011
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/39139/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/39139/2/FULLTEXT.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Skin detection is an important preliminary step in a wide range of image processing applications such as face detection, person identification, gesture analysis and access control. Several techniques have been used for skin detection. In this thesis, the multilayer perceptron (MLP) neural network and histogram thresholding techniques were used. Recent studies have shown that combining skin features and/or skin classifiers can further improve the performance of the skin detection system. Thus, the main objective of this research is to evaluate the effect of several combination strategies on the performance of a skin detection system based on the MLP. To achieve this goal, first the histogram thresholding technique was used to select skin features (chrominance component in a given colour space) that give the highest correct skin detection. These features will be used as inputs to the MLP classifiers. A modified Growing algorithm for finding the number of neurons in the hidden layer of a neural network was also developed it was able to reduce the computational time compared to the conventional Growing algorithm. The combination strategies were done by combining the skin features as well as the skin classifiers. Three skin features (chrominance component from the selected colour space) that gave the highest correct skin detection on a single input MLP classifier were used for these strategies. The strategy of combining skin features or inputs was done using two and three skin features. For combining skin classifiers strategy, several combining rules such as binary operators AND and OR were used to combine two and three classifiers, while combining rules namely Voting, Sum of Weights and New Neural Network were used to combine three classifiers. The Sum of Weights and New Neural Network were the proposed combining rules in this thesis. In order to evaluate the performances of the skin detection systems, the images from Compaq database were used. The strategy of combining two skin features Cb/Cr gave the best performance for combining skin feature strategy with 3.01% more correct detection compared with the best performance given by a single input MLP classifier given by Cb-Cr. The strategy of combining three classifiers using the Sum of Weights gave the best performance for its combining strategy with an improvement of 4.38% more correct detection compared to the best single input MLP classifier given by Cb-Cr. The Sum of Weights strategy also gave 1.37% more correct detection than the best combining skin feature strategy. The other proposed combining strategy called New Neural Network has managed to achieve 82.21% of correct detection. The best performance results obtained in this thesis were considerably good considering the unconstrained nature of the images from the Compaq database. |
|---|