A Ternary Directional Code for Gender Classification
Mario Leonel Quiroa Pimentel
- 원문 URL
Gender classification or recognition for humans may seen one of the most simple things to do. For computers, it is only a binary decision, consequently, it seems to be a really simple job. However, the computer's decision is based on models which are constructed with input data. The computational models have a premise garbage in, garbage out. Hence, the input data needs to be carefully selected. Although much progress has been made in correctly discriminating the face to create a robust and meaningful input data for the model, difficulties to correctly create a pure data vector without the influence of external factors, such as illumination or noise, still remains. The purpose of this thesis is to introduce a series of improvements in the existing research in directional codes to reduce the intra-class difference from people of the same gender and to augment the extra-class difference on people from different gender. By utilizing the edges of the face and manipulating the limit between local and global information, it is possible to build a descriptor which can include both information. Then, a division between two types of data is done; edges and flat areas. It is assumed that the axes can contribute more to discriminate the image than a flat region. Moreover, those directional code with something important to contribute the descriptor are going to be marked. Also, the directions that cannot collaborate to build the descriptor are marked and grouped to don't interfere with the previous group. Finally, in the creation of the final feature vector, which works as input data for the model, the algorithm is taking advantage of the neighboring information to build a new type of vector influenced for more global information to overcome the limitation of few directions. The explained method has been experimentally validated by using publicly image sets. Their results and respective comments are presented in this thesis. This is done by utilizing a person independent approach to do not compromise on the result of the descriptor.