Person recognition from eye circumference images by comparison of different transfer learning algorithms

Machine learning methods are used for purposes such as learning and estimating a feature or parameter sought from that data set by training a data set in solving a particular problem. The transfer learning approach, which is aimed at transferring the ability of people to continue learning from their past knowledge and experiences, to computer systems is the transfer of the learning obtained in the solution of a particular problem, so that it can be used in solving a new problem. Transferring the learning obtained in transfer learning provides some advantages over traditional machine learning methods, and these advantages are effective in the preference of transfer learning. In this study, a total of 1980 eye contour images were collected from 96 different people. These collected data were tried to be classified in terms of person, age and gender. For this, feature extraction was performed with 32 different transfer learning algorithms in the Python program and classified with RandomForest algorithm for person estimation. 30 different classification algorithms were used with the most successful ResNet50 algorithm, and the data were also classified in terms of age and gender. Thus, the highest success rates were obtained as 83.52%, 96.41% and 77.56% in person, age and gender classification, respectively.