Synthetic data generation models and algorithms for training image recognition algorithms using the Viola-Jones framework

The paper describes mathematical models and algorithms of warping grid functions with discrete parameters. For images, three warping models, applied to the generation of extra training data to build face recognition algorithms, are examined: the one based on harmonic functions, the one based on offsetting user-specified control point coordinates, and the one based on the computation of the optical flow between entropy-filtered images. For the training sets, both initial and those synthetically generated using the above three models, learning of face detection algorithms based on the Viola-Jones framework was performed and corresponding detection rates were compared. It is shown that this approach is applicable for synthetic data generation when training image recognition algorithms for recognition of objects characterized by inherent structure.