Training deep artificial neural networks using small data sets

This thesis explores approach which seeks to improve precision of deep neural networks trained on small data sets. At first model was trained on small labelled dataset until convergence. Then detection was performed on remaining images which were treated as unlabelled data. All images with objects which were recognized with confidence above certain threshold were added to new training set. Then additional training was performed using this new dataset. his method proved to be successful in cases where training is performed on small number of classes. For those cases accuracy increased up to 0,05 mAP. When amount of classes increased, improvement was not achieved. Best results were achieved while using threshold in range of 0.5–0.3