A New Augmented Method for Processing Video Datasets Based on Deep Neural Network

Large datasets are required for deep learning to achieve good performance. However, there is a lack of sufficient training datasets in many research fields, which may become a shortcoming of computer vision applications. This article provided a new data augmentation method for making training small datasets, which could be divided into two steps: 1. Unbalanced sampling based on information density. 2. Splicing images to form a dataset. Different information density dataset combinations had been used for testing the model generalization. The enhanced loss function which consisted of label smoothing loss and cross-entropy loss had been used to minimize the model preference during training models. Finally, with the same amount of data, the Mean Absolute Error (MAE) of the model with our sampling method could get 55% increase compared with the traditional sampling method. The best MAE could reach 0.98 if the splicing method had been adopted. The results showed that this augmented method was suitable for scenarios with small sample size, especially video datasets. To get the best performance, the splicing method was a nice choice to optimal model generalization performance.