Data Augmentation Methods for Deep Learning Neural Networks

Standard algorithms face difficulties when learning from unbalanced datasets because they are built to handle balanced class distributions. Although there are various approaches to solving this issue, solutions that create false data represent a more all-encompassing strategy than algorithmic changes. In particular, they produce fictitious data that any algorithm can use without limiting the user's options. In this paper, we present five oversampling methods: Synthetic Minority Oversampling Technique (SMOTE), Random Over Sampling (ROS), K-Means Smote (KMS), Affinity Propagation and Random Over Sampling-Based Oversampling (APROSO), and Self-Organizing Map-based Oversampling (SOMO). We also present four undersampling methods: Random Under Sampling (RUS), Cluster Centroids (CCs), Neighborhood Cleaning Rule (NCR), Near Miss-1 (NM1). To evaluate those over and under-sampling methods, we have used two different Deep Neural Network (DNN) models, i.e., DNN model 1 and DNN model 2. The empirical result shows that all the over and under-sampling methods are providing more effective results on DNN model 2. The result analysis also shows that the oversampling methods are more effective in classifying the Magnoliopsida and Pinopsida images.