Balanced Mini-Batch Training for Imbalanced Image Data Classification with Neural Network

Abstract

We propose a novel method of training neural networks for industrial image classification that can reduce the effect of imbalanced data in supervised training. We considered visual quality inspection of industrial products as an image-classification task and attempted to solve this with a convolutional neural network; however, a problem of imbalanced data emerged in supervised training in which the neural network cannot optimize parameters. Since most industrial products are not defective, samples of defective products were fewer than those of the non-defective products; this difference in the number of samples causes an imbalance in training data. A neural network trained with imbalanced data often has varied levels of precision in determining each class depending on the difference in the number of class samples in the training data, which is a significant problem in industrial quality inspection. As a solution to this problem, we propose a balanced mini-batch training method that can virtually balance the class ratio of training samples. In an experiment, the neural network trained with the proposed method achieved higher classification ability than that trained with over-sampled or undersampled data for two types of imbalanced image datasets.