Imbalanced Image Classi cation Using Adaptive Dynamic Oversampling Framework in Deep Feature Space

Abstract

In real world applications, it is very common to encounter data with high class imbalance. Imbalanced dataset is a challenging issue in practical classi cation problem, as the classi er gets biased towards the majority classes. The traditional techniques like synthetic minority oversampling have great success in traditional machine learning problems with class imbalance, however these techniques fail to perform well in the eld of complex, structured and very high dimensional data like images. In our work we propose a novel dynamic oversampling framework, which is broadly subdivided into three parts. The rst step is the representation learning of the dataset, where a Convolutional Neural Network is used to map the raw input training data into a new feature space. In the second step a modi ed minority oversampling technique is implemented with adaptive k-NN based search between in-class samples in deep feature space. Finally a dense neural classi er is trained on the augmented dataset. To increase the discriminating power of the nal classi er we have trained it with modi ed sample weights. We have also supplemented our work with empirical studies on publicly available benchmark image datasets and have shown that our technique provides a good countermeasure to handle imbalanced image datasets and provides superior performance than existing techniques.