Imbalanced Image Classification by An Enhanced Depthwise Separable Convolutions Network
Lulu Qu, Hai-Qing Zhang, Daiwei Li, Xi Yu, Dan Tang, Lei He · 2022
The field of image recognition is developing rapidly, and many high-performance models have been proposed. However, these models encounter some problems when dealing with imbalanced data, which are, the minority classes have low accuracy and the majority classes have high accuracy. This paper proposes an enhanced depthwise separable convolutions network (called EXception), which takes into account both the learning of the minority classes and the majority classes. EXception contains three components. The data augmentation component improves the generalization by cutting and flipping the pictures. The transfer learning component is a pre-trained model (i.e., Xception) on ImageNet. This paper reuses part of the pre-trained parameters of Xception to reduce training time. And the sampling component contains a novel sampling algorithm that can samples the training data. Before the next iteration, according to F1-scores of the previous iteration, the algorithm dynamically changes the number of samples for each class in the next iteration. To a certain extent, the algorithm ensures that the model can fairly learn the features of the majority classes and the minority classes.