PROJECT TITLE :

Deep Attention-Based Imbalanced Image Classification

ABSTRACT:

In many real-world image classification problems, there is a common issue known as class imbalance. This occurs when some classes have abundant data while other classes do not. In this scenario, the representations of classifiers are likely to be biased toward the classes that make up the majority of the data, and it will be difficult to learn the appropriate features, which will result in unpromising performance. In order to get rid of this biased representation of the features, many algorithm-level methods have learned to pay more attention to the minority classes explicitly based on their prior knowledge of the distribution of the data. An approach that is based on attention and is given the name deep attention-based imbalanced image classification (DAIIC) is proposed in this article. The goal of this approach is to automatically pay more attention to the minority classes in a data-driven manner. In the proposed approach, an attention network and an innovative attention-augmented logistic regression function are utilized in order to encapsulate as many features, which belong to the minority classes, as possible into the discriminative feature learning process. This is accomplished by assigning the attention for various classes jointly in both the prediction and feature spaces. DAIIC will be able to automatically learn the costs of misclassification for the various classes if the proposed object function is implemented. After that, the misclassification costs that were learned can be used to guide the training process in order to learn more discriminative features by making use of the attention networks that were designed. In addition to that, the method that was suggested can be utilized with a wide variety of networks and data sets. The proposed method outperforms several state-of-the-art methods for imbalanced image classification, as shown by experimental results on single-label and multilabel imbalanced image classification data sets. These results demonstrate that the proposed method has good generalizability.