Multi-scale cross-attention for multi-class imbalanced image classification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper addresses the multi-class imbalanced image classification task by a positive and negative multi-scale cross-attention (PNMSCA) network. The proposed PNMSCA method explores the feature-wise compatibility scores between the input image with each of minority and majority classes in multiple scales. To this end, the positive and negative multi-scale cross-attention features are computed from the minority and majority representatives, and then, fused together. The multi-scale cross-attention features are used as weighting factors in a convolutional network where they are injected to the model through residual connections. The PNMSCA method employs a compact, parameter-efficient architecture model, which yields better classification performance with significant difference with respect to several competitors on benchmark SVHN, CIFAR-10 and FER2013 image datasets. In the SVHN dataset, PNMSCA achieves an average Fmeasure improvement of approximately 0.88% over the best-performing competing method across the majority classes, while the improvement increases to approximately 1.87% across the minority classes. In CIFAR-10, PNMSCA also achieves consistent improvements in overall classification performance achieving the highest average Fmeasure compared with the competing methods. For FER2013, PNMSCA achieves an average Fmeasure improvement of approximately 5.96% for the minority class compared with the best non-PNMSCA method. These results demonstrate the effectiveness of PNMSCA for imbalanced image classification, particularly in improving recognition of classes with limited training samples.