A statistical approach to improve CNN classification accuracy

Convolutional neural networks (CNNs) have achieved state-of-the-art performance in image classification tasks. However, they may underperform for specific classes, resulting in misclassifications. To address this issue, the proposed method involves two steps that use the Mann-Whitney U test on generated image distributions. The method is evaluated on the publicly available dataset CIFAR100, utilizing ResNet-50 as the baseline network. The results show that the proposed method is effective in significantly improving the classification accuracy of low-accuracy image classes while preserving the high-accuracy classes.