This study systematically investigates the impact of class imbalance on the performance of convolutional neural networks (CNNs) and compares common methods to address this issue. Class imbalance is a common problem in machine learning, but very limited research exists in the context of deep learning. The study uses three benchmark datasets—MNIST, CIFAR-10, and ImageNet—to evaluate the effects of imbalance on classification and compare methods such as oversampling, undersampling, two-phase training, and thresholding. The main evaluation metric is the area under the receiver operating characteristic curve (ROC AUC), adjusted for multi-class tasks, as overall accuracy is not reliable for imbalanced data. The study concludes that class imbalance significantly harms classification performance. Oversampling is the most effective method in most scenarios, and it should be applied to completely eliminate imbalance. Undersampling is less effective and its optimal ratio depends on the extent of imbalance. Oversampling does not cause overfitting in CNNs, unlike some classical machine learning models. Thresholding is recommended when the overall number of correctly classified cases is important. The study also shows that the effect of imbalance depends on the distribution of examples among classes, not just the total number of examples. For extreme imbalance, undersampling can perform similarly to oversampling, but it reduces training set size, making it a better choice when training time is a concern. Thresholding combined with oversampling is the most effective approach for achieving high accuracy. The study highlights the importance of considering class distribution when evaluating the performance of CNNs on imbalanced data.This study systematically investigates the impact of class imbalance on the performance of convolutional neural networks (CNNs) and compares common methods to address this issue. Class imbalance is a common problem in machine learning, but very limited research exists in the context of deep learning. The study uses three benchmark datasets—MNIST, CIFAR-10, and ImageNet—to evaluate the effects of imbalance on classification and compare methods such as oversampling, undersampling, two-phase training, and thresholding. The main evaluation metric is the area under the receiver operating characteristic curve (ROC AUC), adjusted for multi-class tasks, as overall accuracy is not reliable for imbalanced data. The study concludes that class imbalance significantly harms classification performance. Oversampling is the most effective method in most scenarios, and it should be applied to completely eliminate imbalance. Undersampling is less effective and its optimal ratio depends on the extent of imbalance. Oversampling does not cause overfitting in CNNs, unlike some classical machine learning models. Thresholding is recommended when the overall number of correctly classified cases is important. The study also shows that the effect of imbalance depends on the distribution of examples among classes, not just the total number of examples. For extreme imbalance, undersampling can perform similarly to oversampling, but it reduces training set size, making it a better choice when training time is a concern. Thresholding combined with oversampling is the most effective approach for achieving high accuracy. The study highlights the importance of considering class distribution when evaluating the performance of CNNs on imbalanced data.