Improved Image Data Augmentation Using Dynamic Mode Decomposition

Many real world applications utilize highly imbalanced data, with the target variable located in the minority class. The unequal distribution of data often results in machine’s inability to carry out predictive accuracy in determining minority classes, thereby causing various classification errors. A balanced and feature-enhanced datasets plays a crucial role in ensuring robust and accurate classification outcomes. This study considers Dynamic mode decomposition(DMD) as an oversampling technique for image datasets applied to minority classes that involves creating new examples that are variations of existing ones. This research use the published dataset from Guangzhou Women and Children’s Medical Center. For the considered dataset, the findings suggest that the images generated using DMD based attention-driven image enhancement algorithm are of better quality than ACGAN and normal translations with improved CC, PSNR and SSI. Classification accuracies of various deep CNN pre-trained models were compared when used with different image balancing techniques for a 3 class image dataset. From the test results, it is observed that the proposed methodology offers better class-wise performance with faster execution time while existing methods fails to achieve this. EfficientNetB0, Densenet121 and InceptionV3 when used DMD based balanced data, exhibited an accuracy of more than 90% with classwise accuracy of more than 85% for all three classes while ResNet50V2, MobileNetV2 exhibited accuracies close to 90%.