Data augmentation approaches for polyp segmentation

Image augmentation, and in general data augmentation techniques, can greatly improve the performances of deep neural networks through the creation of artificial patterns, in fact the presence of these new patterns helps the network to generalise thus avoiding overfitting, i.e. the excessive adaptation of the network to the training data. The application of these techniques has become increasingly important, especially in fields where data availability is scarce, such as the medical field. This work reports various non-traditional image augmentation techniques based on filtering and mixup operations in the frequency domain, tone mapping, background change and others. To evaluate the performance and benefits, these techniques are applied to the Kvasir-SEG dataset, which is then used to train the DeepLabV3+ semantic segmentation model.