Data augmentation for deep ensembles in polyp segmentation

The last few years have witnessed a growing interest among computer vision researchers in semantic segmentation, which, in general, learns low-level features and the semantics of an image via an encoder-decoder structure. In brief, the task of the encoder is to extract features by exploiting convolutional layers; the job of the decoder is to generate the same image by applying skip connections in the first layer. This chapter aims to test different data augmentation approaches for boosting segmentation performance by taking DeepLabv3+ as the architecture and ResNet18/ResNet50 as the backbone. The proposed set of data augmentation approaches is coupled with an ensemble of networks obtained by randomly changing the activation functions inside the network multiple times. Moreover, the proposed approach is combined with HardNet-SEG, a recent architecture for semantic segmentation, for a further boost of the performance.