An Empirical Study on Segmentation Methods with Deep Ensembles and Data Augmentation

In the past few years, there has been a growing focus on semantic segmentation, which involves assigning each pixel in an image to a specific label from a given set[65].The use of autoencoder architectures has been explored by numerous computer vision researchers in an attempt to develop models capable of learning both the semantics of an image and a low-level representation of it. When utilizing an autoencoder architecture, the input undergoes encoding to produce a low-dimensional representation. This representation is subsequently leveraged by a decoder to reconstruct the original data. The presented ap- proach involves a combination of convolutional neural networks (CNNs) and transformers to form an ensemble, as detailed in this work. Ensemble methods rely on multiple models being trained and utilized for classification, with the ensemble combining the outputs of individual classifiers. By capitalizing on the varying strengths of each classifier, this approach enhances the overall performance of the system. Distinct loss functions are employed to ensure diversity among the individual networks. The ensemble method em- ploys a combination of the DeepLabV3+, HarDNet, and PVT environments, with varying backbone networks. Additionally, a novel loss function is presented, which integrates the Dice and Structural Similarity Index. To assess the proposed ensemble, a comprehensive empirical evaluation is conducted on six real-world scenarios, namely polyp, skin segmen- tation, leukocyte segmentation, butterfly identification, microorganism identification, and radiology segmentation. The proposed model has achieved state-of-the-art performance on these scenarios.