Abstract
While the transformer architecture is currently the most powerful machine learning architecture, it requires massive amounts of data for training. In addition, the execution requires large amounts of memory since the self attention increases the model size quadratically based on the amount of key, query, and value assignments. The U-Net architecture, which is based on normal convolutions, batch normalization, and residual blocks, is a fast and efficient DNN architecture that can effectively be used for semantic segmentation. In this paper, I evaluate different U-Net modifications and compare them to the state of the art as well as the baseline U-Net with cross connections. Additionally, I propose a novel modification that extends the cross connections together with resizing. The proposed approach slows the model slightly but improves the performance significantly. All evaluations are done on the TEyEDS dataset.