A Symmetric grammar approach for designing segmentation models

Image segmentation is a relevant problem in computer vision present in multiple application domains. One of the most used methods for image segmentation is U-net, a type of convolutional network with additional constraints in its architecture. Studies regarding the U-net usually rely on well-known architectures, which leads to a narrow exploration of the possibilities, and possibly impacting the performance. Genetic Programming approaches have become increasingly popular for designing neural networks due to studies where the generated models were able to achieve results comparable to humans. These approaches can evolve the structure at different levels of abstraction, reducing the need for a specialist. In this paper, we propose the use of Grammatical Evolution for evolving U-net architectures. We propose a mirror grammar, which is capable of generating a variety of flexible U-nets that better explores the search space. We show that the proposed grammar can capture the complex constraints that define the U-nets and achieve comparable results in terms of accuracy, on a benchmark of segmentation problems of varying difficulty.