Automatic Design of Deep Neural Networks Applied to Image Segmentation Problems

A U-Net is a convolutional neural network mainly used for image segmentation domains such as medical image analysis. As other deep neural networks, the U-Net architecture influences the efficiency and accuracy of the network. We propose the use of a grammar-based evolutionary algorithm for the automatic design of deep neural networks for image segmentation tasks. The approach used is called Dynamic Structured Grammatical Evolution (DSGE), which employs a grammar to define the building blocks that are used to compose the networks, as well as the rules that help build them. We perform a set of experiments on the BSDS500 and ISBI12 datasets, designing networks tuned to image segmentation and edge detection. Subsequently, by using image similarity metrics, the results of our best performing networks are compared with the original U-Net. The results show that the proposed approach is able to design a network that is less complex in the number of trainable parameters, while also achieving slightly better results than the U-Net with a more consistent training.