Studying the effect of Drop_Out rate on the performance of U_Net for semantic segmentation image

Abstract

Semantic segmentation in satellite images is a detailed technique to take advantage of space imaging, Deep learning networks are currently considered one of the basic techniques for solving many real-world problems, including processing satellite images, especially Encoder-Decoder Architecture. Designing the architecture of a deep learning network requires searching in a wide range of solutions, ranging from simple structures, which give modest results, to more complex structures, in the hope of reaching better results. These solutions come with different time and technology costs (the hardware on which the software runs).

This research attempts to investigate solutions corresponding to local conditions from the number of available satellite images and processing the raw images before entering them into the semantic segmentation stage, For implementing this stage, the research uses the modified U_Net network in terms of the size of the input layer, the number of convolutional layers, the type of activation function, the type of pixel Up_Sample process in the second path of the network, Moreover, research study's adding Drop_Out layers to the network structure (studying different percentages of Drop_Out and its effect on the model results).