Few shot image segmentation with using U-net

Abstract

Image segmentation is a typical task in the field of computer vision. Compared to the task of image segmentation within the framework of the hypothesis with a large set, recognition with a small data set can not only identify known categories in the training data set, but also assign colored labels to objects. In real-world segmentation tasks, due to various objective factors, it is usually difficult to collect exhaustive training samples for all categories when training a neural network. During training, there is a complete knowledge of categories about the world, and during testing, the algorithm is provided with information about the object (object mask). Segmentation by neural networks is required for accurate segmentation of existing categories in the training data set by assigning labels to them. In recent years, with the development of deep learning, deep learning for image segmentation relies heavily on various neural network models existing today. Although neural networks are very efficient, good results can rarely be explained. A recent method for solving the object segmentation problem involves two main stages: preparation of the initial images of a data set with specially prepared masks; and then training a compact neural network model using a specially prepared dataset. For an instance of a test object, when the probability is below the threshold value of a known class, it is determined that the object belongs to the category seen in the training data set. However, this method does not take into account information about the image data of unknown categories in the process of training the model.