Transforms or operators that map input images to output images can be applied to several areas of study, such as on the analysis of medical images or of document images. The manual creation of operators that perform desired transformations is a task that demands considerable time and effort. Thus, more recently, methods that seek automatic design of these operators and that are based on deep learning techniques are gaining attention. Among such approaches, some are local and seek learning of local characteristic functions that can be seen as individual pixel classifiers, while others are global and based on end-to end, i.e. image-to-image, training methods. Both have distinct advantages and disadvantages. The goal of this project is to study and compare the two approaches, and to create a new technique that joins the interesting features of the local and global approaches. The studied approaches will be applied on two real problems. One of the problems consists of vessel segmentation in retina images, with largely used public datasets, and the second is related to a collaboration with the Medical School of USP.
News published in Agência FAPESP Newsletter about the scholarship: