Paradigmas de aprendizaje profundosoluciones metodológicas de aplicación a problemas de interés

Abstract

In this doctoral thesis a theoretical and experimental study has been carried out in the field of deep learning. The objective of this study has been to improve the current techniques used in problems from different fields of interest of society. The main problems that deep learning models present when implemented in real environments are the lack of precision and the high computational cost of their execution. There are problems in which the accuracy of the model must be as high as possible for this type of solutions to be used. In addition, the processing demands of the models may prevent their use in certain problems where the processing capacity is very limited or a high processing speed is required, as in the case of systems working in real time. Therefore, it is necessary to find new techniques and paradigms to improve the results of the current solutions and to reduce the processing time without losing accuracy. For the development of this thesis, two problems have been chosen in which the need for the mentioned above improvements is evident. On the one hand, the first problem consists of the segmentation of the vascular tree in fundus images. This problem is of special interest since it allows the creation of a tool to support the specialist in monitoring the vascular tree with the purpose of detecting different pathologies. However, as this is a system intended for use in the medical field, any improvement in the model results gives the specialist greater confidence in this tool. On the other hand, the second problem consists in the segmentation by instance of strawberries in images. This problem is a key part in the creation of automatic strawberry harvesters and, therefore, it is necessary to use models that can work in real time on equipment with processing and memory limitations. The result of the work done in this thesis has led to three publications in which solutions have been proposed that address the main limitations of deep learning models in the two problems of interest: performance improvement and processing speed. Specifically, for the case of vascular tree segmentation, a new model based on U-Net (reference model for semantic segmentation) and new techniques for training have been proposed that, as a whole, manage to improve the state-of-the-art results. Thus, the proposed model, with a much more efficient architecture than the original model and without the need to apply processing to the image before or after being processed by the model, presents higher AUC values than those obtained by the most representative models of the state of the art. In the case of strawberry instance segmentation, a modification of Mask R-CNN (one of the reference models in instance segmentation) has been proposed with the goal of improving the processing speed of the original model considerably without notably affecting the results. The proposed modifications have made it possible to work at 10 fps, which means doubling the speed of the original model without notably affecting the mAP value. In addition, a new paradigm has been proposed to address the problem of strawberry instance segmentation, as well as a new model, which for the first time in this problem is able to work in real time (30 fps) with a 15% increase of mAP value compared to Mask R-CNN. The results achieved in the work done in this thesis for the two chosen problems allow us to consider the proposed solutions as the best alternatives for implementation in commercial working environments.