Deep Detection and Segmentation Models for Plant Physiology and Precision Agriculture stars

Resum

Computer Vision is a multidisciplinary field that combines concepts from Artificial Intelligence, image processing, visual perception, and data science to enable computers to understand and analyse visual content in a similar way to humans. In the last years, significant advancements have been made in Computer Vision thanks to the development of algorithms and techniques based on Deep Learning methods. Two areas of Computer Vision that have numerous applications in various fields such as biology, agriculture, and medicine are Object detection and semantic segmentation. Currently, the most successful techniques to tackle these two tasks are also based on Deep Learning methods. However, although these methods have achieved excellent results, using such techniques in contexts outside machine learning can be complex. This is due to the large number of images that are required to train Deep Learning models (which can be difficult to obtain in context like biomedicine or precision agriculture), the process of annotating images (a time-consuming and expertise-demanding problem), and the technical difficulties for training and using Deep Learning models by domain experts. The aim of this thesis is to address these limitations through different theoretical developments, and evaluate the proposed solutions in actual contexts. First of all, we have focused on the development of methods that allow us to improve the performance of object detection models. For this purpose, we have developed an algorithm that improves the accuracy and robustness of object detection models by means of an ensemble method. This algorithm is also the basis to define semi-supervised learning methods that reduce the number of annotated images that are needed to train object detection models. Moreover, to facilitate the create and usage of object detection models, we have developed an open source tool that simplifies the process of creating and using object detection models, thanks to a simple to use graphical interface. Furthermore, we have generalise our work to facilitate the creation and usage of models for any computer vision task. Finally, the developed techniques and tools have served as the foundation for addressing real-world problems in plant physiology, and in precision agriculture. As a summary, this work is a step towards the democratisation of Deep Learning models for users outside the machine learning community.