Deep learning for computer vision: improving generalization for small amounts of data

• Grant number: 24/04500-2
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Effective date (Start): September 01, 2024
• Effective date (End): August 31, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Jurandy Gomes de Almeida Junior
• Grantee: Samuel Felipe dos Santos
• Host Institution: Centro de Ciências em Gestão e Tecnologia (CCGT). Universidade Federal de São Carlos (UFSCAR). Campus de Sorocaba. Sorocaba , SP, Brazil

Abstract

With the popularization of social media platforms focused on user-generated content, a huge volume of images and videos are being generated daily. The cataloging and searching of this content is necessary, however, manually analyzing this immense amount of content is practically impossible, making the use of systems capable of automatically capturing the meaning of this content crucial. The field of computer vision deals with these concepts, encompassing aspects of processing, interpretation, and automatic understanding of scenes. Deep learning methods are able to achieve state-of-the-art performance in computer vision tasks, however, their performance is still far from human perception, with much of the success achieved due to the enormous amount of labeled data available in datasets such as ImageNet. However, in real-world scenarios, many application domains do not have enough labeled data, as labeling is often done manually, which has an enormous human cost. Human intelligence is known for its ability to adapt to new scenarios, generalizing from past experiences, a skill that deep models, despite their excellent results, struggle to replicate as they are usually specialized only in the task they were trained on. Motivated by these aspects, this research project aims to propose methods to improve the generalization capacity of deep learning methods for computer vision tasks, reducing their dependence on human supervision. The main contributions will be the development of new strategies to train deep models with a small amount of labeled data, reduce the human effort required to annotate data, and reduce the number of parameters needed to handle multiple tasks. (AU)