Contextual Similarity Learning Applied to Graph Convolutional Networks for Image Classification and Retrieval

• Grant number: 25/10602-5
• Support Opportunities: Regular Research Grants
• Start date: February 01, 2026
• End date: January 31, 2029
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Lucas Pascotti Valem
• Grantee: Lucas Pascotti Valem
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• City of the host institution: São Carlos
• Associated researchers: Agma Juci Machado Traina ; Daniel Carlos Guimarães Pedronette ; Longin Jan Latecki ; Luis Gustavo Nonato ; mohand said allili
• Associated scholarship(s): 26/16921-8 - Identification of Representative Vertices and Feature Graph Fusion in Graph Convolutional Networks, BP.IC; 26/13387-0 - Graph Modeling with Contextual Similarity in Graph Convolutional Networks for Image Classification and Retrieval, BP.IC

Abstract

The exponential growth of image collections has made the development of effective content-based retrieval systems essential. Deep learning approaches, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViT), show great potential, but training these models typically requires large volumes of labeled data, which is costly and time-consuming to obtain. In this context, unsupervised and semi-supervised methods stand out for reducing this dependency. Additionally, many current approaches still rely solely on pairwise similarity analysis, which poses limitations. Thus, leveraging contextual information, by considering not only the similarity or distance between pairs, but also the incorporation of neighborhood information, is a promising direction. From this perspective, Graph Convolutional Networks (GCNs) have gained attention for applying convolutions over graph structures, enabling the capture of complex relationships beyond the Euclidean space. Despite their potential, GCNs remain underexplored in image collections. This project proposes to investigate the use of GCNs for image retrieval and classification, aiming to enhance the effectiveness of current methods and to propose novel approaches, particularly in scenarios with limited availability of labeled data. The research will explore strategies based on contextual and ranking information to improve five key aspects: (i) input feature vectors; (ii) graph construction; (iii) use of labeled data; (iv) training strategies; and (v) effective representation learning. The proposal aligns with the expertise of the main researcher, who has been active in the field of image retrieval for over a decade. (AU)