Research

Anatomopathological diagnosis is performed through visual findings of lesions in histological sections. The same type of lesion can present itself differently and its variations have diagnostic relevance. A recurrent problem for pathologists is finding images of lesions with such variations on past case bases. A system capable of automatically retrieving histological images that present characteristics similar to a target image presented by the pathologist is of great importance for the refinement of the anatomopathological diagnosis and for the research of histological markers for the diagnosis of diseases. In Nephropathology, these investigations are carried out considering the specific characteristics of renal structures and their relationship with diseases. This project proposes to find, from a region of interest outlined on a query image, other relevant images based on visual evidence and semantic attributes.

The knowledge of the biodiversity of a region is fundamental for the development of effective productive processes along with the minimization of damages to the environment. At the same time, knowing the characteristics of the species allows the definition of proper preservation policies, and the identification and recognition of flora species is a very important task in the activities of many sectors of society. Consequently, the study and application of modern techniques of representation of characteristics and construction of models for recognition are necessary to allow the development of practical tools. Aligned to the worldwide interested community, this project aims to contribute to the activities of recognition of flora by developing effective and efficient methods with special interest on deep learning approaches.

The technological advances in data capture, storage, and processing allowed the construction of large databases, especially for multimedia data sharing. These data are used in many contexts, such as education, medicine, biometry, social networks, entertainment, news, among others. Considering the huge volume of data, providing efficient and effective access is critical. This project explores the use of modern machine learning techniques for multimedia information retrieval.

With the population of spatial data, the interest for new technologies that allow to analyze the influence of places of interest to make strategic decisions grows. Computing the influence of a spatial object is complex and requires analyzing a large amount of data to obtain an accurate result. This project aims proposes new algorithms and techniques to support geospatial analysis and compute the influence of places of interest.

The location of a new Hub (nodal point) has impacts on the cost and the quality of the deliveries. Moreover, it is fundamental to better employ the qualities of a multimodal transport system. On adding a new Hub, there are several objectives that can be pursued such as: 1) balancing the number of packages to be delivered by each Hub or 2) balancing the distance needed to deliver the packages by each Hub. The aim of this project is to investigate and develop a query type that provides information about the location of a new Hub, giving support to the decision making process. The algorithms and tests developed will be shared in the Loggi Bud.

As more and more online applications is becoming part of our daily life, we are being exposed to a large number of options in these applications. The Netflix catalog, for in- stance, has over 13,900 titles, demanding technologies to show the interesting titles for a given user. Recommender Systems have been used to deal with this problem, trying to predict the options that are most likely to be of interest, creating a personalized and even better experience for each user. There is a large number of papers on Recommendation Systems that uses data provided by applications like NetFlix, where the user explicitly indicates whether or not he likes a recommended item. However, when it comes to Ecommerce, the user does not explicitly indicate whether or not he likes a recommended product. This information is inferred from actions taken by the user such as clicking on a product, adding it to the cart or even purchasing it. This increases the amount of data available to get a recommendation and creates challenges in knowing how to select the best information to infer a user's need. This research project focuses on product recommendation techniques for Ecommerce. The main objective is to investigate the quality and performance of current techniques applied to Ecommerce and propose new techniques that are capable of dealing with the large amount of data from current Ecommerce systems, always focusing on the quality of recommendations produced and performance (time required ) to produce the recommendations after the actions have been taken by the users.