Estudo de métodos de re-ranking em imagens

Abstract

The image search processing leads to some issues when the image dataset exponentially grows. Thus, it is necessary to develop techniques to automatically retrieve such images according to the user intention, based on their similarity. To do so, the Content-based Image Retrieval approach is applied. Considering this context, re-ranking algorithms has been used in order to increase de precision of the similarity queries. In this context, algorithms it Re-Ranking has been used in order to increase the accuracy of these systems. This type of algorithm can be used to explore contextual information present in the relationships between some images. The goal of this research was to study and to analyze a Re-Ranking algorithm in order to it in different descriptors to increase its accuracy. The main contribuitions were the analysis and implementation of an algorithm applied to different cases, seeking to improve the precision of CBIR systems. In the experiments we utilized different descriptors, distance functions and comparison of measures in two image databases. The results showed that the algorithm is effective for most cases, however for each case, it was necessary to adequate the algorithm parameters due to the image database and methods employed.