A correlation graph approach for unsupervised manifold learning in image retrieval tasks

Effectively measuring the similarity among images is a challenging problem in image retrieval tasks due to the difficulty of considering the dataset manifold. This paper presents an unsupervised manifold learning algorithm that takes into account the intrinsic dataset geometry for defining a more effective distance among images. The dataset structure is modeled in terms of a Correlation Graph (CG) and analyzed using Strongly Connected Components (SCCs). While the Correlation Graph adjacency provides a precise but strict similarity relationship, the Strongly Connected Components analysis expands these relationships considering the dataset geometry. A large and rigorous experimental evaluation protocol was conducted for different image retrieval tasks. The experiments were conducted in different datasets involving various image descriptors. Results demonstrate that the manifold learning algorithm can significantly improve the effectiveness of image retrieval systems. The presented approach yields better results in terms of effectiveness than various methods recently proposed in the literature.