Visualization techniques are often used to bring the user tools for analysis and exploration of a set of images. In most processes of image analysis is the need for preprocessing, in which are calculated and extracted feature vectors. These vectors are mapped and positioned as points in a plane, for example by multi-dimensional projection techniques. A common difficulty is the large number of characteristics that define a space of high dimension, strongly affecting the performance on visual analysis, as well as the other methods of mining images, such as clustering and classification algorithms. Dealing with this problem usually requires dimensionality reduction techniques or selection of features. Additionally, the various forms of feature extraction produce different vector spaces which can describe the set under severely different forms of analysis.The visual representation is used to reveal useful information to users, such as the extraction of implicit knowledge, the relationship of data, or other standards are not explicit in the data, or simply to determine whether attributes used to representing the set of objects are appropriate for an application.This project intends to give sequence to the works until now developed that use visualization techniques as support to evaluate feature spaces generated by collections of images. The objective is to develop a method based on visual analysis using techniques of positioning of points for multidimensional data, which give supporting to analyst to select a description of the image set. This task involves establishing a criterion of quality for the feature space, which can also be based on statistical and visual methods or by linking this method to the optimal visualization, set a standard to support the user in the process of selecting a suitable vector space for a specific application.
News published in Agência FAPESP Newsletter about the scholarship: