LIFE-VISION is a research project funded by the Ministry of Economy and Competitiveness (MINECO) for the period 2013-2015. The scientific research focuses on learning the relevant image features, manifold structure, and statistics.
This approach is key to the development of innovative and enhanced methods for image and video compression, classification and segmentation, restoration, regression, model inversion, image quality assessment, and perceptual saliency. During the last decade, we have proposed successful image models and representations to properly encode visual information, and investigated the relationship between cortical computing and image statistics. Based on our experience, and on the current trends in image processing and machine learning, we approach the problem by learning the relevant image manifold characteristics, i.e., its local and global behavior, geometry, regularities, and dynamic changes.
Results of our previous projects support this novel manifold-oriented approximation to image processing and the need for cooperation. Manifold learning cuts to the heart of nonlinear feature extraction, while manifold regularization encompasses the inclusion of prior knowledge, learning and encoding invariances, and spatial-temporal-spectral texture characterization. LIFE-VISION focuses on answering key scientific questions that need to be addressed in the context of image processing: the study of the intrinsic dimensionality of images, the error minimization versus information maximization dilemma, the domain adaptation issue, the optimal coding hypothesis, the statistics across image sources, the dynamics of invariances, and the metric learning issue.
The study of these aspects in the manifold structure of images is the guiding principle of this project. Theoretical findings will be assessed in four main application scenarios, in which the members are recognized experts: 1) image and video processing, 2) remote sensing, 3) medical image processing, and 4) computational visual neuroscience. A conceptual view of the project follows.