Material Recognition under General Conditions

The human capacity for recognizing visual scenes relies heavily on discerning shapes and materials. However, the appearance of surface materials can undergo significant changes under natural observation conditions, posing challenges for reliable automatic recognition in various artificial intelligence applications. Real surface textures exhibit intricate dependencies on factors such as illumination angles, viewing perspectives, foreshortening, local shading, interreflections, shadowing, and occlusion of surface elements. The primary goal of this research and development proposal is to pioneer advanced methods for automatically (un)supervised recognizing surface materials across varying geometric and lighting conditions. These methods leverage sophisticated statistical techniques, utilizing seven-dimensional probabilistic visual texture models to approximate state-of-the-art measurable textural representations such as bidirectional texture functions (BTF). Our approach involves the development of intricate multidimensional models, including compound multi-resolution Markov random field models, neural transformers, mixture models, and non-linear reflectance models. The project's focus is on contextual classification of surface materials under realistic observation conditions, with an emphasis on enhancing visual scene interpretation. The outcomes of this project have a substantial impact on industrial applications and medical diagnostics, by advancing the analysis of visual information.