The objective of this thesis is to automatically extract massive amounts of visual and statistical information from large image collections. In particular, the objective which will direct the research is that of the creation of a visual catalog of the objects appearing in an image collection. Such information is of great value for subsequent tasks and will open or facilitate new directions of research. This thesis will also endeavor to highlight, via proofs of concept, the interest of a large-scale automatic extraction for applications (1) societal (understanding of the content posted on social networks), (2) industrial and (3) research, for which we will use the automatically collected knowledge as data useful for statistical learning.