| Resumo: | With the advent of large-scale geo-tagged image sharing on the internet, on websites such as Flickr and Panoramio, there are now large sources of data ready to be mined for useful information. Using this data to automatically create a map of man-made and natural areas of our planet, can provide additional knowledge to decision-makers responsible for world-conservation. The problem of determining the degree of naturalness of an image, required to create such a map, can be generalized as a scene classification task. Experiments were performed to better understand the applicability of each of the identified scene classification techniques to perform the distinction between man-made and natural images. Their advantages and limitations, such as their computational costs, are detailed. With careful selection of techniques and their parameters it was possible to build a classifier capable of distinguishing between natural and man-made scenery with high accuracy and that can also process a large amount of pictures within a reasonable time frame.
|
|---|