Learning Mid-level Image Features for Natural Scene and Texture Classification
Résumé
This paper deals with coding of natural scenes in order to extract semantic information. We present a new scheme to project natural scenes onto a basis in which each dimension encodes statistically independent information. Basis extraction is performed by Independent Component Analysis applied to image patches culled from natural scenes. The study of the resulting coding units (coding filters) extracted from wellchosen categories of images shows that they adapt and respond selectively to discriminant features in natural scenes. Given this basis, we define an image signature as the histogram of the maximum responding filter to the input image. The construction of this histogram takes into account the spatial distribution of these maximal responses within the image. Then, we discuss the construction of a learning database from which filters are extracted, and show that the properties of adaption leading to good discrimination of scenes are preserved when the learning categories are mixed. The entire system is validated by qualitative organization of natural scenes and quantitative categorization of 2500 images.