Rotation before Recognition: Orientation-Invariant Identification of Visual Textures with ARTEX 2
We describe the ARTEX 2 neural network for recognition of visual textures at arbitrary orientations. ARTEX 2 recognizes visual textures by first passing them through a preprocessing stage which rotates them to a canonical orientation. The resulting canonically-oriented visual textures are then classified by a simplified version of the ARTEX texture-recognition algorithm. Several approaches to determining the proper angle for rotation to a canonical orientation are investigated, and their respective performances are compared on a 20-class database of visual textures.