On Rotation Invariant Texture Classification Using Two-Grid Coupled CNNs

On the possibilities of using a class of CNN's for texture classification

“…The network performances have been evaluated using the Brodatz album [6]. The classification rates obtained with five different filter banks based on those eight filters have been calculated.…”

“…A particular architecture of the above kind consisting of second order two-port cells interconnected by two resistive grids which has been shown to exhibit Turing patterns is presented in literature [3][4][5]. Moreover, the architecture proved to be appropriate for building banks of spatial filter useful for texture classification [6]. Unfortunately, because of the complexity and the large number of configuration parameters for this kind of network, the CMOS implementation is less attractive from the area, power consumption and accuracy points of view.…”

“…The applications mentioned above are based on the design of first and second order spatial filters and on the configuration of the network with first and respectively second order neighbor connections. The classification algorithm uses the outputs of a filter bank which covers the whole spatial frequency spectrum [6]. Each input image is successively passed through the bank filters and the mean output energies are used as features for classification.…”

On the posibility of using difference of Gaussians for texture clasification

“…For texture classification purpose, in the preprocessing stage, a bank of circular filters has been used [5], [6]. Although this type of filters doesn't have the same results as Gabor filters for example [7], their main advantage is that for data bases in which the images of the same class are not very different one from another, it offers good classification performance for a bank with a smaller number of filters [5], [6].…”

In-Plane Rotation and Scale Invariant Clustering Using Dictionaries

“…These methods learn a dictionary in a log-polar domain to be invariant to scale and rotation. A cellular neural network-based method for rotation invariant texture has also been proposed in [28].…”