Texture analysis via unsupervised and supervised learning

Abstract

A framework for texture analysis is proposed based on combined unsupervised and supervised learning. The textured input is represented in the frequency-orientation space via a Gabor-wavelet pyramidal decomposition. This representation has optimal localization properties in space and in frequency, and is biologically motivated. In an unsupervised learning phase a neural network vector-quantization scheme is used for the quantization of the feature-vector attributes and a projection onto a reduced dimension clustered map for initial segmentation. A supervised stage follows in which labeling of the textured map is achieved using a rule based system. This information theoretic technique is utilized to find the most informative correlations between the attributes and the texture class specification. Ultimately, a minimal representation for a library of textures is learned in a training mode, following which the classification of new test images is achieved. This combined neural network-rule based approach is suggested as a general framework for pattern classification. Simulation results for the texture classification are given.

Self-Supervised Learning of a Biologically-Inspired Visual Texture Model

“…In the context of texture classification, current human-labeled homogeneous texture databases are few and small, so most deep learning methods transfer features from networks trained with full supervision on an alternative task (typically, object recognition). Some authors have developed limited unsupervised methods based on vector quantization [57,58], and non-negative matrix factorization [59]. Nevertheless, in concert with CNN models, we believe ours is the first competitive self-supervised learning objective for this problem.…”

Novel method and system for pattern recognition and processing using data encoded as Fourier series in Fourier space

“…A vector-quantization learning algorithm defines a mapping from an N-dimensional input vector, X, to an M-dimensional output vector Y involving feature vectors, their quantization, clustering, and rule-based mappings (e.g. Bayesian classifier) (Greenspan et al, 1991). Alternatively, statistical pattern recognition uses measurements and transforms of the pattern structure as feature vectors (Dickhaus and Heinrich, 1996).…”

“…Image relevance is assessed by computing the dot product of each summary vector with the query context vector, and accumulating the results. A Gabor transform may be used to extract features from feature vectors in the frequency and orientation space to form associations (Greenspan et al, 1991). A vector-quantization learning algorithm defines a mapping from an N-dimensional input vector, X, to an M-dimensional output vector Y involving feature vectors, their quantization, clustering, and rule-based mappings (e.g.…”

Face surveillance

“…Homogeneous ensembles of symbolic modules are usually referred to as multistrategy learning (AI) methods. As an example of heterogeneous ensembles, Greenspan [5] has proposed an architecture for the integration of neural networks and rule-based methods using unsupervised and supervised learning for pattern recognition tasks.…”