Supervised texture identification using dictionary based data modelling

• Main Authors: Ranjan, Raju, Gupta, Sumana, Venkatesh, K. S.
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Data models; Dictionaries; Feature extraction; Hidden Markov models; K-SVD; Mathematical model; OMP; Pattern Learning; Sparse Framework; Supervised Texture Identification; Training
• ISSN: 2162-7843
• DOI: 10.1109/ISSPIT.2014.7300625

Texture identification is an important preliminary step in many computer vision applications. There exists supervised and unsupervised approaches to solve this problem. One of the widely used unsupervised technique is KMeans which identifies the region of image based on clustering. The disadvantage of the KMeans technique is that it is an off-line approach that needs all the data prior to processing. In case of supervised techniques such as SVM, Neural Network and HMM the pattern of training data is learnt in space. In this paper, we propose a supervised sparsity based texture modelling. The texture classes are available a priori. For each texture class, feature vectors are extracted and a model is learnt for the set of feature vectors. The sparsity based data modelling attempts to learn the union of subspaces in which feature vectors lie. The set of subspaces are learnt in the form of dictionary. Dictionary contains a set of basis vectors referred to as atoms. Any feature vector can be represented as a weighted linear combination of sparse set of atoms taken from dictionary. A dictionary is learnt for each of the texture classes. Also, a dictionary corresponding to rest of the texture classes is also learnt. The final dictionary is obtained by horizontally cascading both the dictionaries side by side. This is done for each of the texture classes. For decision making, the feature vector corresponding to each pixel in the test image is tested with each of the dictionaries. The residual error is estimated using first half of the sparse code and original dictionary of the texture class. The dictionary which represents the feature vector with least residual error by using the predefined sparsity is said to be the texture class to which the feature vector belongs to. The proposed algorithm is shown to achieve better result than SVM.