Texture Classification Using Sparse Frame-Based Representations

Abstract

A new method for supervised texture classification, denoted by frame texture classification method (FTCM), is proposed. The method is based on a deterministic texture model in which a small image block, taken from a texture region, is modeled as a sparse linear combination of frame elements. FTCM has two phases. In the design phase a frame is trained for each texture class based on given texture example images. The design method is an iterative procedure in which the representation error, given a sparseness constraint, is minimized. In the classification phase each pixel in a test image is labeled by analyzing its spatial neighborhood. This block is represented by each of the frames designed for the texture classes under consideration, and the frame giving the best representation gives the class. The FTCM is applied to nine test images of natural textures commonly used in other texture classification work, yielding excellent overall performance.

Learning Dictionaries of Discriminative Image Patches

“…The proposed method is simple compared to prior work because it contains all texture information in one dictionary. This is different from methods based on learning a dictionary for each class like [11,12,15,18]. Employing an approximate nearest neighbor search makes the encoding complexity O(n log n), and having only one dictionary this search can be done very fast even for large dictionaries.…”

“…Composed textures Our texture segmentation experiments are compared to [11,18] and the results are shown in Table 1. The proposed method performs better than existing methods in more than half of the samples, and follows closely in the rest.…”

“…We aim at adopting these properties in our method, which generally target discriminative problems. Sparse methods based on dictionary learning has previously been used for discriminative problems [11,12,15,18]. These methods are based on one dictionary for each class and the residual error as the discriminant.…”

“…This requires a dictionary for each class, which is computational costly, but these methods achieve very high performance. Initially the chosen dictionaries were optimized for reconstruction [15,18], but significant improvements have been shown by optimizing for discriminative power [11,12]. The procedure that we suggest is different from these methods, by using one dictionary for encoding an image and then inferring label probabilities from the label dictionary.…”

Dictionary Based Image Segmentation

“…High segmentation performance is obtained by utilizing that the texture class used for learning the dictionary can be reconstructed well whereas other texture classes cannot. Methods focusing on optimal reconstruction have been suggested [26,30], and improved performance has been obtained by also optimizing for discrimination [21,20]. Recently Gao et al [13] suggested to use sparse dictionaries together with an active contour for segmentation.…”