A Training-free Classification Framework for Textures, Writers, and Materials

We propose a training-free texture classification scheme, outperforming methods that use training. This we demonstrate not only for traditional texture benchmarks, but also for the identification of materials and writers of musical scores. State-of-the-art methods operate using local descriptors, their intermediate representation over trained dictionaries, and classifiers. For the first two steps, we work with pooled local Gaussian derivative filters and a small dictionary not obtained through training, resp. Moreover, we build a multi-level representation similar to a spatial pyramid which captures region-level information. An extra step robustifies the final representation by means of comparative reasoning. As to the classification step, we achieve robust results using nearest neighbor classification, and s-o-a results with a collaborative strategy. Also these classifiers need no training. Standard texture classification systems aim at i) constructing a rich representation of the image and ii) providing a classification strategy. The representation typically entails local (texture) descriptors, similarity measures, aggregating strategies, and intermediate and global (image level) descriptors. The classification strategy usually adapts its metric to the representation and aims at fixing its flaws. Class models are then built using state-of-the-art classifiers. A literature review is in the paper. We propose a training-free multi-level texture classification framework (Fig. 1). It combines the robustness and simplicity of local descriptors such as BIFs [1], spatial information embedding into the global image representation in a layered fashion similar to SPM or through regions as in [3], the power of comparative reasoning [8], and s-o-a training-free classifiers [6]. Fortes of the framework are: 1) No need for training, and thus data independence. There is no need for learning a dictionary for the local descriptors (such as BIFs [1]). The system performs robustly with a fixed set of parameters on different texture, material and handwritten score datasets. 2) Robustness to intra-class variations. Robustness is provided by the local descriptors, the layered robustified representation, and the classifiers. 3) Layered representation embedding spatial information. Spatial information proved critical for object classification, and so it is for our tasks. 4) Robustified representations by means of comparative reasoning. The power of comparative reasoning (WTA-hash [8]) enhances and robustifies the representations by adding resilience to numeric perturbations. 5) Fast sparse and/or collaborative classification. Lately, sparse and collaborative representation based classifiers performed best at various tasks such as face recognition or traffic sign recognition [6]. Local Texture Descriptor (BIF) Basic Image Features (BIF) [1, 5] are defined by a partition of the filter-response space (jet space) of a set of 6 Gaussian derivative 2D filters up to 2nd order at some scale σ . The Jet space is further partitioned into 7 regions, or BIFs, corresponding to distinct types of local image symmetry. BIFs are rotation invariant. However, we can discretize orientations for BIF codes as in [5], thus obtaining Oriented Basic Image Features (oBIF). To create a more discriminative descriptor, [1] combines the descriptors at different scales on a pixelwise basis and ignores flat regions. BIF with p scales will generate 6p distinct dictionary entries (for p = 4, 1296), while oBIF will generate 22p distinct dictionary entries (for p = 2, 484). Multi-Level Pooled Representation (SPM, BoR) The spatial pyramid matching (SPM) scheme pools regions at 3 or 4 pyramid levels. We continue as long as the cell/region size allows for meaningful histograms. Table 1: Summary of texture, material, and score datasets.