Combining powerful local and global statistics for texture description

Abstract

A texture descriptor is proposed, which combines local highly discriminative features with the global statistics of fractal geometry to achieve high descriptive power, but also invariance to geometric and illumination transformations. As local measurements SIFT features are estimated densely at multiple window sizes and discretized. On each of the discretized measurements the fractal dimension is computed to obtain the so-called multifractal spectrum, which is invariant to geometric transformations and illumination changes. Finally to achieve robustness to scale changes, a multi-scale representation of the multifractal spectrum is developed using a framelet system, that is, a redundant tight wavelet frame system. Experiments on classification demonstrate that the descriptor outperforms existing methods on the UIUC as well as the UMD high-resolution dataset.

A Handcrafted Normalized-Convolution Network for Texture Classification

“…Results on KTH-TIPS-2b: As can be seen from Table 4, our proposed method gets a significant improvement over the original features it inherits from, the enhancement in classification accuracy is approximate 10% compared with [40] 88.05 LBPHF [1] 89.58 Hayman et al [17] 92.00 CLBPHF [52] 92.55 MFS [48] 92.74 BRINT [25] 93.30 MRELBP (SVM) [24] 96.88 ScatNet(PCA) [41] 96.15 VZ-joint [44] 97.83 OTF [47] 97.40 scLBP [39] 98.45 FV-AlexNet (SVM) [6] 99.1 FV-VGGM (SVM) [6] 99.7 FV-VGGVD (SVM) [6] 99.8 NmzNet (ours) 96.97 Table 3. Classification accuracy comparisons with state-of-the-art on the KTH-TIPS-2a database.…”

Fast and Accurate Texture Recognition with Multilayer Convolution and Multifractal Analysis

“…It is well known that the multifractal spectrum encodes rich texture information. The methods in [10,11,12,13] use the box-counting method to estimate the multifractal spectrum. However, this method is unstable due the limited resolution of real-world images.…”

“…As the resolution r is supposed to be very small (r → 0), using small-sized boxes on a relatively lowresolution image results in a biased estimation due to the relatively low-resolution of real-world images [18]. It has been used as the core of various recent multifractal texture descriptors [10,11,12,13] that use the same box-counting method to build the final descriptor. We present a different method to statistically describe textures using multifractal analysis.…”

“…The motivation behind the proposed work comes from the success of multifractal analysis [10,9,11,12,13]. Given an input texture, the image is filtered with a small filter bank for various filter orientations.…”

“…Distributions over these textons are then compared using a proper distance and a nearest neighbor or kernel SVMs classifier [8]. Another alternative to regular histograms consists in using multifractal analysis [9,10,11,12,13]. The VG-fractal method [9] statistically represents the textures with the full PDF of the local fractal dimensions or lengths, while the methods in [10,11,12,13] make use of the boxcounting method to estimate the multifractal spectrum.…”