Repeating average filter for noisy texture classification

Abstract

Abstract. In this paper, it is shown that repeating average lter increases the uniform patterns of noisy textures and, consequently, increases the classi cation accuracy of textures. In other words, for noisy textures, rst, an average lter, such as 3 3 mean lter, is applied to each image; then, a feature extraction method, such as LBP, is used to extract features of the ltered image. The more value of noise, the more repeating of average lter should be applied to textures. Moreover, it is shown that by repeating the 3 3 average lter for textures, the variance of texture decreases, then increases. Thus, average lter must be repeated while the variance of image decreases and when the variance starts increasing, it must be stopped. Using convolution to apply average lter for an image takes so much time; therefore, a simple technique is proposed in this paper that increases the speed of average ltering signi cantly. After noise reduction, by using LBP operator, features of texture are extracted for classi cation. Implementations on Outex, CUReT, and UIUC datasets determine that the performance of the proposed method is better than that of some advanced noise-resistant LBP variants such as BRINT and CRLBP.

Weighted Constraint Feature Selection of Local Descriptor for Texture Image Classification

“…The key point of texture processing is related to extracting a set of discriminative features. In the last two decades, many feature extraction approaches [7][8][9] have been introduced that are used for texture images for classification [10][11][12][13][14] or segmentation [15,16]. Most of these techniques employed statistical local descriptors for texture analysis.…”