ELMP-Net: The successive application of a randomized local transform for texture classification

Abstract

This work proposes a method for texture classification based on the successive application of a local transform presented here for the first time. Such transform comprises two steps: (1) We built a two-layer mapping relating each pixel with its neighborhood, with the weights in the first layer randomly assigned; (2) We use the parameters learned by such mapping to transform the original image. Finally, we extract local descriptors at different stages of the successive application of this transform to compose the texture descriptors. The performance of our method is verified in the classification of benchmark texture databases and compared with state-of-the-art approaches. We also present an application for plant species identification. The results confirm our expectation that a model that is not based on the classical learning-based approach can still be competitive in texture analysis.