A benchmarking: Feature extraction and classification of agricultural textures using LBP, GLCM, RBO, Neural Networks, k-NN, and random forest

Abstract

Agricultural textures are in the interest of classification in image processing. Natural images have unique textural shapes inside which cause a tough problem for classification. This paper tests different feature extraction and classification approaches to serve a benchmarking on several agricultural databases like seeds and leaves. Features are obtained using Local Binary Pattern (LBP), Gray Level Co-Occurrence Matrix (GLCM), and Relational Bit Operator (RBO) independently. Classification is done by Neural Networks, k-nearest neighbor method, and random forest independently, too. LBP counts several binary patterns that occur in the image. GLCM is a kind of statistical approach that uses homogeneity, contrast, energy, and correlation information from pixels. RBO counts the binary relations of neighboring pixels in a box filter to get textural features for image processing. The leading test results are obtained from the LBP method for features and random forest data structure for classification. For example, agricultural seed type classification is obtained with LBP features and random forest classification with an accuracy of 99.5% and leaf classification with 93.5% accuracy. Following sections in the paper start with an introduction and continue with literature review, methods and materials, test results and conclusion.