Comparative analysis of different features and encoding methods for rice image classification

Abstract

Rice is the most widely consumed staple food in Sri Lanka. In this paper, we present a comparative study of different features (SIFT, Multi-resolution Local Patterns, Local Color Histograms, and Random Projections) and feature encoding approaches (Bag-of-visual-words, Sparse Coding, Vector of Locally Aggregated Gradients, and Fisher Vectors) for classifying images containing rice grains. By analysing the performance of a classification model with two-fold cross validation on a dataset of 1000 images containing ten rice categories, we show that SIFT features with Fisher Vector encoding or with Vector of Locally Aggregated Gradients produces the best result (mean class accuracy of 97:9 0:5). We found that increasing the size of the dictionary generally improves the classification performance for all the feature encoding approaches. The dataset we use is made public, and it can be accessed via http://www.csc.jfn.ac.lk/ index.php/dataset/.