Exploring Statistical GLCM Texture Features for Classifying Food Images

Image processing algorithms, which recognize foods from photographs, have been proposed as a method for patients to conveniently track foods they eat. However, food recognition is challenging because 1) the same type of food can differ in shape when prepared differently. 2) the wide variety of food leads to a large number of food categories. Machine learning based food classification has been proposed but its accuracy depends on the types of visual features used. In this work, we compare the accuracy of SIFT to GLCM features for recognition of 4 different categories of foods (apples, burgers, bread and prepared dishes). GLCM features perform better than SIFT overall, and for single-item or homogeneous foods (e.g. Apple) while SIFT performs well for complex dishes.