Image-based Nutrition Composition Analysis with a Local Orientation Descriptor

A local orientation descriptor (LOD) for nutrition analysis by quantity estimation is proposed. By observing nutrition properties, a texture-based LOD is designed to extract discriminant information, frequency and length among food items. Prior to classification, food detection is a challenging problem due to significant variety of backgrounds and containers. Thus, three food region detectors are designed in this study. A detector that employs a modified salient object detection algorithm using prior background knowledge provides promising segmentation results for non-uniform backgrounds. Integrating gradient information to construct graph weights yields more precise segmentation results. In addition, nutrition quantity is estimated using coins as reference objects. Three types of features, normalized colour, density, and symmetry properties are extracted for coin classification. Experimental results show that the proposed LOD outperforms existing object recognition features.