Food recognition improvement by using hyper-spectral imagery

In the past decade, dietary assessment has been one of the most popular topics of research in the food industry, which has resulted in developing several automatic or semi-automatic dietary assessment systems using visible spectrum images for food recognition. However, the main shortcoming of visible spectrum image-based systems is its inability to differentiate foods of similar color. Researchers have added additional features such as shape, size and texture to the color model to improve the overall accuracy. However, the shape and size features are rendered inefficient when recognizing food in the mixed or cooked form. The aim of this research is to show the capability of hyperspectral bands for accurate food recognition based on individual spectral bands. In this work we use a hyperspectral imaging system of 240 spectral bands with the wavelength range between 400 nm to 900 nm. The ReliefF and PCA methods select/extract less, but the most informative features which are important to learn Logistic Regression and Support Vector Machine (SVM) as binary and multiple classifiers, respectively. A total of 20 different food samples in various forms (uncut and cut), shapes, and sizes were used in this study. The prediction results indicate that the hyperspectral images have the advantages of being able to recognize different food items from a mixed form with similar color and similar food types with different colors. In our experiments the highest classification accuracy of 0.6874 with 20 different food samples is produced by SVM multi-classification of ReliefF data with the top 110 hyperspectral features. We are able to obtain approximately 0.90 accuracy, using binary classification on a specific subset of food samples.