An Application for Automatic Classification of Unconventional Food Plants

Unconventional food plants are eatable species, having one or more parts with nutritional potential, but are not commonly used. These plants have attracted considerable attention, as more and more people have become interested and resorted to these natural resources. In order to be actually consumed, unconventional food plants should be known and disseminated. However, although there are many scientific researches dealing with plants identification in the literature, none of them addresses the automatic identification of these species. This work presents a study focused on the identification of unconventional food plants by means of two different strategies: 1) the classical combination of digital image processing-based feature generation and machine learning; and 2) CNN (Convolutional Neural Network) for feature representation and classification. To do so, the authors generated a database of selected species. The paper also details the process of database collecting, its constitution and representativeness, as well as the experiments performed on comparing the two investigated strategies. In the first strategy, we studied 17 features, including shape and texture features and employed Random Forest as classifier. Since we extracted the features from segmented leaves, the paper also details the segmentation process. Finally, the second strategy applies a CNN pre-trained on ImageNet. The comparative study showed that CNNs achieved lower false positive rates and higher and significant accuracy rates. These results show that computer-aided unconventional food plants identification systems are feasible, which may be important tools to allow non-experts to have access to such a valuable information, since the interested public is large and diverse, ranging from professionals to the general public.