Exploring Deep Features and Transfer Learning for Plant Species Recognition

In recent years, with the evolution of the Convolutional Neural Networks, the automatic recognition of plant species from images became a very relevant research topic for scientists, researchers, and students in the field of botany. However, some problems related to the selection of features that best represent the characteristics of a particular species are still challenging due to the great variability of these characteristics within images from the same species and also the similarity of some characteristics between different species. In this sense, we propose a comparative study of Deep Convolutional Neural Networks to extract the feature vectors, here called "Deep Features", from the images of multi-organ plant observations. Moreover, eight variations of the Support Vector Machine (SVM) classifier were used for the assessment of the impact of three different Deep Features on the automatic image-based recognition of plant species. The evaluation protocol adopted for the classifiers was the Stratified 10-fold Cross Validation. As a result, the experiments demonstrate that higher dimensional Deep Features, in our case based on VGG-16 and VGG-19 networks, when exploited with the polynomial kernel SVM classifier and the One-vs-Rest decomposition method presented better classification effectiveness in the proposed study. Beyond it, this work highlights the fact that even in the context of transfer learning with deep features, the adequate selection of the baseline network is extremely important.