Deep Learning for the classification of plants with natural colourants

Master’s thesis by Merlin Sudeepti Savalam, University of Eastern Finland

Biocolourants have a long history of usage and have been utilized by multiple industries, due to their eco-friendly nature and non-risk factors affecting human health. Compared to synthetic colourants, in spite of the limitations hindering the utility of biocolourants, major industrial sectors like food and cosmetics show an increasing trend to opt for them, as they are deemed relatively safe by consumers. However, the inadequate information on the available sources of naturally obtained biocolourants has been a barrier to promoting their utility. Deep learning is a potential machine learning tool for effectively processing large amounts of data, and it has been successfully applied in plant classification and identification with higher accuracy than the traditional botanical taxonomy. Plant species from Magnoliopsida and Pinopsida, the two plant classes containing plants with biocolourant properties, are examined in our work. This study is aimed to classify the plant species from these classes by using two deep learning models, convolutional neural network (CNN) and residual neural network (ResNet). The primary objective is to build an efficient deep learning approach based on the given datasets.

The CNN with a batch size 64 has achieved an accuracy of more than 95%, as illustrated in Fig. 1.

Figure 1. Predicted results with CNN Model

Figure 2 demonstrates that the ResNet with a batch size 64 can yield an accuracy of more than 97%.

Figure 2. Predicted results with ResNet Model