Utilizing machine learning approaches to improve the prediction of leaf counts and individual leaf segmentation of rosette plant images

The segmentation of individual leaves in plant images is still a challenging task, especially in case of leaf overlaps. The exact determination of individual leaf areas could improve the biomass estimation which is a good indicator for plant performance. In addition, the number of leaves is directly related to plant development, leaf counts give insight into changing plant development stages. Machine learning is a powerful tool for solving vision tasks. Here we propose an approach including image analysis methods for extraction of a comprehensive set of image features (based on the software IAP) to predict the number of leaves for rosette plants. In addition, we developed a method to detect exact leaf borders for resolving inaccurate leaf segmentation in case of leaf overlaps. For classifier training we evaluate a broad set of different colour and texture features. The predicted leaf borders are used as input for further image processing methods to complete the leaf segmentation. The approaches were tested on image data of Arabidopsis thaliana and tobacco plants supplied in connection with the Leaf Counting Challenge (LCC) and the Leaf Segmentation Challenge (LSC) of the Computer Vision Problems in Plant Phenotyping (CVPPP) workshop in conjunction with the British Machine Vision Conference (BMVC) in 2015. The results show the methods ability for improved leaf count estimations and for predicting leaf overlap borders, which helps to improve the segmentation of individual leaves.