Wageningen UR Greenhouse technology group is profiling itself as a serious contributor to the rapidly growing field of automated plant phenotyping. The researchers have developed a fully automated method for separate leaves segmentation from RGB images using an independent, blinded from the ground truth and very variable datasets. This method brings accurate high-throughput phenotyping one step closer.
The dataset consisted of RGB top-view images of plants and was constructed as a part of the Leaf Segmentation Challenge (LSC) van de Computer Vision Problems in Plant Phenotyping (CVPPP 2014) workshop. Our novel, automated method outperforms all other methods that are tested on this highly challenging dataset with respect to some of the performance measures.
Per leaf derived growth
For a large-scale breeding, automated methods that measure relevant characteristics and that can deal with variable conditions in the constantly changing environment are essential. One of the most important traits is the plant growth. Determining the growth in an non-invasive manner from the images is a promising approach. For rosette plants non-destructive measurement via images of a plant’s projected leaf area (PLA), i.e. the counting of plant pixels from top-view images is considered a good approximation of plant size and is currently used. However, when considering growth, PLA reacts relatively weakly, as it includes growing and non-growing leaves, but the per leaf derived growth (implying a per leaf segmentation), has a faster and clearer response.
Measuring the growth of each individual leaf from 2D RGB images is a non-invasive, cheap and efficient way, but highly challenging from a computer vision point of view. The image resolution might be very low, presence of moss in the soil is often an obstacle, varying light conditions are always a challenge in agricultural applications but above all the overlapping leaves and the absence of edges between the separate leaves is the biggest challenge.
High processing speed
Wageningen UR Greenhouse technology group has developed an automated method for separate leaves segmentation, that is tested on an very difficult dataset of RGB images of Arabidopsis (Arabidopsis thaliana) and tobacco (Nicotiana tabacum). The plants were imaged in a deliberately difficult conditions with the variability in shape, pose, and appearance of leaves, but also there was a lack of clearly discernible boundaries among overlapping leaves with typical imaging conditions where a top-view fixed camera is used. This fact and the high processing speed make our automated method easily applicable in the real-world phenotyping problems.