Selection of parameters in active contour models for automated plant phenotypic analysis

Modern genomics experiments are often conducted in controlled environments on hundreds of plants at a time. As such, manual and destructive phenotyping techniques are no longer suitable. Recently there has been a trend toward automated high-throughput phenotyping facilities, in an attempt to relieve the phenotyping bottleneck. A majority of these facilities focus on the use of cameras for non-destructive recording of plant data. Hence, the new challenge lies in accurately segmenting the plants from 2D images in an automated manner and then analysing structural and statistical information about them. One such technique for image segmentation is known as the active contour model. Active contour models have seen widespread success throughout a number of applied fields due to their versatility and semiautomated nature. However, a high majority of these models rely on arbitrary parameters that are required to be selected manually. Furthermore, small variations in these parameters can produce substantial variations in the method’s overall accuracy. This makes them unsuitable for use by non-experts and also for the analysis of a series of images that can change significantly over time. For example an image sequence for the growth of a plant. In this paper we attempt to establish relationships between the parameter values of active contour models and the geometry of the objects/shapes that they are segmenting. Our goal is for users to be able to utilise some basic a-priori knowledge about the geometry of the object in order to automatically select a range of suitable parameter values. We analyse the accuracy of active contour models over multiple series of shapes that exhibit some pattern, such as decreasing number of sides or increasing concavity. We present a novel normalization technique so that the parameter values are of a similar scale. We also carefully design an experimental setup that ensures no bias between different shapes or parameter values. We show that over a series of shapes the range of parameters that provide convergence do follow a trend. We also show that not all contours that converge to the objects boundary do so in a stable manner, with a substantial amount oscillating continuously. However more information, such as more shapes and more parameter values, is required to draw meaningful and quantitative conclusions from such an analysis. Future work includes incorporating more of this information along with the application to more active contour models. Another exciting future direction is the use of 2D shape diagrams to quantify relationships between shapes, parameter values and levels of accuracy