Analysis of the Indented Cylinder by the use of Computer Vision

The research summarised in this PhD thesis took advantage of methods from computer vision to experimentally analyse the sorting/separation ability of a specific type of seed sorting device – known as an “indented cylinder”. The indented cylinder basically separates incoming seeds into two sub-groups:
(1) “long” seeds and (2) “short” seeds (known as length-separation). The motion of seeds being physically manipulated inside an active indented cylinder was analysed using various computer vision methods. The data from such analyses were used to create an overview of the machine’s ability to separate certain species of seed from each other. Seeds are processed in order to achieve a high-quality end product: a batch of a single species of crop seed. Naturally, farmers need processed clean crop seeds that are free from non-seed impurities, weed seeds, and non-viable or dead crop seeds. Since the processing is based on physical manipulation of the seeds themselves, their individual shape and size becomes very relevant. The problem of modelling such physical parameters for various species of seed, grown under various environmental circumstances, is a very complex one. The general problem of modelling and controlling seed processing equipment can be expected to be complex. Due to the involvement of seeds, the problem will naturally inherit all their biological complexities. In addition to this, because of the very large number of individual seeds, the problem also involves a granular media and thus inherits all the complexities related to that as well.
The project arrived at a number of results of high scientific and practical value to the area of applied computer vision and seed processing and agricultural technology in general. The results and methodologies were summarised in one conference paper and two journal papers. These three papers, referred to as Paper I, Paper II, and Paper III can be found in Appendix A, B, and C, respectively. These three papers represent the very first examples of published/submitted work that thoroughly analyse and verify the separation ability of the indented cylinder by the use of computer vision (or image analysis). Moreover, the imagery data sets, generated as a result of actual recordings of sorting experiments using the indented cylinder, are novel by their high dimensionality and size. Paper II in Appendix B makes one of these data sets available online as a cite-aware imagery data set. The work summarised in this thesis is very much related to the task of constructing models from observed data. This field is known as empirical model development or more specifically as “system identification”. System v identification deals specifically with estimating mathematical models from observed dynamic states (time series) of inputs and outputs to and from some physical system under investigation. The contribution of the work is to be found primarily within the problem domain of experimentation for system identification. Computer vision techniques were used to acquire observations of a measure of separation efficiency of the indented cylinder. Such techniques for observation are likely to be very relevant for experimentation in a laboratory for system identification purposes. This work should therefore be seen as an important step towards future research in system identification of the indented cylinder. The technical solutions developed are currently novel and represent an ideal platform for future applied research into empirical model development. Finally, this work should also be considered as an early step toward a paradigm shift where the best parameters for the indented cylinder are not mainly determined by “rule of thumb” and other forms of heuristics, but are instead optimized parameters tied to an actual theory of seed separation in the indented cylinder.