Seed identification using machine vision: Machine learning features and model performance

Machine vision offers an alternative to manual identification of seeds with the opportunity of automation and additional information about seed quality aspects like maturity, viability and seed health. In this paper the visual properties that humans perceive in seeds are bridged with machine vision features which have proven useful for the identification of seeds. While morphological and color features bear a high resemblance to human-perceived features, textural features exhibit a higher level of abstraction and spectral features represent an extra dimension of color beyond the human color perception. Hence, they do not correspond as nicely to human perceived properties of seed. Natural within-species variation can be observed within all groups of features and are influenced by factors of genetic, physiological and environmental character. Nevertheless, the reviewed studies on taxonomical classification of seeds have shown that features within all groups are all valuable for the identification of seeds. However, care should be taken not to include too many features in a model as this comes with the cost of the risk of overfitting because of redundant and noisy data and a cost of more calculation power needed. Most studies have focused on a limited number of species, often addressing specific issues such as cereal classification or coffee adulteration, achieving high accuracy rates, even up to 100 %. However, a small number of studies have included a broader range of species (up to 236) while still achieving accuracies above 93 %. While these accuracies cannot be directly compared with the accuracies of humans as human performance is generally evaluated differently, the achieved performances suggest that the technology has now matured enough to be able to assist in real-world seed testing. For the technology to get accepted by large international organizations like the International Seed Testing Association (ISTA) and the Association of Official Seed Analysts (AOSA), stakeholders need to prove that the technology can compete with the accuracies of traditional seed analysts. For this, more research based on recognition of a high number of relevant species using test sets resembling internationally recognized proficiency tests is needed.