Pollen in honey reflects its botanical origin and melissopalynology is used to identify origin, type and quantities of pollen grains of the botanical species visited by bees. Automatic pollen counting and classification can alleviate the problems of manual categorisation such as subjectivity and time constraints. Despite the efforts made during the last decades, the manual classification process is still predominant. One of the reasons for that is the small number of types usually used in previous studies. In this paper, we present a large study to automatically identify pollen grains using nine state-of-the-art CNN techniques applied to the recently published POLEN73S image dataset. We observe that existing published approaches used original images without study the possible biased recognition due to pollen’s background colour or using preprocessing techniques. Our proposal manages to classify up to 97.4 % of the samples from the dataset with 73 different types of pollen. This result, which surpasses previous attempts in number and difficulty of pollen types under consideration, is an important step towards fully automatic pollen recognition, even with a large number of pollen grain types.
Pollen in honey reflects its botanical origin and melissopalynology is used to identify origin, type and quantities of pollen grains of the botanical species visited by bees. Automatic pollen counting and classification can alleviate the problems of manual categorisation such as subjectivity and time constraints. Despite the efforts made during the last decades, the manual classification process is still predominant. One of the reasons for that is the small number of types usually used in previous studies. In this paper, we present a large study to automatically identify pollen grains using nine state-of-the-art CNN techniques applied to the recently published POLEN73S image dataset. We observe that existing published approaches used original images without study the possible biased recognition due to pollen’s background colour or using preprocessing techniques. Our proposal manages to classify up to 97.4% of the samples from the dataset with 73 different types of pollen. This result, which surpasses previous attempts in number and difficulty of pollen types under consideration, is an important step towards fully automatic pollen recognition, even with a large number of pollen grain types.
