Insects are important both in terms of the functioning of the Earth’s ecosystem and in terms of their impact on human society. Accurate identification of a particular type of bag is often not easy and it often takes time. So the Vdc group together with Dominik Vondrek from the Department of Zoology at Charles University College of Science focused on the possibility of creating a specific system automatically using modern technologies.
Nowadays, more than a million species of insects are known, and within one species one can often observe the known variation due to gender, developmental stages and color. Its accurate identification is very important. Insects, known for their biodiversity, play an important and often indispensable role in the functioning of individual ecosystems on Earth. On the other hand, there are many types of insects that are the origin of various diseases or agricultural crops.
For a specific type of knowledge, it is essential to have a long-term, expert knowledge of the matter, and at the same time identification often takes a long time. Therefore, in the past, efforts have been made to automate and simplify insect recognition, which was often imprecise and required a lot of input data. With the development of new methods called spinning nerve centers to enable accuracy and solve most of the problems with automatic insect recognition.
The CNN is a practical image recognition tool, which needs to reduce the input parameters. The chain is automatically converted to vector form automatic processing, which enters the neural networks it works with and is categorized.
Since this is a promising method, scientists have focused on determining the accuracy of automatic insect recognition first with the help of a convoluted neuronal states. They used three different photographic data sets for their research. The first two data sets, consisting of images of different species, were used to determine whether the automatic system could learn to recognize different groups of beetles and biceps beetles. The remaining two datasets were exchanged to identify individuals at the species level.
The first dataset contained 884 individuals from two channels of 11 elks imaged from a frontal view of the head, and the mean here was 92.7%. In beetles (2,936 different individuals out of 14 icicles photographed from a dorsal view), the accuracy was 96.1%. In the other two data sets, the researchers attempted to test the accuracy of automatically recognized species, which is very similar and difficult to identify even for human experts. In this case, 97.3% were spontaneously correctly detected for the three related goldfinch species of the genus Oxythyrea (339 individuals, hilly landscape) and 98.6% for nine larval stalk species (3,845 photographed from a single sample from different perspectives).
Research results are very satisfactory, or they have proven that with the help of modern technologies it is possible to identify specific types of insects with an accuracy of more than 90%. This was possible even in very small datasets, and the results could be compared with evaluations of expert entomologists. By using convolution of neural networks, it will be possible to fully automate the specific taxonomic system.
The use of a bundle of neural networks in the field of biodiversity is still in its infancy, but today it can be observed in a known way. The authors point out that with the help of this method, in the future, in addition to identifying the existing species, it will be possible to discover non-described species, or so-called hidden species, which differ only in a few clear features and are usually recognized, especially DNA .
Valan, M., Makonyi, K., Maki, A., Vondrek, D., Ronquist, F. (2019): Automated taxonomic identification of insects with precision at an expert level using efficient feature transfer from convolutional networks. Systemic Biology 68 (6): 876895.