The better your training image set represents the labels (i.e., the characteristic feature state of the spp), the better the identification results will be. It is particularly important to train your AI on the same type of images (that is, unusual flower, leaf, features) you expect it to receive for identification requests.
For example, creating an image training set with the same background, viewing angle and perfect lighting conditions will produce an AI that would likely result in very poor recognition of images received from the field where the background, lighting and viewing angles can be highly variable.
If the entities you want the AI to recognize are highly variable, it may be better to split the entity into several labels to train on separately, rather than as one entity. For example, if you were wanting to train an AI to recognize some arthropods, it may be better to create labels that represent each distinct instar/life cycle stage for the arthropod, if you weren’t just targeting the adult stage. Take the Coconut rhinoceros beetle (Oryctes rhinoceros) life cycle. It has three very different stages, excluding the egg stage, three instar grub stages, pupa and the adult. Entity labels for this species could be designed as follows:
Coconut rhinoceros beetle (Oryctes rhinoceros) grub
Coconut rhinoceros beetle (Oryctes rhinoceros) pupa
Coconut rhinoceros beetle (Oryctes rhinoceros) adult
It’s also important to describe the aims and the kind of image data you’ve used to build your AI, along with what it’s trying to classify. In some ways an AI can be likened to a hammer looking for a nail. If a user provides an image for identification that’s well beyond its scope (nothing like its training data) or labels, it will still try and classify it, often with puerile results.
It would also be useful to tell the user the best conditions, camera angles and any other relevant information about the type of images they should submit to the AI to get the best results. This information can be added to your project description (See the ‘AI Project’ help topic below for further information, including an example ‘best use’ description).