Brain Builder Knowledge Base
Best Practices: Recognition Brain Training
Direct link to topic in this publication:
Success with Brain Builder Recognition Brains
There are a wide variety of applications for AI vision systems, and there are many types of AI vision solutions that can be used to solve them. These different AI solutions will perform better or worse depending on the details of the use case. To be successful with a Brain Builder Recognition Brain, it's important to make sure your images and use case are the right fit.
Recognition Brains are a simple, powerful type of AI that analyze a whole image and determine if an object on which they have been trained is present in the image. You may have seen Recognition AI systems in your cell phone (classifying photo scenes such as "sunset," "beach," and "portrait") or in novelty websites that try to distinguish different breeds of dogs.
One of the benefits of Recognition is that the labeling is a simple tag applied to the whole image. No bounding box or more detailed annotation are required.
Recognition Use Cases & Best Practices
Use the whole image
Because Recognition systems analyze the whole image, the Recognition Brain will perform best when it's used on images in which the image is predominantly associated with the target class you want the Brain to learn and predict.
A good photo of an electrical insulator
A bad photo of electrical insulators (for a Recognition Brain)
The insulator is prominent and clearly visible in the image.
The insulators are small and not prominently featured in the photo.
One class per image
Brain Builder builds Recognition Brains that are trained to learn a single class for each image. The Brains will not perform well if they are trained with or used on images in which multiple relevant classes of objects occur.
The best use cases for Brain Builder Recognition Brains are those in which the images are very consistent from one to the next and the only significant differences are those that are relevant for distinguishing the different classes.
For example, industrial use cases in which the camera is installed in a fixed position and takes the same types of image of each object on a conveyor belt are a great use case. By keeping everything the same except the objects on the belt, the Recognition Brain can quickly learn the differences between objects and understand that those differences are important for distinguishing between different classes.
Recognition Brains can still be used in less controlled environments, but more training might be necessary to achieve good results.
For example, Brain Builder can be used for drone inspections of infrastructure equipment (such as electrical insulators) to identify which objects are broken and which are not. Because the camera perspective and the area around each insulator changes in each image, Brain Builder needs a larger set of training data to understand the details that distinguish broken insulators and filter out the "noise" of the other changes in each image.
Next: Inference API and SDK >>