Deep Learning Add-on is a breakthrough technology for machine vision. It is a set of five ready-made tools which are trained with 20-50 sample images, and which then detect objects, defects or features automatically. Internally it uses large neural networks designed and optimized by our research team for use in industrial vision systems.
Together with Aurora Vision Studio you are getting a complete solution for training and deploying modern machine vision applications.
Why deep learning from Adaptive Vision?
Typical applications require between 20 and 50 images for training. The more the better, but our software internally learns key characteristics from a limited training set and then generates thousands of new artificial samples for effective training.
A modern GPU is required for effective training. At production, you can use either GPU or CPU. GPU will typically be 3-10 times faster (with the exception of Object Classification which is equally fast on CPU).
Typical training time on a GPU is 5-15 minutes. Inference time varies depending on the tool and hardware between 5 and 100 ms per image. The highest performance is guaranteed by WEAVER, an industrial inference engine.
How is it different from TensorFlow or PyTorch?
TensorFlow and PyTorch are low-level frameworks for programmers and for passionates. If you have your R&D team, you may want to use these tools to build a solution from scratch. It may take a couple of days to create a demo, but at least several months to have a production-ready system, and even more to achieve the highest possible performance. On the other hand, our product is a complete solution, field-tested in over 100 projects and you can use it yourself today.
Why should I use Adaptive Vision if I could use an open source neural network?
We provide you a complete solution – it consists of five optimized neural network designs, but also of: graphical tools for easy data annotation and training, advanced augmentations and automatic balancing of training data, mix of traditional and machine learning methods for data preprocessing, optimized memory management, industrial-grade inference engine for CPU and GPU, tools for deployment, technical support + know-how. We spent many man-years in developing, testing and fine-tuning all of that so that you can bring it to your project instantly, with reasonably small training set, with performance much above the open-source frameworks and at a low cost at the same time.
In the supervised mode the user needs to carefully label pixels corresponding to defects on the training images. The tool then learns to distinguish good and bad features by looking for their key characteristics.
In this application cracks and scratches must be detected on a surface that includes complicated features. With traditional methods, this requires complicated algorithms with dozens of parameters which must be adjusted for each type of solar panel. With Deep Learning, it is enough to train the system in the supervised mode, using just one tool.
Satellite images are difficult to analyse as they include a huge variety of features. Nevertheless, our Deep Learning Add-on can be trained to detect roads and buildings with very high reliability. Training may be performed using only one properly labeled image, and the results can be verified immediately. Add more samples to increase the robustness of the model.
In the unsupervised mode training is simpler. There is no direct definition of a defect – the tool is trained with Good samples and then looks for deviations of any kind.
When a sushi box is delivered to a market, each of the elements must be correctly placed at a specific position. Defects are difficult to define when correct objects may also vary. The solution is to use unsupervised deep learning mode that detects any significant variations from what the tool has seen and learned in the training phase.
Injection moulding is a complex process with many possible production problems. Plastic objects may also include some bending or other shape deviations that are acceptable for the customer. Our Deep Learning Add-on can learn all acceptable deviations from the provided samples and then detect anomalies of any type when running on the production line.
The Object Classification tool divides input images into groups created by the user according to their particular features. As a result the name of a class and the classification confidence are given.
Plastic caps may sometimes accidently flip in the production machine. The customer wants to detect this situation. The task can be completed with traditional methods, but requires an expert to design a specific algorithm for this application. On the other hand we can use deep learning based classification which automatically learns to recognize Front and Back from a set of training pictures.
There may be hundreds of different alloy wheel types being manufactured at a single plant. Identification of a particular model with such quantities of models is virtually impossible with traditional methods. Template Matching would need huge amount of time trying to match hundreds of models while handcrafting of bespoke models would simply require too much development and maintenance. Deep learning comes as an ideal solution that learns directly from sample pictures without any bespoke development.
The instance segmentation technique is used to locate, segment and classify single or multiple objects within an image. Unlike the feature detection technique, this technique detects individual objects and may be able to separate them even if they touch or overlap.
Mixed nuts are a very popular snack food consisting of various types of nuts. As the percentage composition of nuts in a package shall be in accordance with the list of ingredients printed on the package, the customers want to be sure that the proper amount of nuts of each type is going to be packaged. Instance segmentation tool is an ideal solution in such application, since it returns masks corresponding to the segmented objects.
A typical set of soup greens used in Europe is packaged on a white plastic plate in a random position. Production line workers may sometimes accidently forget to put one of the vegetables on the plate. Although there is a system that weighs the plates, the customer wants to verify completeness of the product just before the sealing process. As there are no two vegetables that look the same, the solution is to use deep learning-based segmentation. In the training phase, the customer just has to mark regions corresponding to vegetables.
The Point Location tool looks for specific shapes, features or marks that can be identified as points in an input image. It may be compared to traditional template matching, but here the tool is trained with multiple samples and becomes robust against huge variability of the objects of interest.
The task that seems impossible to achieve with traditional methods of image processing can be done with our latest tool. In this case we use it to detect bees. When it is done we can check whether they are infected by varroosis – the disease caused by the parasitic mites attacking the honey bees. The parasite attaches to their bodies and upon the basis of a characteristic red inflammation we can classify them according to their health condition. Not only does this example show that it is an easy solution for a complex task, but also that we are open to many different branches of industry e.g. agriculture.
In these applications we need to guide a robotic arm to pick up items, most typically from a conveyor belt
or from a container. A good example of such application is picking small stem cuttings and then placing them vertically in pots. Any inaccuracies in detection may result in planting them too deep or upside down, which will result in cuttings not forming roots. Our deep learning tools make it possible to quickly locate the desired parts of the plants and provide accurate results required for this operation.