Deep Learning, when the solution is a neural network
Specialvideo is expanding its knowledge on the topic of Artificial Intelligence and Deep Learning applying these research topics to industrial applications. This learning process, which started with the study of the theory and open source libraries of Tensorflow, led to the introduction of neural networks within our systems.
For each application, we study the problem and choose the neural network model that we consider most effective. We are able to adapt and customize existing neural network models, but at the same time we can create and implement new structures that allow the completion of a specific task.
The field of applications for deep learning is very broad and covers almost entirely the range of tasks that a traditional algorithm can perform.
Neural networks can be used for object recognition, to solve classification problems. Classification can regard two or more classes, without affecting the efficiency of the software. Practical cases for this application can be the sorting of products with different characteristics or quality control, for which the neural network learns to recognize the defective parts.
Localization (object detection)
Another task in which the deep learning approach can be exploited is robot guidance. Indeed, there are neural networks specialized in the localization of one or more types of objects. The identification of the product occurs through bounding boxes that can also be oriented, i.e. indicate the direction in which the object is placed.
Individuation of anomalies (anomaly detection)
The search of anomalies also, can be an ideal case for a neural network application. For instance, there are some models capable of learning and generalizing the characteristics of products that conform to a certain standard, identifying the presence of possible anomalies. Another analogous case can be the search of anomalies inside a regular or irregular pattern of a surface.
Classification and search of defects (segmentation)
The image is divided into parts, that is to say, it is segmented, through the identification of the outlines of the various objects represented in it and a possible subsequent classification. These types of neural networks can be used to implement the subdivision of a surface based on its characteristics or to process some parts of the image recognized as similar to each other.