What is deep learning?
It is an automatic hierarchical or structured algorithm that emulates human learning with the aim of achieving certain knowledge. The system is capable of learning by itself through a previous “training” phase, it is not necessary to be programmed.
Furthermore, it is characterized by processing information through the artificially intertwined neural networks of which it is composed. It is primarily used for the automation of predictive analytics.
Deep learning constitutes one of the fundamental bases of artificial intelligence (AI) and given the great boom that it has had in recent times, interest in deep learning has also increased.
Deep learning techniques have improved the ability to classify, recognize and detect, in short, have given a great boost to the ability to understand.
The classification of images based on deep learning, allows to carry out classification tasks of products or defects through artificial vision techniques.
Through previous learning of the system, and the configuration of a convolutional neural network, with relatively short configuration times, the classification of new products or effects is carried out quickly and efficiently.
By means of semantic segmentation, the trained defect classes can be located with pixel precision. This allows users, for example, to solve inspection tasks, which previously could not be done, or only with significant programming effort.
Today’s electronics make it possible to integrate classification software into embedded systems, enabling the development of compact, low-cost equipment.
Object and defect detection
Object detection locates the classes of trained objects and identifies them with a surrounding rectangle (bounding box). Objects that touch or partially overlap are also separated, allowing for object counting.
We can align these rectangles according to the orientation of the object, which results in more accurate detection, as the rectangles are more closely aligned with the shape of the object.
This technology is very useful when classifying different products that do not have an identifying label or when identifying defects that follow random patterns of appearance and location.
Automatic product sorting
In certain food processes in the meat sector where foods are slaughtered in line, it is common for manual product identification points to be installed at the end of cutting lines. At these points, an operator must classify the product coming from the cutting process and decide which subsequent processing line to go to.
This task is tedious and expensive, and can be replaced by a trained vision system for the automatic classification of products.
In addition, the artificial vision system allows other quality control tasks to be implemented.