AI machine vision

Visual inspection deep learning
food visual inspection deep learning

Machine vision technology is used to extract information from an image to provide automated visual inspection, process control, robots or vehicles guidance. Starting from 2013 deep neural networks and machine learning has pushed the limits of what was possible in this domain and become today the standard approach for the whole field of computer vision. For the automated visual inspection, which is the most popular task for the machine vision, the main advantage of deep learning technology over traditional computer vision is that deep learning systems allow detecting irregular defect types such as cracks, scratches, patches, stains, debris, inclusions, damaged edges, pierces, etc. In other words, manufacturers can automate many inspection routines that until now could to be done only manually.

Deep learning visual inspection to automate routines that until now could be done only manually

In order to make a neural module recognize your target defect it needs to be trained first. Modern artificial neural networks contain from several hundred thousands to tens of millions of internal parameters that need to be adjusted properly to make neural network work. (By comparison, human brain has about ten billion neurons) During training special algorithm loads images of defects’ samples and change neurons’ parameters until the whole network starts recognizing samples properly. True power of neural networks is their native flexibility. It allows detecting objects and patterns that not exactly the same as presented samples but look just similar enough. For example, you collect certain number of sample images of scratches, train the network and it will be recognizing most of the new scratches of various shapes, length, thickness, etc.

There’re three levels in recognition process. The first one is object detection or simplest object classification: recognition software receives image and returns answer, OK or Not OK. The next one includes localization: recognition software returns coordinates where detected objects located. This can be center of an object or its bounding box. Finally, the third level is segmentation, when recognition software finds and shows exactly which pixels form the object. Some machine vision task require only the first level of precision, others, such as medical imaging, usually work with the third level only. Automated visual inspection task usually require the first or second level limited to bounding box precision.

AI machine vision detecting scratches

Multi-purpose recognition platform BitRefine Heads is capable of working at all three levels of presicion and is successfully used in the visual inspection area. The core of recognition platform is interchangeable neural modules. BitRefine offers the whole line of neural modules for various tasks. As every AI machine vision project usually requires detection of special unique objects or patterns, specialists at BitRefine group offer services of selection of the most suitable network architecture from its library and retraining it according to client’s requirements. This end-to-end approach ensures best possible recognition accuracy. After neural module achieved target performance, it gets installed into platform’s image processing pipeline and start doing its job along with other possible modules. BitRefine Heads processing pipeline is a sequence of operations starting from image acquisition and ending with generating system reaction to detected object and saving data into DB. Pipeline can be setup flexibly and include neural network module as well as traditional computer vision tools.

Machine vision data for manufacturers

High flexibility of BitRefine Heads recognition platform offers clear benefits: it can work with typical still images from machine vision cameras with constant light condition and clear background. It also has tracking capabilities and can work with video, track detected objects in complex real-world environment. Due to BitRefine Heads modular expandable architecture the whole number of company’s automation and safety tasks can be solved by single platform with unified control interface.

Manufacturers has been always aiming automating as many internal processes as possible and visual inspection is one of the top tasks in this list. Today deep learning in combination with traditional computer vision techniques offers almost unlimited opportunities for inspection automation. BitRefine’s role is to make transition from semi-automated to fully automated inspection as seamless as possible.

As a data science company BitRefine group offers further integration of big data, collected from visual sources, into corporate data-lake as part of data-driven business transformation

Related Solutions

OCR system

Can a machine read? About a century ago, Emanuel Goldberg, a well-known physicist, developed a reading machine for blind individuals. This invention ...

We’ve got more machine vision to share. Let’s be friends. But you first...