AI object recognition

Object identification

Modern object identification system can work with complex objects

Computer Vision has emerged as a powerful technology over the last decade due to its ability to detect and recognize objects. The goal of object detection is to pull an object from the rest of the image and goal of object recognition to find out what this object is in its essence, i.e. classify it. Traditional computer vision methods didn’t allow us to detect or to recognize complex objects in a real-world environment. The only way was to compare the sequence of frames from video footage. If some cluster of pixels have changed, while background pixels stayed intact, such behavior was interpreted as a signal that there’s a moving “object” detected.

After artificial neural networks concepts came into computer vision, it became possible to determine objects within a single-frame based solely on the object’s shapes. The next big step forward was convolutional neural networks that allowed working with very complex objects, distinguishing them with high precision - cats from dogs, drones from birds, apples from tomatoes, cultivated plants from weeds and so on. Modern object recognition software sometimes works even better than a trained human does. And this powerful technology is now ready to serve almost all world industries, automating business processes, collecting data, and improving safety.

Each AI object recognition module designed to detect a particular set of objects

Thanks to the emergence of deep neural networks along with affordable GPUs high computing power, the object identification performance has truly skyrocketed. Yet, the current level of computer vision technologies doesn’t allow artificial intelligence working with arbitrary visual objects to identify objects as the live brain does. Every neural network module is designed to solve a specific task. Some neural networks detect and classify vehicles, while others classify birds. Data scientists have struggled to design a universal neural recognition module.

To overcome this technology limitation, BitRefine group has designed a multi-purpose recognition software platform as flexible pipelines with interchangeable neural modules, called “Heads”. For instance, when a task requires detecting people, the user loads the people recognition module into the pipeline; when it requires detecting weapons in x-ray images, the client loads corresponding x-ray recognition neural module. BitRefine Head’s pipeline is a sequence of image processing units. Typically, it starts with image acquisition, then there may be one or several preprocessing modules that prepare an image. Then the image is passed to the object recognition neural module that extracts and saves details about revealed objects. While working with videos, the next module usually is the tracker, that assigns an ID to each newly detected object and builds tracks of object’s movement. The tracker also plays the role of a verification agent that collects additional images of the same object. Also, it verifies them with the neural module to ensure that the initially recognized object’s class is correct. The next modules in the pipeline can be, for example, a color detector, various counters, etc. The high flexibility of BitRefine’s recognition software platform allows solving almost any client’s task, extract any required information from the video and save it as structured data into database (DB).

Multi-purpose object recognition AI platform includes library of recognition modules

BitRefine group offers a number of ready-to-use neural recognition modules that deal with standard everyday objects, such as people, faces, vehicles, etc. In addition to this, an important part of BitRefine’s services is designing and training custom recognition modules according to the client’s requirements. BitRefine specialists help clients to collect training data. Then we choose proper neural architecture from BitRefine’s library, train the module, and add required pre- and post-processing modules. In the end, the client receives a complete processing pipeline that takes an image from cameras and returns structured information to the DB or/and notifications.

There’re countless object types, that companies want to automatically recognize and track. BitRefine’s scalable architecture is designed in such a way that each recognition unit can work independently with different neural modules. This means that the whole system can recognize numerous types of objects at the same time. And this is one of the main strengths of multi-purpose object recognition software.

Related Solutions

AI object tracking

Tracking an object is crucial at numerous places such as airports, assembly lines, or stores. This task might seem to be simple because humans, as ...

Drone image processing

In recent years, Unmanned Aerial Vehicles (UAVs) have been adopted to provide services in numerous applications. Due to their ability to cover wide ...

Deep learning X-ray

X-ray security imaging is widely used to inspect luggage, mail, and vehicles to detect and discourage transport of illegal or dangerous items, such as ...

We’ve got more video analytics to share. Let’s be friends. But you first...