Vision systems in workflow automation applications are always about detecting objects or errors based on external features or monitoring certain scenarios. Optical image acquisition devices such as high-resolution cameras and sensors generate a large amount of digital image data for this purpose. This data must be processed and evaluated using software solutions in order to extract useful information and make it available for a wide range of vision applications. In order to achieve high recognition rates, the use of artificial intelligence (AI) is recommended in many cases.
While traditional approaches define and describe the relevant properties of the object to be recognized on the basis of sets of rules, AI works with self-learning algorithms. These comprehensively evaluate large image data sets as part of a training process and recognize certain recurring patterns. The software automatically learns specific features and characteristics of relevant object classes so that new image information and depicted objects can be precisely assigned to a specific class. This allows objects to be identified reliably and with a high degree of accuracy and localized precisely.

The advantages of such robust, AI-based recognition mechanisms are brought to bear in vision solutions such as 'Smart Vision' from Cloudflight. Smart Vision is a technology that uses AI to generate valuable information from image and video data. This process leads from the collection of image and video data with image acquisition devices through a pipeline of various AI modules for data processing and information extraction to the peripheral systems in which this information is used. This end-to-end approach integrates the entire process chain and thus prevents frictional losses and media disruptions.

Information through the AI pipeline

At the start of the data processing pipeline is a camera connector that receives the data in 2D or 3D from the image acquisition devices. The data is then sent to AI modules, which can be flexibly combined with each other depending on the application. One example of such an AI component is the object recognition module: it uses AI to recognize all objects of predefined types in an image and limit them with a bounding box. In machine production, for example, certain manufacturing defects such as painting errors or scratches can be defined as objects to be recognized and taught in. The module can therefore provide information on whether a machine part has a defect, where exactly the defect exists and what type of defect is present. This module is therefore particularly suitable for quality control in production.

In the background of the AI modules used, a higher-level system takes care of passing the data through the individual components and checks the results. Once the data processing is complete, a cloud connector forwards the information obtained to other links in the networked Industry 4.0 process chain, for example to merchandise management or production planning.

Smart sensors through neural networks

Learning algorithms form the technical basis for AI modules. Which specific algorithm is used in each case depends on the type and complexity of the task on the one hand, and on the other, runtime characteristics such as energy consumption and target hardware must be taken into account. In modern computer vision, neural networks are often used, which are further developed in iterative training processes with the help of suitable visual input so that they can automatically generate information from new image material. In object recognition, the YOLO ('You only look once') or R-CNN ('Region based convolutional neural networks') methods are among the best-known options for building neural networks.

With smart vision, optical image acquisition devices are transformed into smart sensors. With their help, data-driven, intelligent services can be implemented and processes can be highly automated - for example, quality control in industrial production. In addition to the object recognition described above, the optical sensors can also perform a variety of other AI-supported vision tasks, such as identifying people, determining the exact position and safe handling of workpieces, precisely measuring objects and observing and evaluating scenarios.

Holistic smart vision solution

Cloudflight offers a holistic service portfolio from a single source for the implementation of a smart vision approach in the company. This begins with advice on possible strategies and technical implementation in order to develop a viable digital business model. Proof-of-concept processes can be significantly accelerated through in-depth data analysis and corresponding AI modules based on pre-trained models. The in-house machine learning platform 'ModelCloud' ensures rapid training and iterative benchmarking of the AI models used according to the principle: train - test - repeat.