NeuroCheck as a Platform

The Basis for your Deep Learning Application

Deep learning is becoming an increasingly important technology in the manufacturing industry. That‘s why NeuroCheck has been providing in-house developed backpropagation networks for image content and object feature classification for over 25 years. The necessary neural network structures have been an integral part of the NeuroCheck image processing software since 1994. With increasing computing power, cloud computing and fast Internet connections, new possibilities for the use of Deep Learning in industrial vision open up. Hard-to-solve inspection tasks become feasible.

Neural networks can be integrated directly into NeuroCheck in the TensorFlow and ONNX format. Other model formats such as CNTK, Pytorch, Keras, Watson, and others are also possible. As a basis for your deep learning application, process integration (Profibus, Profinet, EthernetIP etc.), hardware support (2D, 3D cameras), and data acquisition (image data, XML, data base) have already been part of the standard functionality of NeuroCheck for a long time.

Main areas of application » Successful use when it really makes sense

Sample application » Inspect a rubber sleeve » Pseudo error rate reduced many times over

As part of an automated production step, a rubber sleeve is affixed to a carrier. This can cause pinches and back folding, which in turn may compromise leak tightness and thus the proper working of the assembly group. The challenging part of this inspection task is the high variety of fault appearances that can occur and that it is impossible to establish a quantitative classification of the faults. In addition, the component images differ in size, position, contrast, and brightness due to the production process. Traditional image procession produces a high rate of false alarms with an additional high risk for errors going undetected.

Inspection Rubber Sleeve » Incorrect » NOK

Example Application » Identification and Localization of Screws

Neural Network solves Inspection Task

To control an automated screwing process, precise identification and localization of the screws is necessary. The station processes a high number of various assembly groups with varying numbers of screws, surfaces and geometries. In addition, the screws may be slightly covered or absent, which must lead to a termination of the screwing process. With this application as well, a number of complex environmental conditions need to be dealt with. The example images show various levels of brightness, various sizes of assembly parts, positions and levels of focus.

By using a neural network, the entire task was solved with a single program. In addition, robustness of recognition under changing environmental conditions was greatly improved compared with the traditional process.