Neuralyze® by senswork
Deep Learning Software
In the case of complex test objects, a set of rules for the clear differentiation of features can often no longer be described - classic image processing reaches its limits here. With our machine learning / deep learning software Neuralyze®, a description of rules is no longer necessary.
Neuralyze learns to recognize the smallest errors in any form with high precision and to interpret what is "recognized". A self-learning method based on neural networks is used to assess characteristics. For this purpose, image data of the objects or properties to be recognized are necessary in advance. The assessment of the characteristics is optimized in a training process.
In the video you can see Neuralyze in action in the food industry
Application examples
- Automotive industry: Surface inspection of sheet metal for scratches or traces of paint, detection of cavities in aluminum
- White goods:Inspection of aesthetically visible surfaces for dishwashers or other household appliances
- Food industry: Quality assurance for transparent plastic composite films
- Pharmaceutical industry: Quality assurance for single-use films, e.g. B. for syringes
- Medical technology: Checking infusion bags, cyano bags or other transparent foils for medical technology
- Packaging industry: Defect detection in transparent composite films
Benefits of Neuralyze®
- Efficient inspection software for tasks that cannot be solved with conventional image processing
- Inspection of test objects with transparent, reflective, curved or inhomogeneous surfaces is possible
- Detection of products with a high variance of features
- Dynamic Link Library (DLL)
- Integration in VisionCommander possible
- Simple and easy to use
Model Types of Neuralyze®
Object recognition is one of the oldest applications of classification algorithms and probably also the most widely used approach in the deep learning context.
Object recognition is one of the oldest applications of classification algorithms and probably also the most widely used approach in the deep learning context.
Semantic segmentation is a process with which the affiliation to a object class is determined.
Several models can be combined, each of which continues to work with the results of the predecessor.
Hardware and software from a single source
senswork is not only a recognized software specialist in the area of machine learning / deep learning, but also has many years of experience in the conception and design of camera, lighting and image processing systems as well as special machine construction for image processing systems.
As a result, senswork can offer everything from a single source, from the first concept to the integration of the finished solution. This includes the development and selection of hardware, prototype construction, construction and commissioning as well as the implementation of the software.
Your Contact Person
Markus Schatzl
+49 (0)89 215 298 46 0
markus.schatzl@senswork.com
senswork GmbH
Innovation Lab
Friedenstraße 18
81671 München
FAQs – Frequently Asked Questions
Neural networks can be applied to a wide variety of objects and materials, for example metal, glass, plastic, concrete, ceramic or wood. Limitations in size and dimension do not depend on the deep learning approach itself, but on the recording technology used.
Neuralyze is primarily suitable for the quality inspection of objects with variable size, shape, structure and variable background. Classic image processing reaches its limits in this area.
Our deep learning software is based on data that recognizes laws through machine learning. This training process is used to derive information from data obtained later.
An exclusion criterion is therefore the lack of availability and applicability of image data - without data there is no AI system.
If the number of training data is too small, in which not all properties are represented linearly independently, the learning algorithm could derive incorrect conclusions. It must therefore be ensured that a sufficient amount of training data is available.
Neuralyze is a key component that senswork uses in its own applications. This gives senswork the opportunity to develop a tailor-made system that is 100 percent tailored to the data and framework conditions of the specific application.
The software can be easily and intuitively integrated into existing manufacturing processes. If sufficient suitable image data is available, the training phase for the specific application can be implemented in a few days. If a suitable image processing system has to be designed first, the effort increases.
If Neuralyze is to be introduced into a process, the process is as follows:
- Creating recordings of damaged and intact components
- Definition of error categories
- Classification and labeling of the image data
- Adjustment of the resolutions
- Data augmentation to increase the amount of data
- Training of a neural network for error detection
- Validation of the classification quality on the basis of component photos not used in the training
- Integration of the neural network
- Integration of the system in the ongoing production process
- Staff training.
In the future, the software library will also be available as a stand-alone solution.
The result depends on the algorithm / network. Basically, we always try to return the data in a format that allows for further processing with image processing tools (binary images or polygon lists) as easily as possible. The resolution of the data corresponds to that of the camera system.
The processing speed depends on many factors, including the neural network or the topology itself, the image size, the average number of errors and the computing power of the process computer. In addition, other functions are often called before and after, so a test run is not equivalent to the inference time. Therefore, statements can only be made here individually.
In general, it can be said that the time required for classic image processing tasks is somewhat higher and that inference in classic fields of application with image data from average image processing hardware can be in the range of 50 ms to several seconds.