Author: Stella Enge

Image recognition and understanding

Image-based quality control “Damage Detection”

Advantages

  • Precision & Quality
  • Relief & Comfort
  • Time efficiency & Agility

Functional areas

  • Production / Assembly
  • Storage / Logistics

To the demonstrator

→ To the movie about "Damage Detection"
To overview

Image-based quality control “Damage Detection”

Automated inspection of damage and defects using AI-assisted quality control of reflective surfaces.

Where does the AI application offer the greatest benefit?

The inspection of reflective surfaces, for example of quality defects in production or damage to vehicles, is often still carried out manually and therefore turns out to be time-consuming and requires a high level of expertise on the part of the persons in charge. In order to increase the productivity of the processes, “Damage Detection”, an AI system for the quality control of shiny or diffusely reflecting surfaces, has been developed. It is suitable for a wide range of applications: from industrial production, for example in the automotive industry, to the leasing and insurance industry and to automotive appraisals. The system works automatically and takes less than a minute for surface inspection. The quality defects found are categorized using Deep Learning. The combination of deflectometry, i.e. the contact-free detection of reflective surfaces, conventional image recognition processes and artificial intelligence methods is a unique system.

What are the quality indicators of such AI applications?

  • “Damage Detection” offers low hardware and maintenance costs as a solution. The AI application can be used flexibly and retrofitted into existing productions. Among other things, the solution is able to work under the influence of stray light (e.g. ceiling lighting in a production hall).
  • The system delivers 100% test coverage. It combines high accuracy with low hardware requirements. Currently, defects as small as 0.1 mm can be detected on 1 m component size.
  • By using Artificial Intelligence, the system can be trained to detect various types of defects on the surface.
  • In addition to highly reflective surfaces, the latest developments also allow diffusely reflective surfaces to be tested for their quality.
“The combination of conventional image recognition
methods, AI methods and the detection of reflective
surfaces is unique.”
Dr. rer. nat. Theresa Bick
Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS

What AI technology powers the demonstrator of KI.NRW?

Computer Vision

The solution relies on Convolutional Neural Networks (CNNs) for detection and classification of the surface anomalies. The CNN model belongs to the Artificial Neural Networks, which are suitable for image recognition and understanding.

Informed Machine Learning

By incorporating expert knowledge, the system is able to make dataintensive neural networks practical for industrial production. The amount of training data and the annotation effort can be kept comparatively low.

Combination with classical image processing

The AI is linked to classical image processing. The system uses the advantages of both worlds – fast, approximate algorithms from classical image processing and the powerful methods of deep learning.

What does the AI demonstrator show?

Quality control of shiny or diffusely reflecting surfaces is notable for its simple, mobile hardware design. It operates independently of ambient light and is fully automated. It takes less than one minute for a surface inspection.

Where is more information to be found?

Image recognition in practice

Many “good practice examples” as inspiration for the possible applications of this AI.

To the examples

AI products “made in NRW”

Filter our AI map by “image recognition and understanding”.

To the AI map

AI provider from NRW

Our AI map illustrates who offers AI methods related to image recognition in their portfolio.

To the AI map

Contact us

Dr.-Ing. Stefan Eickeler

Senior Research Engineer

Fraunhofer IAIS
Schloss Birlinghoven
53757 Sankt Augustin

Phone +49 2241 141969

E-Mail senden

Speech and text comprehension

Chatbot with knowledge graph
“Covid Q&A”

Currently, no more data is supplied for our demonstrator. The prototype is temporarily not available.

Advantages

  • Real-time results 24/7

Functional areas

  • Human Resources
  • IT / Infrastructure
  • Marketing / Sales
  • Production / Assembly
  • Support / Customer service
To overview

Chatbot with knowledge graph
“Covid Q&A”

The chatbot demonstrates the functionality of an online dialog system that uses the strengths of a knowledge graph.

Where does the AI application offer the greatest benefit?

A chatbot supported by AI offers a wide range of possible applications and assists companies in any area of communication with employees or customers, regardless of the industry. A classic example are customer-oriented service offerings that are made available twenty-four hours a day. Companies use them on websites, in online stores, on support pages, in apps, or via instant messaging systems to simplify navigation on websites, to answer specific customer inquiries, or to structure access to service and customer support.

However, chatbots can be used successfully not only in external but also in internal corporate communications. Application examples can be found in the onboarding of new employees, in HR or administrative processes such as questions about vacation requests or payroll tax settlements, as well as in the support of complex assembly instructions in production.

Chatbots can also be supplemented by the component of acoustic speech recognition as well as acoustic speech synthesis. In this extended form, we speak of voice assistants (voicebot), similar to Siri or Alexa.

What are the quality indicators of such AI applications?

Understanding texts with Natural Language Understanding (NLU)

NLU methods are based on semantic representations of texts. They can understand and map relationships between words. These semantic representations exceed the possibilities of the classical rule-based methods of text mining.

Access information and prepare dialogs with Dialog Management (DM) and Knowledge Graph (KG)

Knowledge Graphs structure data and knowledge, enable semantic linking, and in many cases are the basis for making artificial intelligence applications explainable and providing results that are comprehensible to humans.

Generate texts with Natural Language Generation (NLG)

Text synthesis is the counterpart to text comprehension. Here, text is generated automatically, which can afterwards be transformed into acoustic speech signals.

“Knowledge graphs that integrate different data sources,
form the basis for many AI applications and assistants.”
Prof. Dr. rer. nat Jens Lehmann
Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS

What does the demonstrator of KI.NRW show?

An AI-based dialog system (“question answering system”) shapes the search for information more efficiently and conveniently for the user. The KI.NRW demonstrator shows such an AI-supported chatbot for querying Corona case numbers worldwide. The animations explain step by step how it operates.

Step 1: Understanding natural language

By using AI, speech recognition is oriented to real conversations of humans and the so-called natural language (“natural language understanding NLU”). The models extract information from the text, which they represent internally in a way that allows further processing. This enables the system to understand less common terms, dialects or everyday language. Associations and connections between words are also taken into account appropriately, for example, that the word invoice is related in content to the word payment.

Step 2: Structure data and knowledge

Knowledge graphs structure data and knowledge, enable semantic linking and, in many cases, are the basis for shaping AI applications explainable. A knowledge graph is able to combine a wide variety of information sources into a dynamic knowledge base. In the case of the KI.NRW demonstrator, Corona case numbers from Johns Hopkins University and the Robert-Koch-Institut are included. They are made accessible with the status of the previous day via a knowledge graph.

Step 3: Generate answer

Finally, an answer is generated that matches the asked question. Thus, this process is the matching counterpart to the first step, the understanding of language. The structured data is now converted into text and output. The output as an acoustic signal can also follow here (for example with the so-called voicebots).

The AI chatbot in action

Try out the AI chatbot for yourself: As an example, the chatbot was set up on data that depicts the worldwide case numbers surrounding the Corona pandemic. Anyone who asks a corresponding question in English via the chat window will receive an immediate answer.

Test questions for the knowledge query can be the following:

  • “Are there new cases in Mexico?”
  • “How many cases were there in total in Germany until 25th October 2020?”
  • “How many new cases were found in Argentina on 10th November 2020?”
  • “Which country had the highest number of cases on 8th November 2020?”

Where is more information to be found?

AI provider from NRW

Our AI map illustrates who offers AI methods related to “knowledge and inference” in their portfolio.

To the AI map

AI methods around knowledge and inference

Many “good practice examples” as inspiration for the possible applications of this AI.

To the examples

AI products “made in NRW”

Filter our AI map by “knowledge and inference”.

To the AI map

Contact us

Roman Teucher

Research Engineer

Fraunhofer IAIS
Zwickauer Str. 46
01069 Dresden

Phone +49 351 85477961

E-Mail senden

To the top