Production of material handling equipment
Company:
Jungheinrich
Services:
Production of material handling equipment
Industry:
Material Handling
Project Duration:
5 Months
Automated pipeline: labeled images to fully trained model
Training and classification process with no-code solutions
Image to classification based on detection from associated model
Web API integrated in client’s production framework
Overview
Challenges
Solutions
Results
Tools
This project involves the development of a Computer Vision tool designed to enhance quality control processes at production sites. The implementation of this Computer Vision tool enabled the client to establish a dependable, automated quality control system, resulting in significant time savings and reduced manpower requirements.
Team: Paul-Louis Pröve, Florian Fabry, Adrian Ebert
The project faced several technical challenges. Firstly, there was a need to develop a web tool capable of annotating images, which had to be seamlessly integrated with a machine learning framework to create an automated pipeline. Secondly, the AI training framework required was not only to be adaptable to each specific project but also reliable and scalable, especially when new images were introduced. Lastly, a critical challenge was to ensure the real-time responsiveness of a web API that could process provided images, incorporating a project-specific postprocessing logic to meet the diverse needs of quality control.
To address these challenges, the team utilized CVAT for image annotation, which was coupled with Azure Blob Storage. This integration allowed for the automatic initiation of training sessions whenever a project was exported. The export of a project triggered a training process within Azure Machine Learning. Subsequently, the trained model was stored in a separate Blob storage. An Azure function was employed to retrieve the model from storage and manage image requests. Additionally, a no-code configuration file was provided, enabling the definition of specific rules for each project. These rules facilitated the generation of straightforward OK/NOT OK responses, streamlining the quality control process.
The client successfully integrated the automated training pipeline into their existing production framework. With a sufficient supply of labeled images, the trained models achieved a remarkable precision rate of over 99%. Furthermore, the implementation of the web API, capable of real-time responses, significantly reduced the time spent on quality control processes. This system not only streamlined operations but also established an additional, reliable control mechanism within the production environment.
-
Ultralytics/YoloV8
-
FastAPI
-
Pytest
-
Git
-
Huggingface
-
Streamlit
-
CVAT
-
MS Azure Functions
-
MS Azure Machine Learning
-
MS Azure Blob Storage