12-05, 17:30–18:00 (UTC), General Track
Tenova, as an innovative engineering company, collaborates closely with its client-partners to create advanced technologies and services that optimize business operations.
This talk discusses the deployment of our image recognition system to identify and mitigate potential hazards on steel plants, specifically focusing on the detection of bulky steel pieces.
The system was deployed on-premise using an edge device and an IP camera, supported by Azure IoT Edge and a Flask API for image processing and prediction.
A recent migration to a RabbitMQ-based architecture using Pika enhanced scalability and communication.
The presentation will cover technical strategies, the challenges (like offline functionality and real-time, low-latency hazard detection) and the positive impact of the system on workplace safety and operational efficiency.
Tenova, as an innovative engineering company, collaborates closely with its client-partners to create advanced technologies and services that optimize business operations.
These solutions contribute to cost reduction, energy savings, environmental impact mitigation, and improved working conditions within the metals and mining industry.
One such groundbreaking solution is an Image Recognition System designed to identify and mitigate potential hazards, specifically focusing on the detection of bulky steel pieces.
Although the system achieved satisfactory accuracy through extensive training on thousands of images, its true impact is realized upon deployment.
Deployment, however, presented specific challenges that needed careful consideration:
- offline functionality: the model is required to operate seamlessly even without an internet connection.
- low latency: quick and efficient response times are crucial for real-time hazard identification.
In response to these constraints, we chose to deploy the Image Recognition System on-premise using an edge device equipped with an IP camera to capture images for analysis.
Leveraging Azure IoT Edge, we implemented a Flask API to expose an endpoint that receives base64-encoded images, processes them, and returns accurate predictions.
As part of our ongoing commitment to technological advancement, we recently transitioned to a Rabbit MQ-based solution, aligning with the broader system architecture.
This shift prompted a comprehensive refactoring of the inference module, leveraging the Pika library.
This enhancement not only streamlines communication within our systems but also contributes to the overall efficiency and scalability of our Image Recognition System.
Join us in this presentation as we delve into the intricacies of deploying an Image Recognition System in a steel plant, exploring the challenges encountered, the innovative solutions devised, and the positive impact on workplace safety and operational efficiency.
No previous knowledge expected
Passionate in Data Science and Machine Learning, involved in projects from ETL through modeling to deployment.
Some of the projects in which I took part are:
- Implementation of custom model for image recognition
- Modelization of physical processes to optimize performances
- Building ML services infrastructure leveraging Microsoft Azure Cloud services such as Azure Machine Learning Workspace, Azure Databricks and Azure DevOps
As side project I built www.whilemodeltrains.com, a little app that serves data related blog posts, presented 3 at a time.
I write about ML (and other stuff) on my blog at www.nicologiso.com.