How computers learn to see – Applying AI to industry

Can AI learn to spot manufacturing defects with just a few thousand images instead of millions? This talk shows how pre-trained models make it possible.

#1about 1 minute

The challenge of quality control in car seat manufacturing

A car seat manufacturer needs to ensure hundreds of special clips are correctly placed on various seat types daily.

#2about 1 minute

Evaluating solutions for automated visual inspection

AI-based inspection is a superior solution compared to manual checks, which are error-prone, or classic computer vision, which requires extensive programming.

#3about 2 minutes

Understanding AI use cases for computer vision

AI uses statistical algorithms for tasks like image classification, object detection, and even generating new images with generative AI.

#4about 4 minutes

Preparing and pre-processing data for a machine learning model

The preparation phase involves clarifying the task, choosing a technical setup, collecting and labeling data, and splitting it into training, validation, and test sets to prevent overfitting.

#5about 2 minutes

A simplified overview of convolutional neural networks

A Convolutional Neural Network (CNN) uses layers like convolutional, pooling, and dense layers to learn features from images and make predictions.

#6about 2 minutes

Building from scratch vs using pre-trained models

Using a service provider with pre-trained models is more time and cost-effective, requiring far fewer images than building a model from scratch.

#7about 3 minutes

Evaluating model performance with a confusion matrix

A confusion matrix helps evaluate model performance by comparing actual to predicted values, highlighting critical metrics like false negatives and the escape rate.

#8about 2 minutes

Iterating and fine-tuning the model for better results

Improving model performance is an iterative process of collecting more data, adjusting pre-processing steps, trying different models, and tuning parameters using the validation dataset.

#9about 2 minutes

Deploying the AI model into a production environment

The final deployed system uses a camera, a classifier to identify the seat type, the inspection model, and a post-processing module to trigger actions like a robot arm.

#10about 1 minute

Key takeaways for applying AI in manufacturing

AI is a feasible solution for visual inspection in manufacturing, especially when using pre-trained models and embracing an iterative, experimental approach.

#11about 1 minute

Q&A on data sourcing and finding pre-trained models

The Q&A covers the necessity of collecting images directly from the customer's assembly line and finding pre-trained models from cloud providers or open-source repositories.