Defect Classification in Steel Sheets

Implementation of an automatic classification system, using convolutional neural networks, for defects in steel sheets produced during the manufacturing process.

Objectives

• Automatically verify the quality of steel sheets
• Increase the quality of the final outcome of the manufacturing process
• Identify the critical stages of the manufacturing process

The steel sheet manufacturing process is particularly delicate. It involves a large number of processes, and small variations in some of them, such as heating or rolling, cause production defects and, therefore, loss of quality or defective parts. For this reason, being able to identify the quality of parts at different stages of the manufacturing process using computer vision ensures that the final quality meets the desired standard, by discarding defective parts and pinpointing the most critical stage.

What we do

We carry out an exploratory data analysis and develop a solution based on the development and training of a deep learning system. In this case, a U-Net convolutional neural network.

Afterwards, an exploratory analysis of the results obtained is carried out again to ensure that the trained model’s accuracy is as high as possible.

Finally, the solution is deployed into production. In addition, if deemed appropriate, the model is retrained using an increasingly updated dataset.

How We Do It

Analysis and classification of input images

A set of images including all the defects to be classified.
A file with the defect type and the manually generated defect location. In this case, possible defects are classified into 4 previously defined groups.

Model-specific training

The Keras segmentation-models library is used with pre-trained models.
U-Net, a convolutional neural network for image segmentation, is used.
The model’s output is a mask with easily identifiable outlines that locate the defects found by the neural network.

Deployment

Once the model has been trained, feeding it an unclassified image returns a heat map that identifies the possible defects contained in the steel sheet.