Steel Surface Defect Classification and Detection

• 

  Manufacturing of Steel

• 

  Classification and Detection of Surface Defects with labels assigned at top on real-world steel surface

• 

  Different Types of Surface Defects

• 

  Detection using bounding boxes

• 

  Pixel-wise Detection of Surface Defects

Inspiration

I am from the Steel City of India and it daily comes to address the problems which come forward in the manufacturing of Steel. So I decided to do a project on Steel Surface Defect to counter this problem and make the product more eco-friendly and with a reduction in cost. With minimum labor, this can classify and detect Steel surface defects.

What it does

A complete steel surface defect classification and detection tasks aim to achieve the specific class and precise location of each defect in an image, which makes it challenging for applying this task in practice.
In this project, a novel defect detection system based on deep learning is proposed which focuses on a practical industrial application: steel plate defect inspection. To achieve strong classification-ability, this system employs a baseline Convolution Neural Network (CNN) to generate feature maps at each stage.

How I built it

I have collected the Steel Surface defects in the Steel Plate from various research organizations and Steel plants. Personally visited India's largest Steel Plant Bhilai Steel Plant and collected the images of the surface defects. Initially, the data is preprocessed and cleaned as it was very noisy and images are of varied sizes. Trained the data using Convolution neural Network pre-trained models like resnet, vgg, etc. I have built the model using the PyTorch framework.

Challenges I ran into

Contacting and collecting data from various steel plants and research organizations was very tedious. Personally collecting data was also very cumbersome. Counter the noise in data was other major issues and training the model with limited computational power was also very challenging.

Accomplishments that I'm proud of

Classification and Detection of Steel surface defects accurately having large noise was very exciting. Classification accuracy of Steel surface defects was around 99.88% and the detection of Steel surface defects pixel-wise in images with the labels assigned in each Steel plate was very soothing.

What I learned

Data Collection of Steel Surface images. Training the data on images with varied sizes and properties with limited computational power. Exploring and learning by applying advance Convolution Neural Network architecture was very exciting.

What's next for Steel Surface Defect Classification and Detection

Integrating the final project and building a prototype as a Mobile Application/Website for the industrial usage by Steel Manufacturing Plants

Built With

• computer-vision
• deep-learning
• python
• pytorch