Detection of surface flaws in a pipe using vision based technique
1. Detection of surface
flaws in a pipe using
vision-based techniques
By Aadil Khan
Guided By Prof. Tejas P. Gotkhindi
2. Introduction
• Pipelines play a crucial role in globally
efficient transportation of materials like
water, oil, and gas.
• However, without regular maintenance,
pipelines can degrade, posing
environmental and safety risks.
• In India, a vast pipeline network exists, and
safety frameworks, including standards like
PNGRB and ASME, are in place.
• The proposal for an undersea pipeline from
India to Dubai, spanning 2,000 km and
costing around $5 billion, exemplifies the
continued importance and expansion of
pipeline infrastructure.
3.
4. Motivation
• The project, "Detection of Surface Flaws in a Pipe
using Vision-Based Techniques," merges AI/ML
and mechanical engineering to address pipeline
flaws.
• Using CNNs and GANs, it employs image
segmentation and realistic image generation.
• Incorporating AutoCAD & FEM adds a
mechanical aspect for holistic structural analysis.
• The project positions the researcher at the
forefront of technology, fostering
interdisciplinary collaboration and aiming to
enhance pipeline safety through innovative
vision-based techniques.
5. What is the problem we are solving ?
• Our project addresses the critical challenge of timely
crack detection in pipelines to prevent catastrophic
consequences like leaks and system failures.
• Traditional manual inspection methods hinder
proactive identification, prompting us to leverage
advanced computer vision algorithms.
• The project aims to autonomously detect and
characterize surface flaws, enhancing efficiency and
contributing to the safety and reliability of pipeline
networks.
• This proactive approach ensures the resilience and
longevity of critical infrastructure.
6. Data Availability
Initially we collect the pictures of road cracks , wall cracks and cracks
of water pipe due to unavailability of gas pipelines using ROV .
The remotely operated vehicle (ROV) navigates a cemented water pipeline
with axial and circumferential cracks, simulating conditions similar to gas
pipelines.
This pictures differs from the real gas pipeline and also publically data is
not available.
9. Generative Tools - WGAN
We come up with a solution to solve data
unavailability of data with custom
generation of images using GAN that looks
like real crack image after few operation .
WGAN it better than normal GAN for
generating images due to improved
training stability, reduced mode collapse ,
and more meaningful gradients
throughout training.
Hence, We use WGAN instead of Normal
GAN
12. Generated Using WGAN
(128x128)
• We have generated images by implementing many combinations of layers in
generator and discriminator listed below
The generator architecture includes (first combination):
• 5 transposed convolutional layers with corresponding batch normalization
layers and ReLU activations.
• 5 normal convolutional layers with corresponding batch normalization layers
and ReLU activations.
• 1 transposed convolutional layer with a tanh activation function for
generating the final output images.
The discriminator architecture includes (first combination):
• 4 convolutional layers with leaky ReLU activations and batch normalization.
• 1 fully connected layer for producing the discriminator score.
Generator
layers
Discriminator
layers Channels Resolution Epochs Results
10 5 3 256 1000 BAD
10 5 3 128 1000 BAD
10 5 3 128 10000 GOOD
10 7 3 128 10000 AVG
12 5 3 128 10000 BAD
12 7 3 128 10000 BAD
13. Tiled and
applying Super
Resolution
After this we follow few processes to make images realistic
1.) We used python library to split 16 images with effective areas.
2.) Applying Super Resolution: Super Resolution using ESR-GAN increases the image resolution and training the segmentation model
becomes easier.
3.) After this we apply Gaussian Blur for smoothing with kernel size = (1,1) and also adjusted brightness and contrast after that
we decided that the images nearly look like the real ones.
14. Creating Ground
Truth using ImageJ
• We use imageJ to manually create ground truth and after that we use
python library to make background black for all 254 images .
16. Observations and results
Manually-created Ground Truths and the predictions by DeepCrack model were
compared .
IOU similarity score was calculated for 28 pairs of images. It turned out to be
around 0.64.
If the data distribution is same, we get a better IOU score, otherwise the results
deviate.
This model applied on the images captured by the ROV tend to perform a little
worse. The IOU score for the dataset captured by the ROV is around 0.57.