Group Actions
Engagement
  • Author Author: uavr&d
  • Date Created: Date Created
  • Views 669 views
  • Likes 3 likes
  • Comments 2 comments
Related
Recommended

#4 UAV Inspection System: CV trouble

uavr&d

If you’re following our previous posts, you might know that we are trying to implement Computer Vision in our Drone Inspection System. Although we have a model ready and trained, there is an issue with the credibility of the same. The model is trained on concrete crack images while we are trying to implement the model on pipe cracks. Both are different in a sense that cracks in pipes will be on a curved surface whereas the training set has all the images with cracks on flat surfaces. This might lead to a decrease in accuracy when implementing the solution on real data.

 

To overcome this, we tried searching for training datasets with cracks on curved surfaces, but we couldn’t get our hands on one. A few of them which we found after rigorous research was proprietary. We thought of making a dataset on our own but encountered some troubles there also. One, we couldn’t locate an industry nearby where we could take pictures of cracked pipes. The ones which we can approach is out of the city that we are located in. Two, even if we visit that industry, we would need at least a month to take a sufficient amount of pictures, post-process and label them. Considering the deadline for the design challenge, we decided not to do so. Therefore, the fact remains that while we have the vision to implement something that is new to the pipe inspection systems, we are unable to do it properly because of the lack of appropriate datasets.

 

We would love to have inputs from the community about the same. We understand that we might not be able to pull it off for this design challenge even if someone suggests us an alternative, but still, any help is welcome. This is the future of the inspection system and working on the same only makes sense.

 

Parents

  • If you cannot get hold of a good dataset for curved crack images, then you can try a different approach: you could try normalising your input images to the training data. This is a technique used quite often to prepare training datasets, and I think you probably can use it also the other way round.

     

    Basically, the idea is once you take the image of the crack, you apply a transformation, to flatten the curvature, so your image will be matching the type you used to train your solution. This way your model will deal with an image of a flat crack, similar to those used to train it.

     

    Your problem now would be to find a algorithm to transform the image, but you can find some good pointers to start from, like this paper on flattening curved text documents.

     

    I'm not saying it is going to be easy, but I think it might be worth a shot.

     

    Fabio

Comment

  • If you cannot get hold of a good dataset for curved crack images, then you can try a different approach: you could try normalising your input images to the training data. This is a technique used quite often to prepare training datasets, and I think you probably can use it also the other way round.

     

    Basically, the idea is once you take the image of the crack, you apply a transformation, to flatten the curvature, so your image will be matching the type you used to train your solution. This way your model will deal with an image of a flat crack, similar to those used to train it.

     

    Your problem now would be to find a algorithm to transform the image, but you can find some good pointers to start from, like this paper on flattening curved text documents.

     

    I'm not saying it is going to be easy, but I think it might be worth a shot.

     

    Fabio

Children