This repository aims to extend dataset size for my thesis, the topic of which is anomaly detection within CT scans of welds using neural networks.
For more information see RADASP repository.
- CT scans of welds - sliced to images
- Few slices are masked
- Only anomalous samples
Create synthetic normal samples using an inpainting algorithm -- removing the defects. This method is effective only for small defects.
Experimented with
-
Inpainting algorithm
-
Inpainting radius
-
Dilatation kernel
-
Image size
Results Only a couple of images were successfully inpainted, as it seems like the defects included in the set are too complex or large for a true expansion of the set using inpainting.
Optical flow takes advantage of sequential order of slices extracted from 3D CT views.
The flow (transformation) is calculated from a Ground-Truth (GT) masked image and an unmasked (unlabeled) image. The computed flow is then applied to the GT mask, resulting in a propagated mask, which corresponds to the unlabeled image.
This process is effective only with slices that are near each other = there are no major changes between them. This means that these images, as well as the masks, are very similar but not identical -- providing new data for the dataset.
Results Using optical flow seems very effective for this dataset expansion. The drawback that comes with it is related to the core principle of this method -- the set will contain multiple very similar samples. Experiments demonstrated that repetitive application of this method does not result in greater NN performance.
The goal of this method is to find a similar dataset that could be used in combination with the original one, resulting in a larger overall set.
Finding a similar dataset is not a simple task, as CT scanning is not as common for welds. Consequently, priority was given to finding a set of X-rays images of welds (since a CT slice is somewhat similar to an X-ray image).
Results The best candidate found is public GDXray dataset -- specifically the "Welds" portion (the set also contains some other categories, such as Nature, Baggages, etc.).
As seen above, the image texture is similar to the original set. Based on the images' format, they are also split in quarters so they better match the original dataset's format. Unfortunately, the set contains only a few masked images of the welds, but by splitting them, the additional size is quadrupled.
A GAN was also considered as a potential method. The issue is that GANs are generally used for generating synthetic normal data -- they effectively learn the distribution of what a normal sample should look like and can replicate it quite well. In this case (due to the absence of normal data), the GAN cannot be trained to produce such samples.
Theoretically, a GAN could also be used to generate defected samples, but raised concerns are that the network would be very difficult (if even possible) to stabilize. It is not clear how well it could produce a relevant (realistic) data when the original set contains such a complex defects.
The expansion is carried out primarily by the Optical Flow method. Additionally, some samples are added through inpanting and by taking samples from the GDXray set, effectively increasing the original dataset size several times. The final count is determined by how many additional samples are masked using the flow for each regular GT sample.