Preprocessing functions for TGS Salt Dataset

[paritosh5feb]

Usually, the identification of Salt using Salt images from the Subsurface is done manually and is a very cumbersome and time taking task. Also, the chances of error are very high. With the advent of CNN architectures such as U-NET, the task of identification of salt from salt images was delegated to Semantic segmentation done by these CNNs. The results were highly accurate and a lot of time was saved. But there is one catch with these architectures. These CNNs are supervised algorithms and thus in order to carry out semantic segmentation, they require pixel-wise labelling in form of boolean masks. Which is another time-taking task. But the advent of Segment-Sort eliminated all these hurdles, as it can also carry out unsupervised segmentation on the dataset without any need for labels or boolean masks. The current code base of SegSort is capable of carrying out unsupervised semantic segmentation on the PASCAL VOC dataset. The code-base is executed via a bash script, and hence carries out the pre-processing task, specific to the PASCAL VOC dataset during the run-time. The pre-processing functions are added in the file 'preprocessing.py' and add value and usability to the code-base as follows:-

1. These functions can be used to prepare the TGS Salt dataset to be used with the code-base.
2. There are separate functions for creating the TensorFlow input queue and training and testing tensors in form of NPY files.
3. Hence the dataset can be pre-processed without executing the whole code-base.
4. These pre-processing functions can be used with any other dataset in their original directory structure.
5. The functions are tested and working.

[jyhjinghwang]

Hi @paritosh5feb,

Thanks for the pull request! Could you please also include in the comment the website link for the TGS Salt Dataset and the pointer to the data that you use? That would be really helpful if someone would like to try out this preprocessing code.

Thanks, Jyh-Jing