Codes of the paper: "Self-Supervised Learning for Seismic Image Segmentation from Few-Labeled Samples" published by IEEE Geoscience and Remote Sensing Letters (Volume 19 - 2022).
Authored by:
- Bruno A. A. Monteiro - Department of Computer Science, UFMG, Belo Horizonte, Brazil
- Hugo Oliveira - Institute of Mathematics and Statistics, USP, São Paulo, Brazil
- Jefersson A. dos Santos - Department of Computer Science, UFMG, Belo Horizonte, Brazil
The key idea of the SSL pretext task consists of pretraining the backbone on the same data domain of the final task.
We have designed and evaluated three pretext tasks based on classical image processing techniques to force the model to learn semantic features of seismic images without using manually labeled data. We then showed that the pre-trained extracted attributes were relevant for a further segmentation task.
- Rotation: the original image is randomly rotated into one of the five possible angles (-8,-4, 0, 4, 8 degrees). Then it is cropped, so no empty values are given as input. The network must then identify which of the rotations was applied.
- Jigsaw: the original image is cropped into 9 regular-sized tiles with a small random gap between them. To avoid an overly complicated task, the possible permutations are limited to 1000, being 500 randomly selected, and for each of them, its pair with the biggest hamming distance is selected. The model is requested to determine the original position for each one of the permuted tiles.
- Frame Order: Six key positions are defined equally distributed within the dataset set to be used as pseudo-classes. Than each randomply picked section (crosslines on the F3 Dataset, Inline on the Parihaka), is passed through the CNN and the final layer classifier predicts the key that is closer to each section.
When training the model to solve the pretext tasks, the entire train set is used to train, and the validation set is used for evaluation.
We randomly selected 1, 5, 10, and 20 labeled seismic sections from the train set, and used only them for fine-tuning. Later on, testing into the entire test set of each dataset.
-
ResNet-50 provided by PyTorch.
-
For the classification tasks, the output of the backbone is connected to an average pooling layer followed by the fully-connected classifier
-
For the segmentation task, two final convolutional layers are used to segment the dataset classes
- Dropout was applied after each one of the four layer sets.
- All training images were augmented using random crop, half-chance horizontal flip, and Gaussian noise addition
- Initial learning rate of
$1 \times 10^{-4}$ for the backbone and$1 \times 10^{-3}$ for the segmenter - For the baseline, the best setup was employing 10x bigger learning rates
- At every stage, we use a weight decay of
$1e^{-4}$ . - Confidence interval calculated assuming a two-tailed paired t-student distribution.
- matplotlib
- numpy
- scikit-image
- scikit-learn
- segyio
- torch
- torchinfo
- torchvision