2. 2
When a seismic wave travels through an interface of two materials, a reflection wave will be generated.
https://en.wikipedia.org/wiki/Reflection_seismology http://geologylearn.blogspot.com/2015/06/marine-and-land-seismic-aquisition.html
3.
4. 4
128
x
128
16
1 16
32 32
16
32 64
64
64 128 128
128 256 256
128
128 128 128
64
64 64
64
32 32 32 32
16
16 16 1
128
x
128
Input
Image
File
Output
Segmentation
Map
conv 3x3, ReLU
copy & concatenate
max pool 2x2
up-conv 2x2
conv 1x1 Softmax
Learn
Context
Structure
Combine
Context &
Local Details
High
Resolution
Shallow
Low
Resolution
Deep
High
Resolution
Shallow
Low
Resolution
Deep
U-Net: Convolutional Networks for Biomedical Image Segmentation
Total Parameters: 2,158,417
5. 5
TGS Salt Identification Challenge
Segment salt deposits beneath the Earth's surface
Training images: 3,200
Evaluation images: 800
IOU = 72%
After 40 epochs of training (2 hours on Kaggle).
Grand Truth Prediction
False Negative (FN) False Positive (FP)
True Positive (TP)
𝐼𝑂𝑈 =
𝐼𝑛𝑡𝑒𝑟𝑠𝑒𝑐𝑡𝑖𝑜𝑛
𝑈𝑛𝑖𝑜𝑛
=
𝑇𝑃
𝑇𝑃 + 𝐹𝑁 + 𝐹𝑃
Image
Mask
Predict
6.
7. 7
Why Does Batch Norm Work? (C2W3L06)
Batch Normalization: Accelerating Deep Network Training
b y Reducing Internal Covariate Shift
Input
Covariant Shift
Normalize
the Input
to Improve
Prediction
Normalize each hidden layer,
so there is less covariant shift for next layer
8. 8
• Without domain knowledge, U-NET can do a reasonable job to analyze seismic data.
• Analyzing seismic data ourselves is using our own brain.
• Appling neural networks is like building a new brain, then train the brain with data to do the job.
• IOU was increased to 77% with the following steps
• Added a batch normalization layer after each convolutional layer
• Replaced Conv2DTranspose with UpSampling2D + Conv2D, making up-sampling more smooth.
• Increased the depth to 5 with a new layer at the bottom of U-Net: 4x4 (512 filters)
• How to improve further?
• Apply conditional random fields (CRF) to enhance pattern recognition, check this paper.
Python code of the project at kaggle: https://www.kaggle.com/dingli/seismic-data-analysis-with-u-net