search

ShaharKatz / Caffe-Data-Augmentation

a branc caffe with feature of Data Augmentation using a configurable stochastic combination of 7 data augmentation techniques.
Other
32 stars 11 forks source link

readme

Caffe-Data-Augmentation

The original repository for Caffe, developed by the Berkeley Vision and Learning Center (BVLC) and community contributors, is at (BVLC\caffe)

This project adds a data augmentation feature to caffe, augmenting the data in 9 several ways.

The ways in which the data is augmentated is explained here:

  1. Image Translation - a random shift in a x and y axis pixels of the entire image. The shift has uniform probability between -7 and 7.
  2. Image Rescailing - shrinking or enalrgin the image (before cropping) by a random unifrom factor between 0.8 and 1.2.
  3. Horizontal Flipping - flipping the image in the horizontal axis.
  4. Vertical Flipping - flipping the image in the vertical axis.
  5. Elastic Deformation with Random Interpolation - dislocate pixels and use OpenCV interpolations method randomly.
  6. Color Noising - adding a small independent noise to each color channel of the image.
  7. Brightness Noising - adding a small noise to the brightness of each pixel.
  8. Small Blurring - convolving the image with small random-sized blurring kernel.
  9. Single Random Transformation - choosing a transformation at random.
  10. Multiple Random Transformations - chooses each transofrmation with probability 1/7, such that the mean is one transformation for every image.

The desired transformation(s) is chosen by parameter transform_type within the prototxt file for the data layer. The value of the parameter corresponds to the transform schemes described above. For example, transform_type=4 uses vertical flipping as it's transformation.

This project was developed by Shani Rehana, Baruch Epstein and Shahar Katz.

The latest version for this project is rc2.