Presentation is loading. Please wait.

Presentation is loading. Please wait.

Lecture 4b Data augmentation for CNN training

Similar presentations


Presentation on theme: "Lecture 4b Data augmentation for CNN training"— Presentation transcript:

1 Lecture 4b Data augmentation for CNN training

2 Data Augmentation Training of large network requires a lot of samples.
The easiest and most common method to enlarge the dataset using label-preserving transformations: image translations re-scale (both up and down) before crop horizontal reflections elastic deformation with random interpolations ((bilinear, area, nearest neighbor and cubic, with equal probability) (Simard, 2003) photometric distortion and altering the intensities of the RGB channels in training images (A.G. Howard. Some improvements on deep convolutional neural network based image classification, 2013)

3 Plankton competition The Plankton competition classify grayscale images of plankton into one of 121 classes. The images from camera were processed by a segmentation algorithm to isolate individual organisms, and then cropped. The size of an organism in the resulting images is proportional to its actual size, and does not depend on the distance to the camera.

4 Plankton: preprocessing and data augmentation
Rescaling all images to the same size zero mean unit variance (ZMUV) normalization Data augmentation: rotation: random with angle in [0°, 360°] translation: random with shift in [-10 ;10] pixels rescaling: random with scale factor [1/1.6 ; 1.6] flipping Shearing (parallel sliding of lines): random with angle [-20° , 20°] stretching: random with stretch factor [1/1.3 ; 1.3]

5 Data Augmentation preprocessed image Augmented image

6 ConvNet with cyclic pooling
add rotation to data layer 4 parallel identical nets to process rotated images root-mean-square pooling  before output layer

7 Pre-defined features Pre-defined features:
Image size (pixels) Image moments Texture features These features were fed into 2 dense layers and combined the CNN features just before output ‘softmax’ layer

8 Test-time data transformations
For each individual model Applied a quasi-random set of 70 transformations to input image compute predictions across various augmented versions of the input image averaged predictions


Download ppt "Lecture 4b Data augmentation for CNN training"

Similar presentations


Ads by Google