Data Augmentation
For a good result in neural network training, the rule is to present a large number of different examples so that the neural network can learn the pattern in the data. However, getting and writing down thousands of examples is an arduous and time-consuming task, and can consume hundreds of hours of a project.
To facilitate this process, Eyeflow.AI offers an extensive range of Data Expansion algorithms (Data Augmentation). It is about inserting disturbances in the images when they are being presented to the neural network in training, thus forcing the neural network to learn to recognize the patterns, without depending only on the static examples that exist in the dataset.
These are several changes such as:
- Changes in optics such as brightness, contrast, color, gamma
- Changes in the way rotations, deformations, positions
- Changes in quality such as blur and noise
Important!
The changes in the images cannot be so high that the object of interest can no longer be recognized. For this, Eyeflow.AI provides an example of these changes for the operator to verify that the changes are not exaggerated. What cannot be seen in these examples, cannot be learned by the neural network.
Important!
It is also useless to stress the changes and generate cases that do not happen in the real operation. It is useless to present an inverted image for the learning of the neural network, upside down, for example, if in real operation this situation will never happen.
Data augmentation parameters
Parameters for Images Data Augmentation
Data augmentation general parameters
General Parameters for all transformations
| Parameter | Values | Default | Description |
|---|---|---|---|
| Interpolation | choice [‘linear’, ‘nearest’, ‘cubic’, ‘area’, ‘lanczos4’] | linear | Interpolation for resize operations |
| Fill Mode | choice [‘constant’, ‘nearest’, ‘reflect’, ‘wrap’] | constant | Fill of null regions |
| Border Value | int [0 - 255] | 0 | The color to fill border regions |
Rotate image
Random rotation of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random rotation |
| Min and Max angle for Rotation | range from -90 to 90 | -20 to 20 | Random rotation of image |
| Rescale image | bool [True - False] | True | If image must be reescaled in rotation |
Translate image
Random translation of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random translation |
| Min & Max horizontal translation | range from -0.4 to 0.4 | -0.2 to 0.2 | Minimum & Maximum percent for horizontal translation |
| Min & Max vertical translation | range from -0.4 to 0.4 | -0.2 to 0.2 | Minimum & Maximum percent for vertical translation |
| Number of trials | int [1 - 6] | 3 | Maximum number of trials without degeneration boxes |
Shear image
Random shear deformation of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random shear |
| Min & Max shear image | range from -60 to 60 | -10 to 10 | Minimum & Maximum values for random shear deformation |
Scale image
Random scale of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random scale |
| Min & Max scale image | range from 0.4 to 1.6 | 0.8 to 1.2 | Minimum & Maximum values for random scale |
Random flip of image
Flip Parameters
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability of Horizontal flip | percent 0% - 100% | 0.3 | Random horizontal flip of image |
| Probability of Vertical flip | percent 0% - 100% | 0.3 | Random vertical flip of image |
Random contrast
Random changes in contrast of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random contrast |
| Min & Max contrast | range from 0.4 to 2 | 0.8 to 1.2 | Minimum & Maximum values for changes in contrast of image |
Random brightness
Random changes in brightness of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random brightness |
| Min & Max brightness | range from -0.6 to 0.6 | -0.2 to 0.3 | Minimum & Maximum values for changes in brightness of image |
Random gamma
Random changes in gamma of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random gamma |
| Min & Max gamma | range from 0.1 to 12 | 0.4 to 1.6 | Minimum & Maximum values for changes in gamma of image |
Random saturation
Random changes in saturation of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random saturation |
| Min & Max saturation | range from 0.1 to 2 | 0.5 to 1.5 | Minimum & Maximum values for changes in saturation of image |
Random hue
Random changes in hue of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random hue |
| Min & Max hue | range from -1 to 1 | -0.05 to 0.05 | Minimum & Maximum values for changes in hue of image |
Random noise
Random insert of noise of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random noise |
| Noise method | choice [‘gauss’, ‘poisson’, ‘speckle’] | gauss | Method for noise insertion |
Random blur
Random blur of image
| Parameter | Values | Default | Description |
|---|---|---|---|
| Probability | percent 0% - 100% | 0.3 | Probability of random blur |
| Kernel size | choice [3, 5, 7, 9] | 5 | The size of kernel to blur image |
Where should I go now?
Feedback
Was this page helpful?
Glad to hear it! Please tell us how we can improve.
Sorry to hear that. Please tell us how we can improve.