Default Feature Presets

Dragonfly's Machine Learning Segmentation module provides default feature presets for both pixel-based training and region-based training. You can use these presets as is, or edit them as required (see Editing Feature Presets). You can also create new presets (see Creating New Feature Presets).

The default dataset features presets listed below are available for creating new pixel-based or region-based classifiers (see Dataset Features for more information about dataset features).

You should note that the default dataset features presets are stored in the directory: %programdata%\ORS\Dragonflyxxxx.x\Data\OrsTrainer\FeaturePresets.

It is usually better to edit the copy of a preset that is added to a model rather than the default preset itself.
Each filter in a dataset feature preset is applied separately during training and classification. The filters in a preset are not applied in a cascade or as a composite.
Dataset features presets
Preset   Category Description
Activation Map 1 Comprehensive U-Net architecture with an input patch size of 512
Activation Map 1 and 2 Comprehensive U-Net architectures with an input patch size of 512
Activation Map 2 Comprehensive U-Net architecture with an input patch size of 512
Anisotropic Diffusion Smoothing Kernel = 2D square of size 3, Option =Inverse, K = 10, Number of Iterations = 20
Bilateral Smoothing Kernel = 2D square of size 3, Sigma color = 1.0, Sigma spatial = 1.0
Canny Edge detection Kernel = 2D square of size 3, Sigma = 1.0
Close Morphology Kernel = 2D square of size 3
DoG Edge detection Kernel = 2D square of size 3, First sigma = 1.0, Second sigma = 16.0, Normalized = True
Erode Morphology 1. Kernel = 2D square of size 19
2. Erode: Kernel = 2D square of size 3
Extrema - Used to diminish the importance of small dark or light elements, this feature preset extracts the minimum and maximum value in a region.

1. Maximum: Kernel = 2D circle of size 5
2. Minimum: Kernel = 2D circle of size 5

Gabor_019 Texture analysis This stack of 19 Gabor filters with 19 different orientations (in radians) can be used to separate elements with oriented patterns.
Gabor_019F2 Texture analysis This stack of 19 Gabor filters with 19 different orientations (in radians) can be used to separate elements with oriented patterns.
Gabor_019F3 Texture analysis This is a stack of 57 Gabor filters with 19 different orientations (in radians) and 3 different frequencies can be used to separate elements with oriented patterns.
Gabor_05 Texture analysis This stack of 5 Gabor filters with 5 different orientations (in radians) can be used to separate elements with oriented patterns.
Gabor_05F2 Texture analysis This stack of 10 Gabor filters with 10 different orientations (in radians) and 2 different frequencies can be used to separate elements with oriented patterns.
Gabors Texture analysis This stack of 19 Gabor filters with 19 different orientations (in radians) can be used to separate elements with oriented patterns.
Gabors_MF Texture analysis This stack of 39 Gabor filters with 19 different orientations (in radians) and 2 different frequencies can be used to separate elements with oriented patterns.
Gaussian_MS Smoothing 1. Kernel = 2D square of size 3, Standard deviation = 1.0
2. Kernel = 2D square of size 3, Standard deviation = 2.0
Hessian Detector Kernel = 2D square of size 3
Kuwahara Smoothing Kernel = 2D square of size 5
Lipshitz Smoothing Kernel = 2D square of size 3, Top Hat= True
Maximum Smoothing Kernel = 2D square of size 9
Median Smoothing Kernel = 2D square of size 9
Membrane Projection Texture analysis 1. Kernel = 2D square of size 3, Projection Type = Sum
2. Kernel = 2D square of size 3, Projection Type = Mean
3. Kernel = 2D square of size 3, Projection Type = Standard Deviation
4. Kernel = 2D square of size 3, Projection Type = Median
5. Kernel = 2D square of size 3, Projection Type = Max
6. Kernel = 2D square of size 3, Projection Type = Min
Moments Texture analysis 1. Kernel = 2D square of size 3, Order = 1.0
2. Kernel = 2D square of size 3, Order = 2.0
3. Kernel = 2D square of size 3, Order = 3.0
Morphological Morphology This stack of 5 morphological filters can be used to uniformize intensity in a region, which can eliminate small areas within a region that can confuse a classifier.

1. Dilate: Kernel = 2D square of size 3
2. Close: Kernel = 2D square of size 3
3. Erode: Kernel = 2D square of size 3
4. Morphological Gradient: Kernel = 2D square of size 3
5. Open: Kernel = 2D square of size 3
Neighbors Convolution Stack of 3 x 3 editable filters with only one element of the kernel to 1.

[1 0 0, 0 0 0, 0 0 0]
[0 1 0, 0 0 0, 0 0 0]
[0 0 1, 0 0 0, 0 0 0]
[0 0 0, 1 0 0, 0 0 0]
[0 0 0, 0 0 1, 0 0 0]
etc...
Non Local Means Smoothing Kernel = 2D square of size 3, Smoothing = 1.0, Rician noise = False
OppOfSeqChamfer
(Opposition of Sequential Chamfer)		 Texture analysis Kernel = 2D square of size 3, Projection Type = Variance of all
Otsu Thresholding Kernel = 2D square of size 3
Self Intensity - The image itself.
Smoothing Features Smoothing Smooth noisy images, which helps a lot when working on pixels.

1. Median: Kernel = 2D square of size 3
2. Mean Shift: Kernel = 2D square of size 3, Maximal distance X = 3, Maximal distance Y = 3, Noise spread = 1.0
3. Mean: Kernel = 2D square of size 3
4. Maximum: Kernel = 2D square of size 3
5. Bilateral: Kernel = 2D square of size 3, Sigma color = 1.0, Sigma spatial = 1.0
6. Percentile: Kernel = 2D square of size 3, Percentile value = 50.0
7. Rank: Kernel = 2D square of size 3, Rank value = 0
Sobel Edge detection Kernel = 2D square of size 3, Apply Gaussian Before Sobel = False

The default region features presets listed below are available for creating new region-based classifiers (see Region Features for more information about region features).

You should note that the default region features presets are stored in the directory: %programdata%\ORS\Dragonflyxxxx.x\Data\OrsTrainer\RegionFeaturePresets.

It is usually better to edit the copy of a preset that is added to a model rather than the default preset itself.
Region features presets
  Description
Geometry This stack of 6 filters can be used to describe the geometrical characteristics of the regions.

Note You should have regions big enough to embrace a complete object to add this region features preset. The background is often very random in opposition with the elements to separate which have a more regular geometry.
Histograms

The default settings of this stack of 3 filters is listed below.
1. Bins: 32
2. Bins: 16
3. Bins: 8

Note The number of bins is crucial, with an irrelevant number of bins, two classes could end up classified in the same bin.

Neighbors Distance This stack of 9 filters can be used when you want to include or exclude dark regions in a bright surrounding or bright regions in a dark surrounding.
Statistics This stack of 5 filters describe properties of statistical distribution.