Aivia Software
Features (Pixel Classifier)
A feature describes a unique element in the image appearance, such as the object morphology, edges or curvatures, for use in the Pixel Classifier. These features are derived from image processing filters that transform and enhance certain aspects of the image. In Aivia, there are four feature groups:
Zeroth order derivative: describes the original morphology of the image
First order derivative: describes the edges on the image
Second order derivative: describes the curvatures on the image
Specialized features: describes unique features, such as thin line patterns
Within each of the feature groups, there are several different image transformation functions that are used to extract specific image features. Each of these transformations enable the Pixel Classifier to create a feature vector for each pixel to be used in the classification.
Zeroth order derivative
This feature group describes the original morphology of the image. There are six features within this group as described in the table below.
Feature Type | Description |
|---|---|
Original | This feature type uses the original image and requires negligible processing time. For multi-channel images, each image channel is used individually as a feature. For RGB images, this feature will convert the image into the HSV (hue, saturation, value) color space. |
Gaussian Blur | This feature applies a Gaussian filter to the image at multiple scales and uses each filtered image as a feature. The Gaussian Blur function is useful for describing image intensity at local and regional scales on the image. More information about the Gaussian blur function can be found here. |
Max | This feature applies a greyscale dilation to the image at multiple scales to determine the maximum intensity value at local and regional scales on the image. More information about the dilation function can be found here. |
Min | This feature applies a greyscale erosion to the image at multiple scales to determine the minimum intensity value at local and regional scales on the image. More information about the erosion function can be found here. |
Mean | This feature applies an average filter to the image at multiple scales to determine the mean intensity value centered at each pixel. The Mean function is useful for describing the overall intensity information at local and regional scales on the image. |
Median | This feature applies a median filter to the image at multiple scales to determine the median intensity value centered at each pixel. |
First order derivative
This feature group describes the edges on the image. The edges (boundaries between abrupt changes in intensity) are enhanced by applying an image transformation function with respect to the specified feature. There are four features within this group as described in the table below.
Feature Type | Description |
|---|---|
Variance | This feature applies a variance filter to the image at multiple scales to compute the intensity variance centered at each pixel. The Variance function is useful for describing the local and regional intensity uniformity. |
Difference of Gaussian | This feature applies two Gaussian filters with different processing kernel sizes at multiple scales to compute the difference between the two results. The Difference of Gaussian function is useful for describing edges on the image within local and regional contexts. More information about the difference of Gaussian function can be found here. |
Sobel | This feature applies a Sobel filter at multiple scale to describe the edges on the image within local and regional contexts. More information about the Sobel filter can be found here. |
Structure Tensor | This feature calculates a number of sub-features (determinant, trace, largest and smallest eigenvalues) from the structure tensor matrix at multiple scales. The Structure Tensor function is useful for identifying edges as well as sharp corners on the image. More information about the structure tensor can be found here. |
Second order derivative
This feature group describes the curvatures on the image to identify complex patterns such as blobs (large, flat disks) and tubes (long, thin edges). There are two features within this group as described in the table below.
Feature Type | Description |
|---|---|
Hessian | This feature calculates a number of sub-features (determinant, trace, largest and smallest eigenvalues) from the Hessian matrix at multiple scales. The Hessian function is useful for identifying complex object shapes. More information about the Hessian matrix can be found here. |
Laplacian | This feature applies a Laplacian of Gaussian filter to the image at multiple scales to identify blob-like (large, flat disks) objects on the image. The Laplacian function generates similar results to the determinant sub-feature of the Hessian matrix. More information about the Laplacian of Gaussian filter can be found here. |
Specialized features
This feature group can be used for enhancing very fine line-like or filamentous objects (such as microtubules or dendritic extensions). These features are processing-intensive and are recommended only for special circumstances. There are three features within this group as described in the table below.
Feature Type | Description |
|---|---|
Membrane Projection | This feature applies a line-shaped processing kernel to the image at multiple scales and at multiple rotation angles. Intensity statistics (such as mean, variance, max and min intensities) are calculated on the aggregate of the rotated images at each scale. |
Gabor | This feature applies a Gabor filter to the image at multiple scales to generate different directional images that describe the pixel orientations within local and regional contexts. More information about the Gabor filter can be found here. |
Gabor Aggregated | This feature combines the Gabor filtering results at multiple scales and at multiple rotation angles. Intensity statistics (such as mean, variance, max and min intensities) are calculated on the aggregate of the rotated images at each scale to remove the orientation aspect of the Gabor feature. |
3D Features
For processing 3D images, it is possible to use the 3D versions of select features above for improved pixel calculation. Instead of using a 2D (x-y) image processing kernel, the 3D features use a 3D (x-y-z) image processing kernel that incorporates adjacent voxel information across z-planes when calculating the feature vector. Seven of the features above have a corresponding 3D feature type:
Gaussian Blur
Difference of Gaussian
Hessian
Structure Tensor
Sobel
Mean
Laplacian