pixel classification

This repo contains code to generate a make a binary classification decision for pixels in 2D images. Eventually C-code is generated for the classifier and it is compiled into a mex file. The compiled classifier is about 50x faster (single core) than the one built in to MATLAB in the few benchmarks that I've performed.

This code is not written to be easy to use or useful for you, most likely you would should use the excellent Ilastic.

Pixel Features
Classifier
Example usage
To Do
References

Pixel Features

Just like Ilastic, features are extracted at multiple scales. The built in features are:

Gaussian (1)
Laplacian of Gaussians (1)
Gradient magnitude (1)
Structure tensor eigenvalues (2)
Hessian eigenvalues (2)

For details see/edit px_features_2d.m.

Classifier

Uses the class TreeBagger in MATLAB for classification. There is also templateTree/fitensebmle, see this comparison.

Example usage

The general procedure is something like this:

Choose a training image.
Create a copy of the training image and annotate it in your favorite drawing program. Use red and green to indicate background and objects.
Extract features from the labeled pixels in the training image
Build a random forest to classify the pixels.
Export the classifier as C-code.
Compile the classifier.
Use the classifier on other images.

Here is example code (in MATLAB) for classifying a single image:

image = 'image.tif';            % Image to classify
image_mask = 'image_mask.png';  % Annotated image
outFolder = 'test-classifier/'; % Where to store the results
mkdir(outFolder)

% Create binary masks for objects and background
mask = imread(image_mask);
bg = mask(:,:,1) > 200 & mask(:,:,3) < 50;
nuclei = mask(:,:,2) > 200 & mask(:,:,3) < 50;

% Read the image train on
I = df_readTif(image);

% Generate and train a random forest
px_gen_classifier(I, max(double(bg), 2*double(nuclei)), outFolder)

% Generate C-code and compile it into a mex file
model_to_c([outFolder 'Mdl.mat'])

% Apply it to the image
C = px_classify_image([outFolder 'Mdl.mat'], image);
C = double(C);
C = C/max(C(:));
imwrite(C, [outFolder 'classification.png'])

TODO

Parallelized classifier.
Generation of standard shared objects.
Improved logistics and usage in general.

Files:

px_interactive -- interactively mark two classes in an image and then classify it.
px_gen_classifier -- generate a classifier from an image and a label image.
px_classify_image -- classify an image (by name) using a classifier/model generated by px_gen_classifier. This function processes the image in tiles (with some overlap) so that it doesn't consume too much memory.
px_classify_region -- classify a image matrix. Used by px_classify_image.
mode_to_c.m -- convert a TreeBagger model into a compiled mex function cMdl
px_cleanup.m -- Clean up classification results with watersheds etc.
createComposite -- merge channels and segmentation result.

References

Ilastic A very user friendly program for 2D and 3D pixel classification. This is probably what you really want.

elgw/pixelClassifier