FeatureExtractor

FeatureExtractor is a matlab script that extracts variety of visual features from image files.

List of features

1. Adaptive color quantization -- reduces colors on each image to N colors using adaptive bin size.
   • Number of features: 112
   • Output: PREFIX_acq.txt
2. Basic statistics on light channel -- mean, standard deviation, etc.
   • Number of features: 7
   • Output: PREFIX_light.txt
3. Grayscale histograms -- 8 and 32 bins
   • Number of features: 40
   • Output: PREFIX_light.txt
4. Basic statistics on RGB channels -- mean, standard deviation, etc.
   • Number of features: 21
   • Output: PREFIX_rgb.txt
5. RGB color histograms -- 4 and 8 bins
   • Number of features: 36
   • Output: PREFIX_rgb.txt
6. Basic statistics on HSV channels -- mean, standard deviation, etc.
   • Number of features: 21
   • Output: PREFIX_hsv.txt
7. HSV color histograms -- 4 and 8 bins
   • Number of features: 36
   • Output: PREFIX_hsv.txt
8. Block difference of inverse probabilities -- 2x2, 3x3, 4x4
   • Number of features: 29
   • Output: PREFIX_spatial.txt
9. Block variation of local correlation coefficients -- 2x2, 3x3, 4x4
   • Number of features: 29
   • Output: PREFIX_spatial.txt
10. Entropy
    • Number of features: 1
    • Output: PREFIX_texture.txt
11. Grey-Level Co-occurrence Matrix texture measurements (GLCM) -- constrast, correlation, energy, etc.
    • Number of features: 16
    • Output: PREFIX_texture.txt
12. Sobel edge energy
    • Number of features: 1
    • Output: PREFIX_texture.txt
13. Gabor features -- 4 orientations (0,45,90,135) and 4 sizes (1x,0.5x,0.25x,0.125x)
    • Number of features: 16
    • Output: PREFIX_gabor.txt
14. Similarity-based color segmentation -- 4 block sizes (1x1,2x2,3x3,4x4)
    • Number of features: 34
    • Output: PREFIX_segment.txt

Total number of features: 399

How to run

There are two different ways to specify input images. 1. Using absolute or relative path to the directory of the images 1. Using path to a text file containing a list of images

The second approach is useful when pausing/resuming is necessary for the processing job.

Option 1: Use the wrapper scripts

Simple mode: ./runFeatureExtractor.sh {path-to-images-or-image-list} {prefix}

Or, also process sub-directories: ./runFeatureExtractor_subdirs.sh {path-to-images-or-image-list} {prefix}

Option 2: Run from matlab interactively

1. Start matlab without GUI matlab -nodisplay
2. In matlab, set script path path(path,'PUT-YOUR-SCRIPT-PATH-HERE');
3. Run FeatureExtractor FeatureExtractor('PATH-TO-IMAGES-OR-IMAGE-LIST','PREFIX');
4. Exit matlab exit;

Option 3: Run matlab in batch mode

matlab -nodisplay -r "path(path,'PUT-YOUR-SCRIPT-PATH-HERE'); FeatureExtractor('PATH-TO-IMAGES-OR-IMAGE-LIST','PREFIX'); exit;"

Performing PCA on the output files

The PCA script is located under matlab/PCA.

The purpose of the script is to process the text files that have been created by FeatureExtractor and calculate PCAs of the data. It calculates a PCA for each sub file (e.g. one set of PCA values for light, one for texture etc.) and also its calculates one PCA for the entire data set.

The command consists of: 1) the full path to the script itself 2) the directory containing FeatureExtractor data files 3) the 'prefix' that was previously used by FeatureExtractor in creating the data.

For example: if FeatureExtractor used "all" as the prefix, and output was 'all_light.txt' 'all_texture.txt' etc., the PCA command uses 'all' in order to analyze those data files.

There are two ways to run the PCA script.

How to run

cd matlab/PCA ./runPCA.sh {PATH-TO-RESULT-FILES} {PREFIX}

1. Start matlab without GUI matlab -nodisplay
2. In matlab, set script path path(path,'PUT-YOUR-SCRIPT-PATH-HERE');
3. Run PCA script runPCA('PATH-TO-RESULT-FILES','PREFIX');
4. Exit matlab exit;

{PATH-TO-RESULT-FILES} is the path to the output files of FeatureExtractor

Outputs

Different PCAs are performed: 1. all - using all features 1. composition - using _segment.txt and_spatial.txt 1. color - using _light.txt,_rgb.txt and _hsv.txt 1. texture - using_texture.txt 1. orientation - using gabor.txt

For each PCA, there are two output files: 1. *.coeff.txt - coefficients of the top three principal components 1. *.score.txt - coordinates of each image in the space spanned by the top three principal components