Code covered by the BSD License  

Highlights from
Trainable COSFIRE filters for keypoint detection and pattern recognition

image thumbnail

Trainable COSFIRE filters for keypoint detection and pattern recognition



05 Jul 2012 (Updated )

A COSFIRE filter detects features or patterns of interest, e.g. traffic signs in complex scenes.

Gabor(img, lambda, aspectratio, bandwidth, orientation, shift, halfwave, supPhases, inhibition, fastmode, thinning)
function [result oriensMatrix data filterkernel] = Gabor(img, lambda, aspectratio, bandwidth, orientation, shift, halfwave, supPhases, inhibition, fastmode, thinning)
% Get the Gabor filter response for the given img and for the given
% parameter values

  % set all the initial data-variables  
  oriensMatrix = [];  
  if iscell(orientation)
      data.orientation = cell2mat(orientation);
      data.nroriens = length(data.orientation);
      data.orientation = 0;
      data.nroriens = orientation;
  if max(img(:)) > 1
      data.img = double(double(img) ./ 255);
      data.img = double(img);
  data.wavelength=lambda; % wavelength
  data.phaseoffset=shift; % phase offset (0 and 90 degrees)
  data.aspectratio=aspectratio; % aspectratio
  data.bandwidth=bandwidth; % bandwidth
  data.sigma=0; % sigma
  data.hwstate=1; % if this value is 1, than half-wave rectification is applied, otherwise none (checkbox)
  data.halfwave=halfwave; % percentage for use with half-wave rectification
  data.halfwave_old=0; % old percentage (if it is unchecked and checked again)
  data.selRight=2; % defines the option which is selected in the display-selection list
  data.inhibMethod=inhibition.method; % the used inhibition method (1= none, 2 = iso, 3 = aniso)
  data.alpha=inhibition.alpha; % alpha, parameter for inhibition
  data.k1=1; % k1, param for inhibition
  data.k2=4; % k2 param for inhibition
  data.facilMethod=1; % if facilitation should be used (1=don't use, 2=use) 
  data.beta=2; % alpha, parameter for facilitation
  data.f1=0.4; % k1, param for facilitation
  data.f2=2; % k2, param for facilitation
  data.power=2; %power, param for facilitation
  data.thinning=thinning; % if 1 then thinning is applied
  data.hyst=0; % if 1 then hysteresis thresholding is applied
  data.tlow=0.1; % tlow, threshold parameter
  data.thigh=0.2; % thigh, threshold parameter
  data.supPhases=supPhases; % method of phases superposition (1: L1, 2: L2, 3: L-inf, 4: none)
  data.supIsoinhib=3; % method of superposition for isotropic inhibition (1: L1, 2: L2, 3: L-inf)
  data.changeRight=0; % defines if the output image should be displayed
  data.changeLeft=0; % defines if the original image should be reloaded
  data.fastMode=fastmode; % this value makes sure no already available values are recalculated
  data.orienchanged=0; % used to define if in the gui 'Orientation' or 'Number of orientations' is changed => a new orientationlist should be displayed
  data.error=0; % this remembers the errors in the inputvalues (typed in in the GUI)
  data.groundtruth = 'no image selected'; % the filename of the groundtruth
  data.invertGT = 0; % this stores if the groundtruthimage is an inverted version of the output as displayed (black-on-white vs. white-on-black)
  data.performanceValue = NaN; % this stores the performance value
  data.invertOutput = 0; % this stores if a white background should be used (inverting the original result) 

  % apply the calculations. fastMode defines a number which
  % defines the speedup, e.g. fastMode == 6: no speedup (all calculations are made), 

    % initialization and calculation of convolutions
    % note that the list of orientations is sorted and contains no
    % duplicate values
    [data.img, data.orienslist, data.sigmaC] = readandinit(data.img, data.orientation, data.nroriens, data.sigma, data.wavelength, data.bandwidth); % initialisation
    [data.convResult data.filterkernel] = gaborfilter(data.img, data.wavelength, data.sigma, data.orienslist, data.phaseoffset, data.aspectratio, data.bandwidth);
    data.oriensdisp = data.orienslist;
    data.selection = (1:size(data.oriensdisp,2)); % the indexes of the orientations which should be calculated
    result = data.convResult;
    filterkernel = data.filterkernel;

  if (data.fastMode > 0)
    % calculation of half-wave rectification
    data.hwResult = calc_halfwaverect(data.convResult, data.orienslist, data.phaseoffset, data.halfwave);
    result = data.hwResult;
  if (data.fastMode > 1)
    % calculation of the superposition of phases
    data.superposResult = calc_phasessuppos(data.hwResult, data.orienslist, data.phaseoffset, data.supPhases);
    result = data.superposResult;
  if (data.fastMode > 2)
    % calculation of the surround inhibition
    data.inhibfacilResult = calc_inhibfacil(data.superposResult, data.inhibMethod, data.supIsoinhib, data.sigmaC, data.alpha, data.k1, data.k2, data.facilMethod, ...
                                            data.beta, data.f1, data.f2, data.orienslist, data.power);
    result = data.inhibfacilResult;                                        
  if (data.fastMode > 3)
    % calculation of the orientationmatrix (maximum orientation response
    % per point) and merges the images per orientation to one image
    [data.viewResult, data.oriensMatrix] = calc_viewimage(data.inhibfacilResult, data.selection, data.orienslist);
    result = data.viewResult;
  if (data.fastMode > 4)
    % calculation of the thinned image
    data.thinResult = calc_thinning(data.viewResult, data.oriensMatrix, data.thinning);
    result = data.thinResult;
  if (data.fastMode > 5)
    % calculation of the hysteresis thresholded image
    data.hystResult = calc_hysteresis(data.thinResult, data.hyst, data.tlow, data.thigh);
    result = data.hystResult;

Contact us