letter_compare.m

这是用matlab编写的有关图像识别分类方法的源代码

     %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
   %%                                                            %%
 %%       Prof. Sclaroff's CS585 Image avd Video Processing       %%
%%                         Project  ONE                            %%
%%           C H A R A C T E R   R E C O G N I T I O N             %%
%%                                                                 %%
%%                    by Stanislav Rost                    %%
 %%                       ID:  31764117                           %%
  %%                                                             %%
   %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

function result = letter_compare(ocr_letters, document_letter)
% LETTER_COMPARE.M
%
% function result = letter_compare(ocr_letters, document_letter)
%
% Parameters:  ocr_letters is an array (1..4) of structures
%		with fields:  Mean, Covariance
%	       Subscripts of array mean 1=O 2=E 3=A 4=U
%	       document_letter fields is a structure which has
%		with fields:  Image, EulerNumber, Area, BoundingBox
%
% This function compares a letter extracted from the image paragraph
% to each of the four vowels.  Returns the index of the vowel
% which the letter from the paragraph matches, if any.
% If no vowels match up, return 0.
%
% Assumptions:  
%  - the letter form the document/paragraph is properly oriented


% Calculate the data necessary for the matching
perImg = bwperim(document_letter.Image);
perArea = bwarea(perImg);

compactness = perArea^2/document_letter.Area;

moments = invmoments(document_letter.Image);

% Calculate the top/bottom area ratio

midpoint = floor(document_letter.BoundingBox(4)/2);
topPart = document_letter.Image(1:midpoint,:);
bottomPart = document_letter.Image((midpoint+1):end,:);

partRatio = bwarea(topPart)/bwarea(bottomPart);

% Final vector for the doucment image which will be used for comparison
vector = [ moments(1) moments(2) moments(3) compactness partRatio ];

% Go through all four letters
for i = 1: 4,

	% Closeness will be Mahalanobis distance without the log term
	closeness(i) = (vector - ocr_letters(i).Mean)*...
			inv(ocr_letters(i).Covariance)*...
			(vector - ocr_letters(i).Mean)';

	% Distance is Mahalanobis distance with the log term
	% The log term equalizes the distance space so that
	% we can compare which one of the four letters
	% is closer to the letter extracted from the paragraph
	distance(i) = closeness(i) + log(det(ocr_letters(i).Covariance));

end


% Find one of the vowels with the smallest distance to the document letter
minimumDistance = distance(1);
closestMatch = 1;

for j = 1:4,
	if distance(j)<minimumDistance
		minimumDistance = distance(j);
		closestMatch = j;
	end
end


% Now we have to determine if the closest match is noise or not
variance = [ ocr_letters(closestMatch).Covariance(1,1) ...
	     ocr_letters(closestMatch).Covariance(2,2) ...
	     ocr_letters(closestMatch).Covariance(3,3) ...
	     ocr_letters(closestMatch).Covariance(4,4) ...
	     ocr_letters(closestMatch).Covariance(5,5) ];

% Standard deviation
std = sqrt(variance);

matchCloseness = closeness(closestMatch);

% Use Mahalanobis distance to figure out the allowed range for
% the closeness within which the letter is accepted as a match
allowedRange =  6.5*(std) * inv(ocr_letters(closestMatch).Covariance) * (std)';


% Closeness is within allowed interval
if (matchCloseness <= allowedRange)


% ######################  SECONDARY CHECKS  ################
	if closestMatch == 1
		% It thinks it is an O
		if (document_letter.EulerNumber ~= 0)
			% Reject
			closestMatch = 0;
		end
	elseif closestMatch == 2
		% It thinks it is an E
		if (document_letter.EulerNumber ~= 0)
			% Reject
			closestMatch = 0;
		end
	elseif closestMatch == 3
		% It thinks it is an A
		% Euler# can be 0 or 1

		if (document_letter.EulerNumber ~= 0) & ...
			(document_letter.EulerNumber ~= -1)
			% Reject
			closestMatch = 0;
		end

		% Ahh, but it might be an S !
		% S's have about equal distribution
		% of weight about their midpoint

		if (partRatio > 0.87 )
			% S rejected because the weight of the top
			% half is about equal to the wight of the
			% bottom half
			closestMatch = 0;
		end

	elseif closestMatch == 4
		% It thinks it is a U
		% Euler# can be 0 or 1

		if (document_letter.EulerNumber ~=1 )
			% Reject
			closestMatch = 0;
		end

		% Also, calculate the Eulers for the top and
		% bottom slices

		midpoint = floor(document_letter.BoundingBox(4)/2);
		topPart = im2bw(document_letter.Image(1:midpoint,:), 0.5);
		bottomPart = im2bw(document_letter.Image(midpoint+1,:), 0.5);

		topFeature = imfeature(double(topPart), 'EulerNumber');
		bottomFeature = imfeature(double(bottomPart), 'EulerNumber');

		topEuler = topFeature(1).EulerNumber;
		bottomEuler = bottomFeature(1).EulerNumber;

		% For u's, the top Euler number is 2, and the bottom
		% part has Euler number of 2
		if (topEuler~=2) | (bottomEuler ~= 2)
			% Can't be an U according to Euler checks
			closestMatch = 0;
		end

	end

	result = closestMatch;

else
	result = 0;
end