OCR手写数字识别的MATLAB仿真-源码

% this is a script for the OCR demo on the USPS data set % Multi-classification using one-versus-the-rest approach clc; clear; close all; warning off; addpath(genpath(pwd)); % load the training data - we are only using the first 2007 samples % because of memory limitations in MATLAB ... (X, y) load usps_digits_training.mat X = X(1:2007,:); y = y(1:2007,1); whos, % show what variables are in the workspace % load also the testing (validation) data, (Xt yt) load usps_digits_testing.mat whos, % show now what variables are in the workspace % the kernel parameter used is the degree of the polynomial d = 2; % now we are now going to create 10 binary classifiers we % will assign +1 to the digit we desire to learn and then % -1 to the rest, the digit we want to learn is called digit for digit = [0 1 2 3 4 5 6 7 8 9], % display some info disp(sprintf('Now learning to recognize digit %d', digit)); % create the corresponding output vector yb = zeros(size(y)); Iminus = find(y ~= digit); Iplus = find(y == digit); yb(Iminus) = -1; yb(Iplus) = 1; % fire up our kernel based perceptron algorithm [alpha, R] = perceptron_dual_kernel(X, yb, d); % now compute the test error, first what is the correct output ytb = zeros(size(y)); Iminus = find(yt ~= digit); Iplus = find(yt == digit); ytb(Iminus) = -1; ytb(Iplus) = 1; % now compute the estimated ouput for this test data Xt % note that we really need the support vectors Isv = find(alpha>0); % the bias is computed by b = alpha(Isv)'*(yb(Isv)*R^2); % compute the kernel K = kernel(X(Isv,:),Xt,d); % now compute the corresponding estimate of the output yest = sign((yb(Isv).*alpha(Isv))'*K + b)'; % how does this estimate correspond to the output error = sum([ytb ~= yest])/length(ytb)*100 % now let us keep this recognizer in a data structure called OCR OCR(digit+1).X = X(Isv,:); OCR(digit+1).y = yb(Isv,:); OCR(digit+1).alpha = alpha(Isv); OCR(digit+1).R = R; OCR(digit+1).LABEL = digit; end % OK so how do our binary classifiers perform when they compete disp('we now classify the whole test using all of our perceptrons'); for i=1:length(OCR), % compute margin for each perceptron, make sure its the geometric one! K = kernel(OCR(i).X,Xt,d); b = OCR(i).alpha'*(OCR(i).y*OCR(i).R^2); Yest(:,i) = ((OCR(i).y.*OCR(i).alpha)'*K + b)'; end % now display the result of this experiment: [dummy,I] = sort(-Yest,2); character = I(:,1)-1; total_test_error = sum([yt~=character])/length(yt)*100 % lets show at least the first 50 we failed to classify correctly Iwrong = yt~=character; plotdigits(Xt(Iwrong,:), yt(Iwrong), I(Iwrong,:)-1)