Merge branch 'master' of github.com:jervisfm/Digit-Recognizer

README

@@ -1,8 +1,45 @@
-Jervis Muindi
-Varun Ravishankar
-Biometric Project
-28th November 2011
-
+Jervis Muindi (jjm2190)
+Varun Ravishankar (vr2263)
+Biometric Final Project
+12th December 2011
+---------------------------
 About
 =======
-This is a handwritten digit recognizer that uses a nearest neighbor classifier to recognize handwritten single digits. 
+This is a handwritten digit recognizer that uses a nearest neighbor classifier to recognize handwritten single digits. 
+
+
+Files:
+=======
+readDATA.m          ->  reads in the MNIST dataset. The data should be in a subdirectory called 'MNIST'
+loadHndDATA.m       ->  reads in the manually collected images. These images should in a subdirectory called 'img'
+procTD.m            ->  Processes the MNIST training set and bounds the images to 28x28 bounding box.
+proc.m              ->  Processes a test image and bounds it in a 28x28 bounding box.
+approach1.m         ->  Does classification using raw features
+approach2.m         ->  Binds digits to a bounding box and then applies nearest neighbor classifer
+approach3.m         ->  Does nearest neighbor classification with K = 3.
+extract.m           ->  Finds the bounding box in an image with a digit and returns that bounding box.
+binarze.m           ->  Converts an image to a binary image of 0s and 1s.
+pca.m               ->  Implements PCA
+dopca.m             ->  Tests PCA apporach to see how well it does. 
+darken.m            ->  Converts white background in a image to a black background. 
+
+
+
+Before Running
+==============
+Make sure that the directory structure is as follows: 
+
+1) MNIST : This directory should contain all of the MNIST raw data set and it should be in extracted form. You can download the dataset from http://yann.lecun.com/exdb/mnist/
+2) img   : This should contain the 200 scanned images.
+
+If you do not have this structure, you will get a file not found error as the scripts expects the data to be in these locations.
+
+
+Independent Dataset
+===================
+For the classification of the independently gathered dataset, the relevant files are :
+approach1.m ; approach2.m and approach3.m and dopca.m.
+
+When you run these you scripts you will get a current estimate of how well the classifier is doing on test data. Also note that
+the scripts output estimated time of completion when carrying out computationally intensive task that can take a while to run. 
+Refers to the files description above to infer what each file does and for more details please see the report of Jervis Muindi. 

approach1.m

@@ -16,7 +16,7 @@ function   approach1()
 
 sizeTrain = size(train_data,2); 
 sizeTest = size(test_data,2); 
-sizeTest = 50;
+sizeTest = 200;
 
 %Do a quick nearest neighbor test classification. 
 errors = 0; 
@@ -64,9 +64,9 @@ function   approach1()
     if( test_labels(i) ~= result(i) )
 
         fprintf('Curr Accuracy: %f | %d,%d.\n', curr_acc, test_labels(i),result(i) );
-        figure;
+        %figure;
         im = [test_data{i} train_data{minIdx} ];
-        imshow(im);
+        %imshow(im);
         continue;
     end
 

approach3.m

@@ -1,8 +1,6 @@
-% Jervis Muindi
-% Does classification on test data set. 
-% By applying nearest neighbor directly
-% on extracted digits that have been bounded to 
-% a box and also thickening the test images by 2 pixels. 
+%Jervis Muindi
+% Biometrics
+% K-nearest neighbor with K = 3. 
 function   approach3()
 % Does a test for classifying the handwritten images. 
 
@@ -13,94 +11,6 @@ function   approach3()
 
 [test_data, test_labels] = loadHndDATA();
 
-sizeTrain = size(train_data,2); 
-sizeTest = size(test_data,2); 
-sizeTest = 50;
-
-%Do a quick nearest neighbor test classification. 
-errors = 0; 
-for i = 1:sizeTest
-
-    fprintf('test %d - ', i); 
-    tv = test_data{i}; 
-    %process the test image so that it's also bound to a box of 28 x 28 and also thicken it by adding 2 pixels to digit outline..
-    tv = proc2(tv); 
-    tv = tv(:); 
-
-    minIdx = 0;
-    minDist = Inf; 
-
-    for j = 1:sizeTrain
-
-        av = train_data{j};
-        av = av(:); 
-        diff = tv - av; 
-        dist = norm(diff,2); 
-
-        if(dist < minDist)
-            minDist = dist;
-            %disp(j);
-            minIdx = j;
-        end
-        %disp('running');
-        %disp(dist);
-
-    end
-    %error('s');
-
-    %save result; 
-
-    result(i) = train_labels{minIdx}; 
-
-
-    %do accuracy checking in line
-    if( test_labels(i) ~= result(i) )
-        errors = errors + 1;
-
-    end
-
-    tot = i;
-    curr_acc = (tot - errors) / tot; 
-
-    if( test_labels(i) ~= result(i) )
-
-        fprintf('Curr Accuracy: %f | %d,%d.\n', curr_acc, test_labels(i),result(i) );
-        figure;
-        im = [test_data{i} train_data{minIdx} ];
-        %imshow(im);
-        continue;
-    end
-
-    fprintf('Curr Accuracy: %f.\n', curr_acc);
-
-end
-
-
-
-end
-
-
-
-
-
-
-
-
-
-
-
-%{
-K NEaREST NEIGHTBOR
-
-function   testhnd()
-% Does a test for classifying the handwritten images. 
-
-disp('reading data');
-% assume, we have loaded
-[train_data, train_labels ] = readDATA();
-
-[test_data, test_labels] = loadHndDATA();
-
 sizeTrain = size(train_data,2); 
 %sizeTrain = 10000;
 sizeTest = size(test_data,2); 
@@ -114,6 +24,7 @@ function   testhnd()
 
     fprintf('test %d - ', i); 
     tv = test_data{i}; 
+    tv = proc(tv); %process the test image so that it's also bound to a box of 28 x 28.
     tv = tv(:); 
 
 
@@ -175,8 +86,8 @@ function   testhnd()
 
     if( test_labels(i) ~= result(i) )
 
-        fprintf('Curr Accuracy: %f | %d,%d.\n', curr_acc, test_labels(i),result(i) );
-        figure;
+        fprintf('Curr Accuracy: %f | %d,%d.\n', curr_acc, test_labels(i),result(i) ); % print out actual number | classified number.
+        %figure;
         continue;
     end
     fprintf('Curr Accuracy: %f.\n', curr_acc );
@@ -186,5 +97,3 @@ function   testhnd()
 
 
 end
-
-%}

darken.m

@@ -1,4 +1,6 @@
-
+% Jervis Muindi
+% Biometrics
+% Dec 2011
 
 function r = darken(A)
 %convert from white background to black background

dopca.m

@@ -0,0 +1,113 @@
+function [r,pc]  = dopca()
+%UNTITLED9 Summary of this function goes here
+%   Detailed explanation goes here
+
+
+
+% with 60,000 samples , its too much. don't have enough memory
+% on hand to compute (need 24gb).
+%
+% so just look at 10,000 samples -> this needs 0.8Gb which we have.
+%
+
+[train_data, train_labels, test_data, test_labels] = readDATA(); 
+%train_data = procTD(train_data);
+[test_data, test_labels] = loadHndDATA();
+
+
+length = size(train_data,2); 
+length = 2000;
+data = zeros(784,length); 
+
+
+for i =1:length
+    if(mod(i,1000) == 0)
+    disp(i);
+    end
+
+    img = train_data{i}; 
+    img = img(:); 
+    data(:,i) =  img;
+
+
+end
+
+disp('doing pca...');
+
+[r, pc] = pca(data); 
+%[r, pc] = pca2(data'); 
+disp('pca done');
+disp('proj data');
+size(r)
+disp('principal comp')
+size(pc)
+
+%error('s');
+disp('doiing testing of pca results...'); 
+
+train_size = length;
+train_data = r;
+errors=0;
+
+for i = 1:200
+
+
+    tv = test_data{i};
+    %tv = proc(tv);
+    tv = tv(:);
+    %tv = tv'
+    ptv = pc' * tv;
+    tv = ptv;
+
+     minIdx = 0;
+    minDist = Inf; 
+    for k = 1:train_size
+        j=k;
+         av = train_data(j);
+        av = av(:); 
+        diff = tv - av; 
+        dist = norm(diff,2); 
+
+        if(dist < minDist)
+            minDist = dist;
+            %disp(j);
+            minIdx = j;
+        end
+
+    end
+
+
+    result(i) = train_labels{minIdx}; 
+
+
+    %do accuracy checking in line
+    if( test_labels(i) ~= result(i) )
+        errors = errors + 1;
+
+    end
+    tot = i; 
+        curr_acc = (tot - errors) / tot; 
+
+
+    if( test_labels(i) ~= result(i) )
+
+        fprintf('Curr Accuracy: %f | %d,%d.\n', curr_acc, test_labels(i),result(i) );
+        %figure;
+        %im = [test_data{i} train_data{minIdx} ];
+        %imshow(im);
+        continue;
+    end
+
+    fprintf('Curr Accuracy: %f.\n', curr_acc);
+
+
+end
+
+
+
+
+
+
+end
+
+

loadHndDATA.m

@@ -39,6 +39,10 @@
     img = img / max(img(:)); % normalize values to have maximum of 1. 
     img = darken(img);  %binarize image and convert to a black background.
 
+    %do extraction and make it fill up entire 28x28
+    %img = extract(img);
+    %img = imresize(img, [28 28]);
+
     test_images{i} = img; 
     test_labels(i) = mod(i,10); 
 
@@ -48,23 +52,8 @@
 
 end
 
-% Converts from from white background to black background and
-% Binarizes the image in the process. 
-function r = darken(A)
-
-    rows = size(A,1);
-    cols = size(A,2);
-
-    for i = 1:rows
-        for j = 1:cols
-            if(A(i,j) == 1)
-                A(i,j) = 0;
-            else
-                A(i,j) = 1; %now we're binarizing img. 
-            end
-        end
-    end
-
-    r = A; 
-end
+
+
+
+
 

pca.m

@@ -1,3 +1,6 @@
+% Jervis Muindi
+% Biometrics
+% Dec 2011
 function [ signals, PC, V ] = pca( data )
 % Perform PCA. 
 % Data is the data matrix having dimensions in the columns and each image vector in the rows
@@ -24,7 +27,7 @@
 size(data)
 disp('computing conv');
 
-covariance = 1 / (N-1) * (data * data') ; 
+covariance = 1 / (N-1) * data * data' ; 
 
 disp('finding eig vals');
 %find the eigen vectors and eigen values
@@ -39,8 +42,8 @@
 
 V = sorted; 
 disp(' e v');
-V
-rindices = rindices(1:10, :);
+%V
+rindices = rindices(1:20, :);
 %size(rindices)
 
 %PC = PC(:, rindices);