Figure of image recognition

Progarm Description

get_tiny_images.m

It is a function to implement the tiny images representations. The tiny images is very simple method. The "tiny image" feature function implements the image representations by resizing each image to a small, fixed resolution.

nearest_neighbor_classify.m

It is a function to implement the nearest neighbor classify images representations. In this function, first, find the distance between each pairs of the test images and the train images, then find the "nearest" train images to every test images, assign the categories of train images to the test images

build_vocabulary.m

The function is used to build the vocabulary of visual words.First, randomly choose some train images from every categories. Then find the sift features for each images, at last, cluster all features with kmeans.

get_bags_of_sifts.m

This function is used to implement the images representation of Bag of words models. First, for each images, we get their sift features,then by assigning features to each cluster in visual word, we build a histogram of image. After normalising, the histogram will be used as the features of image

Code Example

load('vocab.mat')
vocab_size = size(vocab,2)


N=size(image_paths,1);
image_feats=zeros(N,vocab_size);

for i=1:N
    s = num2str(cell2mat(image_paths(i)));
    A=imread(s);

    A = mat2gray(A);
    [locations, SIFT_features] = vl_dsift(single(A),'step',5,'size',6,'fast');

    D=vl_alldist2(single(SIFT_features),vocab);
    [Y,I]=min(D,[],2);
    his=[];
    binranges=1:vocab_size;
    his=histc(I,binranges);

    nor_his=his/norm(his);
    image_feats(i,:)=nor_his;
    
end

svm_classify.m

This function is used to implement the 1-vs-alllinear SVMs.For each categories, we will get a linear function, then use this function to score every images, after this, for each images, get the largest score in all categories, assign this categories to this image

Results

Basic Result

Performance with tiny images and nearest neighbor classifier	Performance with tiny images and linear SVM classifier	Performance with bag of SIFT and nearest neighbor classifier
0.191	0.529	0.723

Result in different parameters

images num used in building vocabulary	vocabulary sizes	lambda for vl_svmtrain	step(GBS)	size(GBS)	step(BV)	size(BV)	Accuracy	Time
150	400	0.0001	5	6	8	10	0.709	25min
600	400	0.0001	5	6	8	10	0.705	32min
300	400	0.0001	4	5	8	10	0.703	41min
300	300	0.0001	4	5	8	10	0.701	31min
300	400	0.00001	5	6	6	8	0.670	27min
300	400	0.0001	5	6	6	8	0.723	27min
300	400	0.001	5	6	6	8	0.655	27min
300	400	0.01	5	6	6	8	0.492	27min
300	400	1	5	6	6	8	0.437	27min
300	400	10	5	6	6	8	0.477	27min
300	400	0.0001	5	6	8	10	0.703	27min
300	400	0.0001	6	8	8	10	0.691	19min
300	400	0.0001	8	10	8	10	0.667	12min
300	400	0.0001	25	30	25	40	0.585	2min

Result in different vocabulary sizes

(Get_bags_of_sifts: setp 5 size 6; build_vocabulary: step,8,size,10)

vocabulary sizes	SVM	1NN	Time
10	0.473	0.387	2min
20	0.577	0.467	3min
50	0.645	0.489	5min
100	0.681	0.493	8min
200	0.691	0.509	14min
400	0.696	0.525	21min
500	0.709	0.525	34min
1000	0.712	0.525	66min
10000	0.727	0.493	6h36min

Report Analysis

Based on the reports above, we can get the result that the tiny images and nearest neighbor classifier is the simpler and less method to implement Image Recognition,with Accuracy about 10%~20%, the linear SVM classifier is better than nearest neighbor classifier, it can improve the accuracy to 50%~60%. The bag of SIFT is better than tiny images, though much slower. It has accuracy of 60%~70%.With turing parameters(number of clusters, SVM regularization,number of pathces sampled when building vocabulary, size and step for dense SIFT features) well. When using bag of SIFT and nearest neighbor classifier, the, lager vocabulary number and lower step and size number will help to increase the accuracy, but slow down running time of the whole project. With different vocabulary sizes, the size increased, the accuracy and running time will increased fast, when the size reaches about 400, the accuracy for SVM will reach about 70%,and 1NN will reach about 50%, the speed of increasing will slow down.