Classifiying a set of Images into Classes

Question

I have the problem that I get a set of pictures and need to classify those.

The thing is, i do not really have any knowledge of these images. So i plan on using as many descriptors as I can find and then do a PCA on those to identify only the descriptors that are of use to me.

I can do supervised learning on a lot of datapoints, if that helps. However there is a chance that pictures are connected to each other. Meaning there could be a development from Image X to Image X+1, although I kinda hope this gets sorted out with the information in each Image.

My question are:

1. How do i do this best when using Python? (I want to make a proof of concept first where speed is a non-issue). What libraries should i use?
2. Are there examples already for an image Classification of such a kind? Example of using a bunch of descriptors and cooking them down via PCA? This part is kinda scary for me, to be honest. Although I think python should already do something like this for me.

Edit: I have found a neat kit that i am currently trying out for this: http://scikit-image.org/ There seem to be some descriptors in there. Is there a way to do automatic feature extraction and rank the features according to their descriptive power towards the target classification? PCA should be able to rank automatically.

Edit 2: I have my framework for the storage of the data now a bit more refined. I will be using the Fat system as a database. I will have one folder for each instance of a combination of classes. So if an image belongs to class 1 and 2, there will be a folder img12 that contains those images. This way i can better control the amount of data i have for each class.

Edit 3: I found an example of a libary (sklearn) for python that does some sort of what i want to do. it is about recognizing hand-written digits. I am trying to convert my dataset into something that i can use with this.

here is the example i found using sklearn:

import pylab as pl

# Import datasets, classifiers and performance metrics
from sklearn import datasets, svm, metrics

# The digits dataset
digits = datasets.load_digits()

# The data that we are interested in is made of 8x8 images of digits,
# let's have a look at the first 3 images, stored in the `images`
# attribute of the dataset. If we were working from image files, we
# could load them using pylab.imread. For these images know which
# digit they represent: it is given in the 'target' of the dataset.
for index, (image, label) in enumerate(zip(digits.images, digits.target)[:4]):
    pl.subplot(2, 4, index + 1)
    pl.axis('off')
    pl.imshow(image, cmap=pl.cm.gray_r, interpolation='nearest')
    pl.title('Training: %i' % label)

# To apply an classifier on this data, we need to flatten the image, to
# turn the data in a (samples, feature) matrix:
n_samples = len(digits.images)
data = digits.images.reshape((n_samples, -1))

# Create a classifier: a support vector classifier
classifier = svm.SVC(gamma=0.001)

# We learn the digits on the first half of the digits
classifier.fit(data[:n_samples / 2], digits.target[:n_samples / 2])

# Now predict the value of the digit on the second half:
expected = digits.target[n_samples / 2:]
predicted = classifier.predict(data[n_samples / 2:])

print("Classification report for classifier %s:\n%s\n"
      % (classifier, metrics.classification_report(expected, predicted)))
print("Confusion matrix:\n%s" % metrics.confusion_matrix(expected, predicted))

for index, (image, prediction) in enumerate(
        zip(digits.images[n_samples / 2:], predicted)[:4]):
    pl.subplot(2, 4, index + 5)
    pl.axis('off')
    pl.imshow(image, cmap=pl.cm.gray_r, interpolation='nearest')
    pl.title('Prediction: %i' % prediction)

pl.show()

Answer 1

You can convert a picture to a vector of pixels, and perform PCA on that vector. This might be easier than trying to manually find descriptors. You can use numPy and sciPy in python. For example:

import scipy.io
from numpy import *
#every row in the *.mat file is 256*256 numbers representing gray scale values
#for each pixel in an image. i.e. if XTrain.mat has 1000 lines than each line
#will be made up of 256*256 numbers and there would be 1000 images in the file.
#The following loads the image into a sciPy matrix where each row is a vector
#of length 256*256, representing an image. This code will need to be switched
#out if you have a different method of storing images.
Xtrain = scipy.io.loadmat('Xtrain.mat')["Xtrain"]
Ytrain = scipy.io.loadmat('Ytrain.mat')["Ytrain"]
Xtest = scipy.io.loadmat('Xtest.mat')["Xtest"]
Ytest = scipy.io.loadmat('Ytest.mat')["Ytest"]
learn(Xtest,Xtrain,Ytest,Ytrain,5) #this lowers the dimension from 256*256 to 5

def learn(testX,trainX,testY,trainY,n):
    pcmat = PCA(trainX,n)
    lowdimtrain=mat(trainX)*pcmat #lower the dimension of trainX
    lowdimtest=mat(testX)*pcmat #lower the dimension of testX
    #run some learning algorithm here using the low dimension matrices for example
    trainset = []    

    knnres = KNN(lowdimtrain, trainY, lowdimtest ,k)
    numloss=0
    for i in range(len(knnres)):
        if knnres[i]!=testY[i]:
            numloss+=1
    return numloss

def PCA(Xparam, n):
    X = mat(Xparam)
    Xtranspose = X.transpose()
    A=Xtranspose*X
    return eigs(A,n)

def eigs(M,k):
    [vals,vecs]=LA.eig(M)
    return LM2ML(vecs[:k])

def LM2ML(lm):
    U=[[]]
    temp = []
    for i in lm: 
       for j in range(size(i)):
           temp.append(i[0,j])
       U.append(temp)
       temp = []
    U=U[1:]
    return U

In order to classify your image you can used k-nearest neighbors. i.e. you find the k nearest images and label your image with by majority vote over the k nearest images. For example:

def KNN(trainset, Ytrainvec, testset, k):
    eucdist = scidist.cdist(testset,trainset,'sqeuclidean')
    res=[]
    for dists in eucdist:
        distup = zip(dists, Ytrainvec)
        minVals = []
    sumLabel=0;
    for it in range(k):
        minIndex = index_min(dists)
        (minVal,minLabel) = distup[minIndex]
        del distup[minIndex]
        dists=numpy.delete(dists,minIndex,0)
        if minLabel == 1:
            sumLabel+=1
        else:
            sumLabel-=1
        if(sumLabel>0):
            res.append(1)
        else:
            res.append(0)
    return res

Answer 2

I know I'm not answering your question directly. But images vary greatly:remote sensing, objects, scenes, fMRI, biomedial, faces, etc... It would help if you narrow your categorization a bit and let us know.

What descriptors are you computing? Most of the code I use (as well as the computer vision community) is in MATLAB, not in python, but I'm sure there are similar codes available (pycv module & http://www.pythonware.com/products/pil/). Try out this descriptor benchmark that has precompiled state-out-the-art code from the people at MIT: http://people.csail.mit.edu/jxiao/SUN/ Try looking at GIST,HOG and SIFT, those are pretty standard depending on what you wanto to analyze: scenes, objects or points respectively.