Chapter 9. Classical Machine Learning Methods in Image Processing

In this chapter, we will discuss the application of machine learning techniques in image processing. We will define machine learning and learn about two of its algorithms, supervised and unsupervised. Then, we will continue our discussion on the application of a few popular unsupervised machine learning techniques, such as clustering, and problems such as image segmentation.

We will also be looking at applications of supervised machine learning techniques for problems such as image classification and object detection. We will be using a very popular library, scikit-learn, along with scikit-image and Python-OpenCV (cv2) to implement machine learning algorithms for image processing. This chapter is going to give you insight into machine learning algorithms and the problems they solve.

 The topics to be covered in this chapter are as follows:

• Supervised versus unsupervised learning 
• Unsupervised machine learning—clustering, PCA,...

Supervised versus unsupervised learning

Machine learning algorithms can primarily be of two types:

• Supervised learning: In this type of learning, we are given an input dataset along with the correct labels, and we need to learn the the relationship (as a function) between the input and the output. The handwritten-digit classification problem is an example of a supervised (classification) problem.
• Unsupervised learning: In this type of learning, we have little or no idea what our output should look like. We can derive structure from data where we don't necessarily know the effect of the variables. An example is clustering, which can also be thought of as segmentation, in image processing technique where we do not have any prior knowledge of which pixel belongs to which segment.

A computer program is said to learn from experience, E, with respect to some task, T, and some performance measure, P, if its performance on T, as measured by P, improves with experience, E.

For example, let's say that...

Unsupervised machine learning – clustering, PCA, and eigenfaces

In this section, we will discuss a few popular machine learning algorithms along with their applications in image processing. Let's start with a couple of clustering algorithms and their applications in color quantization and the segmentation of images. We will use the scikit-learn library's implementation for these clustering algorithms.

K-means clustering for image segmentation with color quantization

In this section, we will demonstrate how to perform a pixel-wise Vector Quantization (VQ) of the pepper image, reducing the number of colors required to show the image from 250 unique colors down to four colors, while preserving the overall appearance quality. In this example, pixels are represented in a 3D space and k-means is used to find four color clusters.

In image processing literature, the codebook is obtained from k-means (the cluster centers) and is called the color palette. In a color palette, using a single byte, up to...

Supervised machine learning – image classification

In this section, we will discuss the image classification problem. The input dataset that we will use is MNIST (http://yann.lecun.com/exdb/mnist/), which is a classic dataset in machine learning, consisting of 28 x 28 grayscale images of handwritten digits.

 

 

The original training dataset contains 60,000 examples (the handwritten digit images along with the labels to train machine learning models), and the test dataset contains 10,000 examples (the handwritten digit images along with the labels as ground-truths, to test the accuracy of the models learned). Given a set of handwritten digits and images along with their labels (0-9), the goal will be to learn a machine learning model that can automatically identify the digit from an unseen image and assign a label (0-9) to the image:

1. A few supervised machine learning (multi-class classification) models (classifiers) will be trained using the training dataset
2. Then, they will be used to predict...

Supervised machine learning – object detection

So far, we have demonstrated how to use the classification model to classify an image, for example, to use binary classification to find if an image contains the handwritten digit 1 or not. In the next section, we will see how to use supervised machine learning models, not only to check whether an object is in an image or not, but also to find the location of the object in the image (for example, in terms of a bounding box—a rectangle the object is contained in).

Face detection with Haar-like features and cascade classifiers with AdaBoost – Viola-Jones

As we have discussed briefly in Chapter 7, Extracting Image Features and Descriptors, (in the context of Haar-like feature extraction), the Viola-Jones' object detection technique can be used for face detection in images. It is a classical machine learning approach, where a cascade function is trained using a training set of positive and negative images, by using the hand-crafted Haar-like features...

Summary

In this chapter, we discussed a few classical machine learning techniques and their applications in solving image processing problems. We started with unsupervised machine learning algorithms such as clustering and principal component analysis. We demonstrated k-means and spectral clustering algorithms with scikit-learn and showed you how they can be used in vector quantization and segmentation. Next, we saw how PCA can be used in dimension reduction and the visualization of high-dimensional datasets such as the scikit-learn handwritten digits images dataset. Also, how the PCA can be used to implement eigenfaces using a scikit-learn face dataset was illustrated.

Then, we discussed a few supervised machine learning classification models, such as kNN, the Gaussian Bayes generative model, and SVM to solve problems such as the classification of the handwritten digits dataset. Finally, we discussed a couple of classical machine learning techniques for object detection in images, namely...

Questions

1. Use k-means clustering for thresholding an image (use number of clusters=2) and compare the result with Otsu's.
2. Use scikit-learn's cluster.MeanShift() and mixture.GaussianMixture() functions to segment an image with mean shift and GMM-EM clustering methods, respectively—another two popular clustering algorithms.
3. Use Isomap (from sklearn.manifold) for non-linear dimension reduction and visualize 2-D projections. Is it better than linear dimension reduction with PCA? Repeat the exercise with TSNE (again from sklearn.manifold).
4. Write a Python program to show that the weighted linear combination of a few dominating eigenfaces indeed approximates a face.
5. Show that eigenfaces can also be used for naive face-detection (and recognition) and write Python code to implement this (hint—refer to this article: https://sandipanweb.wordpress.com/2018/01/06/eigenfaces-and-a-simple-face-detector-with-pca-svd-in-python/).
6. Use PCA to compute eigendigit-based vectors from the MNIST dataset (this is similar...

Further reading

• Viola, Paul, and Michael J. Jones, Robust real-time face detection. International journal of computer vision: http://www.merl.com/publications/docs/TR2004-043.pdf
• Navneet Dalal and Bill Triggs, Histograms of Oriented Gradients for Human Detection: https://lear.inrialpes.fr/people/triggs/pubs/Dalal-cvpr05.pdf