Day 56 of 100DaysofML

Relation between Image Processing and Machine Learning. This is a topic which I read about after one of the comments on my LinkedIn mentioned about. So, what exactly is the difference or relation between Image Processing and Machine Learning.

I shall mention a few opinions based on what I read and found online and then put in my own opinion.

Computer Vision and Machine Learning are two core branches of Computer Science that can function, and power very sophisticated systems that rely on CV and ML algorithms exclusively but when you combine the two, you can achieve even more. That is to say, you don’t need to learn one before the other or even simultaneously and there are specific use cases or domains where you would consider a marriage between the two.

Machine learning and computer vision are closely related. Computer vision uses techniques from machine learning and, in turn, some machine learning techniques are developed especially for computer vision.

The main difference is in focus (heh): machine learning is more broad, unified not by any particular task but by similar techniques and approaches. Many machine learning algorithms and systems are pretty agnostic to what the machine is working on — you just need to provide the right set of inputs and features to the algorithm. Very similar classifiers can be used to block spam or identify pictures of cats. Computer vision, on the other hand, is unified by a set of tasks: dealing with images. This requires a fair amount of different technologies — a fair bit of machine learning, to be sure, but also things from AI and signals processing and other fields. Since the two fields share techniques and applications, they are pretty close. Chances are someone working in computer vision also has a fair amount of experience with machine learning, and somebody in machine learning has at least some exposure to computer vision. (On the other hand, you wouldn’t expect either one to have any experience with programming language theory, for example.)

Traditionally, “machine vision” has focused more on the use of cameras and video in industrial settings where light and motion are controlled and where the objects to be viewed are already known and almost all observed events are predictable. But the term seems to be used less often now than in years past, so maybe it’s less useful to differentiate MV from the larger umbrella term of CV.

Computer vision includes MV but also deals with aspects of 2D and 3D images and vision that are uncontrolled, often unpredictable, and where objects, their activities, and the surrounding world extend much further into the unknown. Machine learning is now used extensively in CV yet less in MV, I suspect. In recent years CV has expanded its ambitions significantly from mere image/video analysis, into 3D shape reconstruction, CAD/CAM model creation, human pose estimation, scene interpretation, and a wealth of topics useful to robotics like mapping of routes, and the discovery and characterization of physical surroundings and scenery.

Image Processing and Computer Vision normally go hand in hand. A large number of concepts do overlap, and there is no clear demarcation between the two fields.

In general, one can consider this scenario to be the difference between Image Processing and Computer Vision. Loosely speaking, those applications where you want the output of an algorithm to be another image which is mainly used for visual applications can be regarded as Image Processing. Algorithms of Computer Vision mostly provide you an “objective” output — it could be a set of features (e.g., SIFT, HoG etc.), or it could be a set of coordinates of a tracked object, or the identification of a face or any recognized object.

Having said this, not all algorithms that provide an image can be considered as image processing. The best example to prove this statement is Image-Based-Rendering (IBR), which provides an image at some virtual viewpoint (where the camera wasn’t there during acquisition stage). Though the output of IBR is an image, the concepts, assumptions and methodologies involved originate from the theory of Computer Vision.

That basically summarizes everything I wanted to mention. Check out the below given link to understand more about the topic.

That’s it for today. Thanks for reading.

Cheers.