*Author's Note: I'm working on an application that uses an edge detection library at Microsoft. Obviously, as a lowly undergraduate intern that I am, all I need to do is call a method and it just magically detects edges on images. I wanted to understand more about edge detection and not just use "a library". This blog post is part of my "for dummies" series, where I try to explain hard concepts in a simple way without super-sophisticated language, so that the average person (like me) can understand the concept.*
References for all the information I'm about to explain is at the bottom of this blog post.

Edge detection is what it sounds like. You have an image, and you want to detect "edges". Pictures are worth a thousand words, so the below figure should answer any of your questions on what "edge detection" is.

In the study of edge detection and image processing, an edge is formally defined as a "significant local change in the image intensity,
usually associated with a discontinuity in either the image intensity or the
first derivative of the image intensity". This sounds really hard, but bear with me here.

Let me first talk about image intensity, and a bit later in this post, I'll get back to the "first derivative of image intensity". Image intensity is a bit hard to describe, but for a normal greyscale image, like the one I'm about to show you below, it can be thought of as "dark vs light", or colour. If you want to learn more about image intensity, you can find it

here.

Now I'll try to explain more about what they mean by "discontinuity" and "first derivative of image intensity". Discontinuities in the image intensity
can be visualized as step discontinuities, where the image intensity value literally changes from one value to a totally different value. You can represent this as a step function. Discontinuities can also be in the form of line discontinuities, where the image intensity also abruptly changes value, but then returns to the starting value within some short distance. Think of a very thick, vertical, black line you draw in a paint application. As you go across the line, you go from the paint application's default white background, to black, and then to white again.

Now, in a perfect image, these discontinuities would be sharp and we can easily detect them. However, because we don't live an ideal world and because camera apps sometime like to "smoothen" captured photos, sharp discontinuities rarely exist in real photos. The step and line discontinuities you will actually obtain from processing the image will instead look like "ramp functions" or "roof functions". Examples of these two functions are drawn in the below figure.

In addition to sharp discontinuities not existing, another reason why developing a a good edge detection algorithm is hard is because images have "noise" in them, and we need to make algorithms pretty good at being able to filter out this "noise".

As you can see in this above figure, there is noise on the left most and right most part of the graph, where both parts seem to have this "ramp" function... but as you can tell from the actual image above the graph, those ramp-like parts are just noise. There is nothing on the left and right parts of the actual image.

I'm going to now try to write a short list of definitions of some commonly used terms, which will be used later in this blog post. I will try to make these terms as understandable as possible.

**Definitions**
__Edge Point__ - I'm sure you all know what a point means. An edge point is basically a point in the image that is part of what is considered to be an edge by whatever edge detection algorithm you use. So, basically, it is an (x, y) coordinate at a location where there is a significant image intensity difference.

__Edge Fragment__ - An edge fragment is basically an edge point, along with that particular point's "edge orientation". An edge orientation is basically what direction/orientation the edge is. For example, I could draw a line horizontally on my paint application, and that edge would have a horizontal orientation. Obviously, this edge orientation is an actual angle/mathematical value, but I want to make things simple.

__Edge__ - When image processing experts talk about "edges", they

*actually* mean "edge points" or "edge fragments". Read the definition below on an edge detector if you don't know what I meant from that.

__Edge Detector__ - an edge detecting algorithm that actually produces "edges" (a.k.a. edge points or edge fragments) from an image.

__Contour__ - A contour is a list of "edges" (a.k.a. edge points/edge fragments) or the mathematical curve that
models the list of "edges" (a.k.a. edge points/edge fragments), into something most of us define as an "edge".

__Edge Linking__ - "Edge linking is a process of forming an ordered list of "edges" from an unordered list. By convention, "edges" are ordered by traversal in a
clockwise direction". To simplify, edge linking is by linking edge points together on a small scale (e.g. within a few pixels of each other, we determine that edge point A and edge point B can be linked).

__Edge Following__ - "Edge following is the process of searching the {filtered} image
to determine contours." To simplify, edge following is where you look at the whole image by filtering out noises, and then determining the lines. It is edge detection on a big scale, rather than a small scale.

Edge coordinates could be in the coordinate system of the original image (e.g. an image of pixel width X and pixel height Y), though more likely, it will be in a different coordinate system as many edge detectors use a process called "filtering" in order to filter out noise. Filtering an image actually changes the coordinate system of the image, which may translate or scale the image coordinates.

Now, edge detectors cannot be perfect. The types of edges that can be obtained from an edge detector are: correct edges (so a human testing an edge detector can verify that "yep it identified this edge correctly"), false edges (where the algorithm mistaken a non-edge for an edge), and missing edges (where the algorithm should have identified an edge but didn't). False edges are called false positives, and missing edges are called false negatives.

Now, remember when I talked about the "first derivative of image intensity"? What does this mean? Well, let's take the example of a step edge associated with a significant difference in image intensity. The first derivative, if you've took an intro to calculus course, is just that - the first derivative. The slope. The below figure, taken from a neat website I referenced at the bottom of this blog post, has two images: the first one is the step function (though more like a continuous ramp function), and the graph below that is the derivative of that ramp function - a.k.a. the "first derivative of image intensity".

I will now define the "gradient". The gradient, as the figure above describes, is the first derivative of image intensity, and can be used to detect edges. How? Well, we know that the first graph can symbolize an edge because it goes from a low value to a high value very abruptly. Every time a change like this occurs, the first derivative has a local maxima! So, that means everytime a big enough local maxima occurs, we can determine that there is an edge point!

So, you can imagine that an image is just a huge 2D array of these continuous functions of image intensities, and by interpreting all of these graphs, we can eventually detect the edges in a whole image.

Now, the figure above only represents a 1D gradient and a 1D image intensity. Pictures are 2D, so we need to somehow define a 2D gradient. Below is a screenshot of a page from a textbook that describes the "formal definition" of 2D gradient:

Most edge detection algorithms uses something called a "mask" in order to approximate the gradient values in an image. An example of some different types of masks is shown in the below figure.

How do you use one of these masks, you might ask? The method is pretty simple. Let's say we are using the Sobel mask (which has a mask in the x-direction and the y-direction). For each 3x3 group of pixels, we apply Sobel's x- and y- masks by a process known as "convolution". An easy example of how to do convolution is in

this link.

These matrices that get convoluted with the original source image's matrix are not only known as masks, but also filters, because they filter out noisy signals to get the signal that we want - the edges.

There are other types of filters as well, such as ones with sizes of 2x1, 1x2, et cetera. But these these are a bit hard to think about because there isn't really a "center" pixel in that mask/matrix thing. Some of these mask matrices can be very huge as well, which we will talk about below in this blog post.

The simplest numerical approximation for gradient would be:

Gx = f[i,j + 1]- f[i,j]

and Gy = f[i, j] - f[i + 1,j]

j corresponds to the x direction and i to the negative y
direction.

The corresponding masks for this simple gradient approximation are illustrated below:

Algorithms for edge detection generally contain these three steps:

__Filtering__: Images have noise. Filtering is used to reduce this noise. However, there's actually a trade-off, because by filtering to reduce noise, you could potentially reduce edge strength in the images as well. More filtering to reduce noise results in a loss of edge strength.

__Enhancement__: Enhancement emphasizes pixels where there are local "first derivative intensity value" maxima and are usually performed by computing the gradient magnitude. There are several ways to compute the gradient (e.g. Roberts, Sobel, etc.), which we will talk about below

__Detection__: In an image, there might be lots and lots of local maxima when looking at the first derivative of image intensity. How do we actually determine what is an edge and what isn't? Typically, making a threshold of e.g. "if the maxima is above X value, then it's an edge!" is used.

Many edge detectors have been developed in the last two decades. I will discuss now talk about some commonly used edge detectors.

**Roberts Operator**

The Roberts cross operator provides a simple approximation to the gradient
magnitude:

Ok, maybe it doesn't look simple. You don't need to formally know the definitions, as you can just "use the mask". Here's the mask that you would use in the following figure though:

**Sobel Operator**

The Sobel operator is actually probably one of the most commonly used edge detectors, quoting from the below screenshot of the textbook I read. Also, the below screenshot gives you the Sobel matrix:

**Second Derivative Operators**

Now that I've talked about how to do edge detection using the gradient of image intensity, I will now try to explain how you can also use the "second derivative of image intensity" in order to do edge detection. The edge detectors I talked about earlier calculated the first derivative and, if it
was above a certain X threshold, the presence of an edge point was assumed. This actually results in the detection of too many edge points. A better approach would be to find *only*
the points that have local maxima in gradient values and consider them edge
points. This means that at edge points, there will be
a peak in the first derivative and, equivalently, there will be a __zero crossing__
in the second derivative.

Thus, edge points may be detected by finding the
zero crossings of the second derivative of the image intensity. This is the second major way of edge detection, instead of just the "gradient" way.

There are two operators in two dimensions that correspond to the second derivative: the Laplacian and second directional derivative. This sounds hard, but basically, there are two ways at looking at the second derivative, and correspondingly, two different ways of making an algorithm for edge detection using the "zero crossing" method.

I won't explain too much into this, other than the fact that Laplacian and second directional derivative can give you masks that you can use in your edge detection code, just like the "gradient" methods:

I will make a follow up blog-post that goes more into zero-crossing methods of edge detection, including talking about the canny-edge detector algorithm.

-------------

REFERENCES