In this chapter, we will cover the following recipes:
Detecting image contours with the Canny operator
Detecting lines in images with the Hough transform
Fitting a line to a set of points
Extracting the components' contours
Computing components' shape descriptors
In order to perform content-based analysis of an image, it is necessary to extract meaningful features from the collection of pixels that constitute the image. Contours, lines, blobs, and so on, are fundamental image primitives that can be used to describe the elements contained in an image. This chapter will teach you how to extract some of these important image features.
In the previous chapter, we learned how it is possible to detect the edges of an image. In particular, we showed you that by applying a threshold on the gradient magnitude, a binary map of the main edges of an image can be obtained. Edges carry important visual information since they delineate the image elements. For this reason, they can be used, for example, in object recognition. However, simple binary edge maps suffer from two main drawbacks. First, the edges that are detected are unnecessarily thick; this makes the object's limit more difficult to identify. Second, and more importantly, it is often impossible to find a threshold that is sufficiently low in order to detect all important edges of an image and is, at the same time, sufficiently high in order to not include too many insignificant edges. This is a trade-off problem that the Canny algorithm tries to solve.
In our human-made world, planar and linear structures abound. As a result, straight lines are frequently visible in images. These are meaningful features that play an important role in object recognition and image understanding. The Hough transform is a classic algorithm that is often used to detect these particular features in images. It was initially developed to detect lines in images and, as we will see, it can also be extended to detect other simple image structures.
With the Hough transform, lines are represented using the following equation:
The ρ parameter is the distance between the line and the image origin (the upper-left corner), and θ is the angle of the perpendicular to the line. Under this representation, the lines visible in an image have a θ angle between 0 and π radians, while the ρ radius can have a maximum value that equals the length of the image diagonal. Consider, for example, the following set of lines...
In some applications, it could be important to not only detect lines in an image, but also to obtain an accurate estimate of the line's position and orientation. This recipe will show you how to find the line that best fits a given set of points.
The first thing to do is to identify points in an image that seem to be aligned along a straight line. Let's use one of the lines we detected in the preceding recipe. The lines detected using cv::HoughLinesP are contained in std::vector<cv::Vec4i> called lines. To extract the set of points that seem to belong to, let's say, the first of these lines, we can proceed as follows. We draw a white line on a black image and intersect it with the Canny image of contours used to detect our lines. This is simply achieved by the following statements:
int n=0; // we select line 0
// black image
cv::Mat oneline(contours.size(),CV_8U,cv::Scalar(0));
// white line
cv::line(oneline,
...Images generally contain representations of objects. One of the goals of image analysis is to identify and extract these objects. In object detection/recognition applications, the first step is often to produce a binary image that shows you where certain objects of interest could be located. No matter how this binary map is obtained (for example, from the histogram back projection we did in Chapter 4, Counting the Pixels with Histograms, or from motion analysis as we will learn in Chapter 11, Processing Video Sequences), the next step is to extract the objects that are contained in this collection of 1s and 0s.
Consider, for example, the image of buffaloes in a binary form that we manipulated in Chapter 5, Transforming Images with Morphological Operations, as shown in the following figure:
We obtained this image from a simple thresholding operation followed by the application of open and close morphological filters. This recipe will show you how to extract...
A connected component often corresponds to the image of an object in a pictured scene. To identify this object, or to compare it with other image elements, it can be useful to perform some measurements on the component in order to extract some of its characteristics. In this recipe, we will look at some of the shape descriptors available in OpenCV that can be used to describe the shape of a connected component.
Many OpenCV functions are available when it comes to shape description. We will apply some of them on the components that we have extracted in the preceding recipe. In particular, we will use our vector of four contours corresponding to the four buffaloes we previously identified. In the following code snippets, we compute a shape descriptor on the contours (contours[0] to contours[3]) and draw the result (with a thickness of 2) over the image of the contours (with a thickness of 1). This image is shown at the end of this section...
© 2014 Packt Publishing Limited All Rights Reserved
