Corner detection – find the features of the image

Corner detection, as its name suggests, detects corners in images. A corner is essentially a local feature that can be easily located in an image and is widely found in man-made objects such as tables, chairs, cars, houses, etc. (There are many curves in nature, which you can observe while walking). Also useful is the fact that corners can be accurately detected (even with sub-pixel accuracy).

As for the use of corners, essentially, corners are used as a feature of the image. We can imagine a jigsaw puzzle. Obviously, corner information is of great value for image stitching and alignment, and changes in images or objects in images, such as image detection and motion tracking, are also of great value.

Harris corner detection comes from the basic hypothesis of human eye corner point recognition, we choose a small window for the image as part of our actual observation, if a dot moving up and down or so after he is bigger, the variance of the surrounding so he should be a corner point, if the direction where a large change in one direction change is small, then he is likely to be a line. So we can get the corner points in the region by setting up a fixed window and comparing its left and right changes.

Also, therefore, we need to set three parameters: the size of the window, the size of the domain (i.e. the corner range, or how big the corner might be), and the threshold. For the Harris algorithm, we can call the cornerHarris algorithm in OpenCV directly, and the three key parameters are

• BlockSize – The size of the field to consider in corner detection.
• Ksize — the window size used in Sobel derivatives
• K — Free parameter in Harris corner detection equation, value parameter is [0.04,0.06].

Here is the code. Here we used an image from OpencV’s Github project. You can download the original image from OpencV’s Github project or download all the code and images from Fongtian’s Github Project — Talking About Computer Vision:

img = cv2.imread('/home/fonttian/Data/image/OpenCV/blox.jpg') gray = cv2.cvtColor(img, Cv2. COLOR_BGR2GRAY) # cornerHarris Gray = NP. Float32 (Gray) blockSize = 2 # Field size ksize = 3 # Aperture size k = 0.04 # Harris DST = Cv2.cornerharris (SRC =gray, blockSize=blockSize, ksize=ksize, k=k) # cv2.cornerHarris(SRC =gray, blockSize=blockSize, ksize=ksize, k=k) # Set the pixel of the image to True, otherwise set it to False. A = DST > 0.01 * dst.max() img[a] = [0, 0, 255] # show cv2.imshow('corners', img)Copy the code

In the above code, we don’t use the focus data directly. Instead, we filter the data and draw the corner points, which avoids the problem of too many corners. In addition, we can simplify our code by using a more suitable corner detector: the goodFeaturesToTrack function. See the following code for details.

What we have said before is direct corner detection, and corner is essentially a feature point of an image. In addition to these algorithms that directly obtain corner points, we also need many other feature point detection algorithms that can be directly used, and these feature point detection algorithms are also inseparable from various basic operators. Take the ORB algorithm.

RB feature is a combination of FAST feature point detection method and BRIEF feature descriptor, which is improved and optimized on their original basis.

Firstly, it uses the FAST feature point detection method to detect feature points, and then uses Harris corner point measurement method to pick out N feature points with maximum Harris corner response value from FAST feature points.

In addition to the above description, ORB algorithm itself also involves a lot of content, such as rotation invariance (FAST feature itself does not have), multi-scale invariance, non-maximum suppression, etc. It requires a lot of ink to understand the feature extraction algorithm similar to ORB algorithm in detail, so we will talk about it in the next chapter. However, using OpenCV is quite simple:

img = cv2.imread('/home/fonttian/Data/image/OpenCV/blox.jpg')

orb = cv2.ORB_create()

kp = orb.detect(img, None)

kp, des = orb.compute(img, kp)

img = cv2.drawKeypoints(img, kp, img, color=(0.255.0), flags=0)

cv2.imshow('p', img)
Copy the code

The first line of code above reads the image, the second line creates the ORB, then uses Detect for feature detection, compute for feature calculation, and finally draw. Des is not actually used here, and we’ll talk about how to use it later. But only from the code analysis, we can clearly see that only one line of code is needed to create, detect, and draw, and draw this step, we even need to call OpenCV encapsulated classes, can be said to be very simple. The effect is as follows. We can also modify the specific drawing effect by parameters:

Similar to ORB, there are many other useful feature detectors in OPECV that are also easy to use. We’ll cover that in the next article.

• Starting with noise removal, what kind of PHOTOSHOP do you need with OpenCV?
• The cornerstone of computer vision – filtering
• Discover your body shape – OpenCV image contours
• Image edge detection, detection or simplification