Introduction to image processing algorithms

In the early days of the field of machine vision, a common research paradigm was to view images as two-dimensional digital signals and then use methods from digital signal processing, which is image processing. Today, I will introduce to you the related technologies of image processing.

Point Operation

Some operations only operate on each pixel of the image to produce a new image. Binarization is a typical example. When the threshold is set in advance, the output of the binarization depends only on the value of the point, so

Where fA and fB are the input and output images respectively. This operation can be done once the image data is passed sequentially using a lookup table, LUT. Various grayscale corrections are also such operations, the difference is that binarization produces a binary image, while correction still produces a grayscale image.

Local Operation

The output of the local operation is still an image, and the value of each pixel in the output image depends on the corresponding pixel in the input image and its surrounding pixels.

N in the formula represents the local sub-image centered at [i, j]. An example of this operation is shown in the figure below. This operation appears in almost every machine vision system. Its input image can be a grayscale image or a binary image; it can be used for smoothing, sharpening, denoising, thinning, edge detection and other operations. There are many ways to select local pixels, such as cross, square, honeycomb, etc., but the most commonly used is square. For example, a 3×3 or 5×5 or 7×7 square template is used. Each position in the template has a kernel coefficient. Using this template to convolve each pixel and its adjacent pixels is the most commonly used operation. Others such as median filtering are also good algorithms for removing noise.

Global Algorithm

Some algorithms are based on the entire input image, which are called global operators.

An example of this type of algorithm is shown in the figure below.

Its output can be an image or a symbolic output. Histogram, Fourier transform, and generalized Hough transform are all global algorithms.

Object-level algorithms

In most machine vision applications, it is required to calculate the characteristics of objects in the image. In order to identify these objects, size, average grayscale, perimeter, center of gravity, shape and other characteristics are often used, which are directly calculated on the target object; in defect detection, they are also obtained by detecting the target object. This leads to a very difficult, but crucial question in machine vision systems: What is the target object? Where is it?

Many algorithms in machine vision revolve around the purpose of knowing where objects are in an image. Objects in an image can also present difficult decisions.

For example, we must use all points that belong to the target object to calculate the characteristics of the object; but we must also use these characteristics to distinguish whether these points belong to the object.

Therefore, successfully segmenting foreground objects from background pixels is the key to the success of the visual system. In order to fully understand the content of the image, machine vision must perform multiple operations on the target object in order to make the correct segmentation.

The following figure demonstrates the chain code obtained after tracking the edge of an object. The chain code can be used to directly classify the shape; the chain code can also be transformed into a Fourier transform to highlight the component reflecting the shape. From this example, it can be seen that the object-level algorithm is not as regular as the three types of algorithms described above. It can be performed in the order of pixel scanning, which is convenient for using special chips to complete in real time. Object target-level algorithms are often more complex and are only suitable for processors to execute, such as PC CPUs, dedicated signal processors DSPs, etc. At the same time, the operation time of these algorithms is not often fixed, but varies with the complexity of the image content.