What Is Computer Vision Technology?

Computer vision is a technique that converts still image or video data into a decision or a new representation. All such transformations are done to accomplish a specific purpose.

The input data may contain some scene information, such as "the camera is mounted on a vehicle" or "the radar detected a target one meter away". A new representation that means converting a color image to black and white, or removing the effects of camera motion from an image sequence.

human vision

Because we are creatures endowed with vision, it's easy to mistake "computer vision is also a very simple task". How hard is computer vision?

Please tell me how you observed a car from an image. Your initial intuition can be highly misleading. The human brain divides visual signals into many channels to allow different streams of information to enter the brain. The brain has been shown to have an attentional system that, in a task-based manner, examines the estimates of other regions through important parts of the image. There is a huge amount of feedback in the visual information flow, and little is known about this process until now.

The extensive interconnectedness of muscle-controlled perceptrons and all other senses allows the brain to exploit the cross-associations that arise from many years of human experience living in the world. Feedback loops in the brain deliver feedback to every process, including The human body's sensory organ (the eye) regulating the retina's perception of surfaces by physically controlling the amount of light through the iris.

computer vision

In a machine vision system, however, the computer receives a grid of numbers from a camera or hard drive, which means, crucially, there is no pre-established pattern recognition mechanism for the machine vision system. There is no automatic control of focal length and aperture, nor can years of experience be tied together. Most visual systems are still in a very primitive stage.

Figure 1 shows a car. In this image, we see that the mirrors are located next to the cab. But to the computer, all it sees is numbers arranged in a grid. All the numbers are given in the grid also have a lot of noise, so each number can only give us a small amount of information, but this grid of numbers is all the computer can "see". Our task becomes to convert this noisy digital raster into the perceptual result "rearview mirror".

If the solution of a mathematical and physical definite solution problem exists, is unique and stable, the problem is well-posed; if not, the problem is ill-posed.

In fact, this problem, as we have mentioned before, is not enough to describe it as "difficult", and it is simply impossible to solve in many cases.

Given a two-dimensional (2D) observation of the 3D world, there is no single way to reconstruct the 3D signal. Even if the data were perfect, the same 2D image could represent any one of an infinite combination of 3D scenes.

Also, as mentioned earlier, data can be contaminated with noise and distortion. Such contamination arises from many aspects of real-life (weather, light, refractive index, and motion), as well as circuit noise in the sensor and other circuit system effects, as well as effects on image compression after acquisition.

Under the influence of this series, how can we push things forward?

In classical system design, additional scene information can help us improve the quality of information obtained from the sensor level.

Scene information can aid computer vision

Consider an example where a mobile robot needs to find and pick up a stapler in a building. Robots might take advantage of the fact that decks are usually placed in offices and staplers are usually stored in desks. This also gives an inference about size: the stapler must be large enough to fit on the table.

Taking it a step further, this can also help reduce the chance of misidentifying the stapler in places where the stapler is unlikely to be present (such as ceilings or windows). The robot can safely ignore the 200-foot-tall stapler-shaped airship because the airship does not satisfy the prior information being placed on a wooden tabletop.

In contrast, in tasks such as image retrieval, all stapler images in the dataset are from real staplers, so unreasonable sizes and some strange shapes will be implicitly eliminated when we collect images— —Because photographers would just go and shoot regular size staplers. People also tend to shoot with the subject in the middle of the image and tend to shoot at the angle that best characterizes the subject. As a result, there is usually a lot of unintentional additional information that people unintentionally add to their photos.

Scene information can also be modeled (especially by machine learning techniques). Implicit variables (such as size, the direction of gravity, etc., which are not easily observed directly) can be found and inferred from labeled data sets. Alternatively, you can try to use additional sensors to measure the value of implicit variables, such as using lidar to measure depth, so as to accurately obtain the size of the target.

Use statistical methods to combat noise

The next problem faced by computer vision is noise, and we generally use statistical methods to combat noise.

For example, it is difficult for us to judge whether it is an edge point by a single pixel and its adjacent pixels, but if we observe its statistical regularity in an area, edge detection will become easier.

A true edge should appear as a series of independent points within an area, all oriented in the same direction as their closest point. We can also suppress noise through cumulative statistics over time, and of course, there are methods to eliminate noise by establishing noise models through existing data. For example, because lens distortion is easy to model, we only need to learn a simple polynomial model to describe the distortion to almost perfectly rectify distorted images.

Based on camera data, computer vision prepares to make actions or decisions that are performed in the context of a specific purpose or task scene. We might want to remove noise or fix damaged photos, so a security system can alert on dangerous behaviors like trying to climb a railing or count the number of people crossing a certain playground area.

And the vision software of a robot roaming the building will take a completely different strategy than the security system because the two strategies are in different contexts. In general, the tighter the constraints of the environment a vision system is exposed to, the more we can rely on those constraints to simplify the problem, and the more reliable our final solution will be.

The goal of OpenCV is to provide tools for problems that computer vision needs to solve. In some cases, advanced functions in function libraries can effectively solve problems in computer vision. Even when faced with problems that cannot be solved in one go, the base components in the library are sufficiently complete to enhance the performance of the solution for arbitrary computer vision challenges.

In the latter case, there are also some reliable ways to use libraries, all of which start by using as many different component libraries as possible. Often, after developing the first rough solution, you can discover what flaws the solution has and fix those flaws with your own code and ingenuity (better known as "solve the real problem, not your imagined problems"). After this, the level of improvement can be evaluated using the rough solution as a judging criterion. From this point, you can solve arbitrary problems.

Source: Labkom99