"Traditional image algorithms have formed a very mature methodology in the development of the past decades, but the deep learning technology developed in recent years, with the increasing practicability of the theory and the increasing number of landing scenes, has constantly impacted the status of traditional image algorithms in visual applications."

In today's industrial visual application, when we want to solve the production through visual recognition, measurement, guide and detection, etc., can we have a traditional algorithms of image and deep learning two technical direction, the choice of these two technologies, respectively, represents the experience summary of computer image technology and the calculation method of the era of big data. Although the application of deep learning is increasing, the role of traditional image algorithms cannot be ignored.

1. Traditional image algorithm

Image storage in the computer is usually divided into image size and color information.

Image size: indicates the number of horizontal and vertical pixels of the image. For example, if the image has 2500 horizontal and 2000 vertical pixels, the pixel accuracy of 5 million is constituted. Then at least 5 million values need to be stored in the computer to represent all pixels

Image color: Based on whether the image is a color image, the color image will contain the three primary colors of red, green and blue, but there is only one gray value for the black and white image. Therefore, each pixel of a color map needs to store three values to represent the component of each color, while a black and white image only needs to store one value to represent the size of the gray scale.

The traditional image algorithm is a kind of efficient computational analysis based on the above size and color information.

Taking the defect detection scenario as an example, when we need to test product defects, defects of a certain size and depth will usually be detected according to the requirements of process quality inspection. The processing method of the traditional image algorithm is to traverse the photographed defect image with a pixel value, and judge whether there is an uneven area on the surface by comparing the color difference between each pixel and the surrounding pixels. If the color difference is too large, it indicates that there may be an uneven area at this location, that is, a defect. At the same time, the number of pixels with large difference can be counted and the location of pixels can be recorded, so as to obtain the area size and location of defects. This is the basic core idea of the traditional image algorithm in the detection application.

So when we in the use of traditional algorithms of image processing visual applications, through the characteristics of the project location for size and color combinations to design different characteristics, such as: distance, color, shape, aspect ratio, color gradient pixel statistics, etc., through different decision rules, can deal with a variety of test object and test requirements.

The traditional image algorithm has simple processing methods and clear decision rules, but it also has some shortcomings:

· Feature design depends on the experience of engineers and the requirements of detection content, which makes it difficult to realize the overall algorithm reuse.

· Feature design requires too many parameters for complex objects, and the logic complexity of rules exceeds the ability of manual design.

· The algorithm is sensitive to the difference of color gray scale and has high requirements for optical stability of the imaging environment

· The algorithm has no correlation with the spatial position, and it needs to fix the area to reduce the computation amount of pixel traversal in the whole image and reduce the interference of background noise.

2. Deep learning

If we look at a picture of a cat and a dog and want to distinguish them correctly, the human brain quickly analyzes the differences in features between them: Eyes, nose, ears, coat color, tail, etc., our thinking and experience will classify the various features seen by the eyes. When all the features are consistent with the impression that you have seen a cat, then we can think with high probability that this is a cat.

Deep learning is just like the human brain. When seeing an object with complex features, it can parse out each feature layer by layer and carry out corresponding quantification. Of course, this is achieved by the multi-layer neural network constructed by continuous trial and error.

When all the features are decomposed and quantified from a picture, a regression classification of the feature distribution can be carried out through the corresponding mathematical statistics. After several regression calculations, the boundary line that can better distinguish these features can be fitted.

After that, when similar objects are input into the analysis, according to the distribution of the dividing line, a proportion of all features distributed in different areas can be calculated, so as to calculate the probability of the object being classified into a certain category.

Therefore, when deep learning technology is used for visual application, feature analysis, quantification and probability calculation of feature distribution of the detected target can be carried out through a set of network structure, and then the target to be detected can be judged based on probability statistics, so as to achieve reuse at the algorithm level. At the same time, due to the extraction of a variety of features, the background has a relatively strong anti-interference.

However, deep learning also has obvious disadvantages:

· The empirically constructed feature extraction network structure and probabilistic calculation method are uninterpretable to a certain extent for visual recognition, and it is not easy to determine the adjustment direction. For example, there are obvious differences in the identification of features, similar features can be identified.

· There needs to be predictable labeling and distribution of sample features, which is commonly referred to as the need to label data and learn. Data are often limited by the difficulty of obtaining, so it is easy to make it difficult for deep learning technology to form an evaluation standard.

· It requires a large amount of calculation and a high requirement on the computing power of the equipment, so it is not suitable for high-speed detection.

3. How to choose industrial vision applications

Traditional image algorithm is suitable for the scene with single target and stable features. In industrial applications, it can avoid the background environment interference and reduce the complexity of rule design by means of tooling, which can meet the requirements of high-speed detection. This is also the technical means adopted by most visual manufacturers.

Based on the application of traditional image algorithms, deep learning can better cope with the diversity of targets, and the anti-interference to the background also makes the use of deep learning technology does not need to consider too much tooling design, can effectively reduce the complexity of structure design, making the whole vision application scheme better adapt to flexible production.

Traditional image algorithms will not be eliminated due to the development of deep learning, and deep learning also needs the assistance of traditional image algorithms to add regular conclusions to its probabilistic results.

Traditional image algorithms and deep learning complement each other and learn from each other to build high-performance, highly usable and flexible industrial vision solutions.