Texture and Surface Analysis

Texture analysis is the process of analyzing surfaces such as fabrics, leather, rubber, wood, etc. to find errors or defects. Although it is commonly used in the textile industry, it also has applications in many different sectors such as metal industry, plastics, food industry, and packaging.

Texture analysis, also known as novelty detection and texture inspection in the literature, is one of the most challenging areas of image processing. Identifying errors that can be easily distinguished with the naked eye can be difficult to communicate to a computer through mathematical modeling. Since errors may not always have the same form, a general approach may not always work.

While it may be easy to find errors on a surface like fabric, wood, leather or rubber in some cases, a program that can capture almost any type of error needs to use advanced artificial vision techniques.

In places where uniform production is carried out, finding certain types of errors may be easy. However, companies that produce with dynamic production models generally work with different types of models and patterns. Both the surface to be controlled (fabric, wood, leather, etc.) and the error to be captured (oil, stain, color difference, tear, cut, etc.) vary greatly, making it almost impossible to develop a general algorithm.

For example, in the above picture, faulty areas can be easily seen with the naked eye. With a classic approach, finding the black areas and checking the area would be a simple approach that would yield good results. However, when production is carried out with different colors and patterns, area and color checks will have to be changed every time, and in some cases, these checks can result in labeling a good product as faulty or a faulty product as good.

For such situations, at Pi Robotics, we use a machine learning solution based on a classification technique. In this method, 4-5 different filters are applied to the image. (Usually, first and second derivatives are taken.) The softened and Gaussian filtered images obtained are subject to a classification process. Afterwards, the system will automatically detect each type of error. (Without the need for any parameter such as color, size, area, etc.)

Generally, classification methodologies known as GMM (Gaussian Mixture Model), MLP (Multi Layer Perception), and SVM (Support Vector Machine) are applied. While approaches supported by CNN (Convolutional Neural Networks) have also been tried in recent years with the development of deep learning technologies, they are less preferred because they require a lot of sampling.

As can be seen, in tissue analysis problems that seem easy at first but can become quite complicated when delved into, we use methods that are as simple as possible from the end user's perspective but technologically advanced.

Depending on the type of fabric and the desired error to be captured, images can be illuminated from the front or back, and line scan or area scan cameras can be used.

The systems we build by analyzing tissues are generally quality control systems that capture the following errors:

Weft-Warp Errors

Stains such as oil, dirt, soot, etc.

Pattern Distortions

Color Differences

Errors such as tearing, melting

Defects such as insect holes, surface wounds, scratches

Errors such as embossing, sinking.

Classification methodologies commonly known as GMM (Gaussian Mixture Model), MLP (Multi Layer Perception), and SVM (Support Vector Machine) are applied. Although CNN (Convolutional Neural Networks) supported approaches have been tried lately with the development of deep learning technologies, they are less preferred because they require a lot of sampling.

As seen, in tissue analysis problems that may seem easy at first but can become quite complex when you dive in, we use methods that are as simple as possible for end-users but technologically advanced.

Images can be front or backlit, line scan or area scan cameras can be used depending on the type of fabric and the type of error to be captured.

The systems we build by analyzing tissues are generally quality control systems that capture the following errors:

We determine the sensitivity for these errors according to customer needs. The camera, lens, and lighting are selected at a resolution suitable for customer sensitivity. Multi-camera systems are often set up. Production speed is an important parameter to consider.

In the end, a solution is presented based on the error types to be captured as a result of meetings with the customer (especially the quality control department). The solution we offer is an artificial vision solution that can self-learn, adapt to many different patterns, and uses deep learning techniques.

Surface analysis is similar to tissue analysis methods. Generally, the surface is not as complex as the tissue. It is often used in the white goods, glass, automotive, and metal sectors.

The basic usage areas are to find errors such as surface scratches, cracks, fractures, migration, or bulging.

Surface analysis is used for the detection of errors on the external surfaces of white goods such as refrigerators or washing machines, paint control on surfaces such as automotive body bumpers, and controls on surfaces such as metal press sheets with broken or torn holes.

Two-dimensional analysis methods are used when the surface is planar (as in the refrigerator example). When the surface is three-dimensional (usually metal press surfaces are like this), surface analysis becomes difficult. In this case, the best solution is sought with alternatives such as using a 3D camera, setting up a multi-camera system, or moving the camera on a robot gripper.

Surface with scratches on a metal part (left) and the captured image of detected errors (right)

In surface analysis, filtering techniques are mostly used in image processing.

Transformations such as Fourier transform, inverse Fourier transform, convolution, and Gaussian filters are commonly used structures.

Lighting will play an even more important role in finding surface defects. Especially with photometric stereo and deflectometry methods, many errors that are difficult to see even with the naked eye can be detected.