Quality Control of Leather

High-quality products must not only satisfy many material-relevant material requirements but also have excellent optical properties. These requirements apply to leather in particular, since leather is mostly used for visually appealing products (e.g. upholstery, car seats, steering wheels). Because leather is a natural product, possible quality-degrading properties, so-called "surface defects" are manifold and hardly avoidable.

For this reason, a quality control is carried out several times during the production of leather along the value chain. This surface test is challenging and usually takes place as a manual 100 percent visual inspection. We have developed software modules which, in conjunction with suitable hardware, can support or fully automate this quality control.

Modules have been developed for

  • Contour detection
  • markers identification
  • as well as for the detection and classification of defects.


As a rule, the necessary assortment of the skin material takes place in the so-called wet-blue stage.
As a rule, the necessary assortment of the skin material takes place in the so-called wet-blue stage.

Contour Detection Process for Leather Hides

A cow hide has an approximate size of 3 m x 3 m, but this varies considerably with respect to its geometry. For the processing process, it is important to know which leather area is available at all.

Contour detection is associated with some difficulties because of the strong natural variability of leather. Like most natural products, cow hides are unique, which is why assumptions about the shape of the hides are not possible.

  • The color of foreground and background varies greatly.
  • The hides throw wrinkles.
  • They show cracks, holes and discoloration..

Standard procedures work only to a limited extent. Therefore, a special contour detection process for leather hides has been developed. In this case, the images of the color camera are first transformed into a special color space, which has a contrast-rich two-dimensional subspace ("gray value image"). A morphological contour reconstruction takes place under the assumption that the center of the image is within the hide. This increases the gray values, which are similar to the center point.

In the next step, an adaptive threshold value can be calculated by means of which the foreground is separated from the background. In order to achieve a continuity of the contour and to eliminate outliers, further morphological operations are carried out, for example the watershed transformation for the detection of connected areas. This results in a stable segmentation of the leather pieces with a high degree of regularity, which is largely independent of the actual color of the hide and the background.

Originalbild Lederhaut
Originalbild Lederhaut
Erfasste Kontur der Lederhaut
Erfasste Kontur der Lederhaut

Detection of Manually Applied Markers

Automatic defect detection systems are rarely used. Alternatively, the manual application of markers at defect spots is widespread. These markings can be digitized by comparatively simple camera systems and fed to the subsequent cutting optimization step. Nevertheless, the automatic recognition of the markers is associated with some difficulties, for reasons similar to the contours detection.

In order to determine with which color in the image the markings have been applied and which color correspond to the unlabeled leather hide, the color gradient of the image is first determined by means of an eigenvalue decomposition. As a rule, it is thereby possible to find a global threshold for the separation of marking and background (leather). However, due to the inhomogeneous leather structure, a series of pseudo-markings can arise. That is, areas are recognized as marks, but are in reality leather structure. This over-segmentation is eliminated by a subsequent hysteresis process. A morphological skeleton provides the vectorization of the markers, which are finally converted into polygons for further processing in the process.

Lederhaut mit markierten Fehlern
Lederhaut mit verschiedenen händisch markierten Fehlerregionen und Resultat der Erkennung.

Automatic Defect Detection

Due to the steadily increasing quality requirements and the growing performance of the image processing systems, automated testing will be even more successful in the future.

The design of such a test system varies according to product type and requirements. Typically, two to four line cameras with special lenses and as many illuminations are used. As a rule, a camera operates in the incident light and the other cameras in the dark field, whereby direction / angle and color of the illumination vary. Because of the large amount of data (depending on configuration, some gigabytes per hide), a powerful evaluation unit is required. These requirements are currently best met by PCs with multi-core processors.

For textured surfaces (as in the case of leather, for example), robust evaluation algorithms are of decisive importance. Roughly, the evaluation can be divided into the following four steps:

  1. Preprocessing
  2. Defect detection
  3. Clustering
  4. Defect Classification

The main difficulty with defect detection on leather surfaces lies in the irregular natural structure of the leather. In order to distinguish between defects (e.g. scratches) and natural irregularities (e.g. veins), elaborate pre-smoothing steps are required. This allows for easier detection of these defects in subsequent steps and increases the robustness against disturbances in the data (e.g. noise). On the other hand, shock filters are used which not only preserve, but also strengthen surface defects in the image.

An essential step for detecting defects is finding edges. These can be located in the image data using an edge detector. Since simple gradient-based edge detectors do not produce the desired result, more complex edge detection procedures have been implemented. The basis is a scaled view using wavelets. The edges are detected over several scales. Structures in the image are not only localized, but also receive information regarding their size and shape. In addition, topological gradients, specially adapted erosion and dilatation as well as geodetic reconstruction are used.

Ergebnis des Clusterings einer Menge von potenziellen Defekten.

The described methods are so-called detection methods, i.e. algorithms that detect potential defects on the hide. These algorithms also generate pseudo-defects due to the surface structure. In order to eliminate these, the potential defects have to be classified in a further step. Since a direct classification is too faulty, another method was introduced, the so-called "clustering". Clustering algorithms use descriptors to calculate a similarity hierarchy of all potential defects found. This process is fully automatic and does not need any human interaction. The agglomerative hierarchical clustering compares all found defects in pairs and groups them hierarchically.


Typische Deskriptoren für den Klassifikator
Typische Deskriptoren für den Klassifikator. Derzeit werden zwischen 50 und 100 Deskriptoren benutzt.

The result of the clustering is now used for the classification by including the generated hierarchy in the list of descriptors. Descriptors describe the defect characteristics determined from the image and are decisive for the quality of the classification.

Typically, a training record is used for the classification, in which the defects were marked manually. The creation of the training data record is a complex process, since the data record must be sufficiently large on the one hand and, on the other hand, be as consistent as possible. Consistent means that only real defects are marked, but also no defects are been disregarded.

There exists a wide range of classifiers, typically support vector machines are used. The classifiers trained on the different classes of defects are executed several times and not only distinguish between defects and pseudo defects, but also supply the defect type, such as scratches, pigmentation, warts, dirt.