EP0661108A2 - Method for optically sorting bulk material - Google Patents

Abstract

The invention relates to a method for the optical sorting of bulk material in a color sorting machine, in that it is conveyed via a conveyor belt and moves past an observation head with a light source and a product signal receiver arranged in the vicinity of the light source, the reflected light of the pixels of the test material being passed through different Color filters of adjacent detection elements of a line of the receiver are broken down into several spectral ranges and the test material is sorted on the basis of the color values (measured value of the intensity in the respective color). According to the invention, in order to improve the detection rate, it is provided that, in the case of test material mixed with rejects, the color values of the product are examined in several selected sub-areas, in that a classifier in each sub-area determines connected areas of pixels with color values falling in the respective sub-area and according to predetermined criteria from the Geometry and the size of these detection areas carries out a classification.

Description

The invention relates to a method for the optical sorting of bulk material according to the preamble of patent claim 1.

It is already known that test material is conveyed on tapes and its image is recorded for testing with a diode line camera or a television camera. The signal is preferably recorded in flight, e.g. the test material is transferred from one belt to another belt. When recording signals in flight, the test material can be examined from several sides with a defined background.

In modern systems, the color is also recorded during image acquisition. The color is used to detect conspicuous areas in the image.

The image of the test item is evaluated in step-by-step with the image scanning, so that a test item can be classified immediately after it has passed through the measuring station. This makes it possible to eject the parts in flight using flaps or air nozzles.

A disadvantage of the known methods is that the detection rate for color-heterogeneous products is low if the detection of conspicuous pixels is limited to the detection of color values that are not contained in the product because there are many different color values in the product. If the detection is expanded to include color values which are also contained in the product, an unbearably high proportion of the error-free product is generally detected as a reject even when extending to colors which rarely occur in the product.

It is an object of the invention to improve the method for the optical sorting of bulk material in such a way that foreign bodies to be detected with color-heterogeneous bulk material are recognized with a very low error rate.

The method with the characterizing features of patent claim 1 serves to achieve this object.

During the image acquisition, the light of each pixel is blocked by color filters in front of the detection elements of a line e.g. broken down into the three color components red (R), green (G) and blue (B). It is thereby achieved that a detection of conspicuous pixels (points with color values which rarely occur in the fault-free product) is possible by evaluating the color values (intensities of the respective color components) measured by the line elements. The geometry is then evaluated in terms of local clusters of conspicuous pixels.

First, the entire range of possible color values in the color space is divided into several sub-areas, the color space being spanned by the different color components that are measured for each pixel. For example, the three color components red, green and blue form a three-dimensional color space. Allow classifiers, ie means for evaluating the measured values on the basis of predetermined criteria a classification of the measured color values, with a classifier concentrating on only one sub-area and thereby recognizing detection areas in this sub-area for the color-heterogeneous product, that is, coherent areas of conspicuous pixels.

If the color values of a color-homogeneous reject part are preferably in the selected sub-area, the reject part is detected as a relatively large area of pixels of the color values of the selected sub-area and can be recognized by the classifier by evaluating this detection area. On the other hand, in the case of the defect-free product within such a sub-area, large areas of conspicuous pixels are generally found only in rare cases, and the number of incorrect detections thus remains small. This improvement in the classification is used in practical application by dividing the committee into typical classes and setting up a classifier for each class, with the classifiers working in parallel during the test.

In a preferred embodiment, the distribution of their color values is learned in the sub-areas in which reject parts are suspected by showing reject parts.

Fig. 1

zeigt beispielhaft eine eindimensionale Farbwertverteilung mit den Bereichen für gutes Prüfgut.

Fig. 2

zeigt ein eindimensionales Beispiel zur Klassifizierung mit parallelen Klassifikatoren bei der Erkennung von Ausschuß, dessen Farbwerte sich mit den Farbwerten des Produktes überlappen.

Fig. 3

zeigt ein eindimensionales Beispiel zur Korrektur eines Klassifikators durch das Nachlernen.

Fig. 4

zeigt ein Beispiel für die Farbverfälschung an Objektkanten bei Verwendung einer Kamera mit Bildpunkten, bei denen die Farbsensoren nebeneinander liegen.

The invention is explained in more detail below with reference to drawings:

Fig. 1

shows an example of a one-dimensional color value distribution with the areas for good test material.

Fig. 2

shows a one-dimensional example of classification with parallel classifiers in the detection of rejects, whose color values overlap with the color values of the product.

Fig. 3

shows a one-dimensional example for the correction of a classifier by re-learning.

Fig. 4

shows an example of the color falsification at object edges when using a camera with pixels, in which the color sensors are next to each other.

In a color sorting machine, the bulk material preferably moves in flight past an observation head with a light source and a product signal receiver arranged in the vicinity of the light source. The reflected light from each pixel of the test material is transmitted through different color filters of adjacent line elements of a camera line, e.g. a CCD line, the receiver divided into the three colors red (R), green (G) and blue (B). The line elements thus measure the brightness of the pixels, also called color values, in their respective spectral ranges. This results in a three-dimensional distribution of color values, the evaluation of which is discussed below using one-dimensional examples.

With reference to FIG. 1, the test material is measured without rejects in a pre-learning process and the frequency distribution 1 of the color values is determined.

In a re-learning process, the test material is also measured without rejects and in a first step a color value range for good test material is determined by placing a threshold 2 based on experience on the frequency distribution 1 of the color values, the intersection between the threshold 2 and the curve of the frequency distribution being derived from the intersection points 1 result in the limits of the test material color value range.

With the selected setting of threshold 2, pixels will also appear in the case of the error-free test material, which are classified as conspicuous. However, these pixels would erroneously serve as a committee if they clustered into large areas classified. Experience has shown that such an agglomeration in turn occurs preferentially in certain color value ranges. In order to measure these color value ranges, a large area detected in the error-free test material is stored in the learning process and the distribution of its color values is measured. This distribution is introduced as threshold 3 after standardization. All color values at which the threshold 3 exceeds the color value distribution 1 of the test specimen parts, ie the color values in the interval between the intersections of the threshold 3 with the curve of the color value distribution 1, are interpreted as belonging to the test specimen and therefore do not lead to an error detection.

When measuring test material with rejects, the color value ranges of the product are divided into sub-ranges. With reference to FIG. 2, in this example each of the classifiers A, B and C working in parallel concentrates only on one sub-area. If the color components of the color-homogeneous reject part are preferably in the selected sub-area, the reject part is detected as a relatively large area and can be recognized by evaluating the detection areas. Here, too, the distributions of the color values of these large areas are measured and introduced as thresholds after their normalization. All color values at which these thresholds 4, 5 and 6 exceed the color value distribution 1 of the test material parts are interpreted as rejects and lead to an error detection.

It is also possible that in the case of a defect-free product, large-area detection areas are detected in a color value range covered by a classifier, and the defect-free product is thus classified as a reject. In a further learning process, these color values in particular, which lead to large-area detection areas in the fault-free product area, are learned and recognized as good test material by changing the thresholds. 3, the threshold 8 shows the Color value distribution of a reject part. Within the color value range determined by threshold 8, error-free test material is classified as a reject. Through the re-learning process, the color value distribution of this large-area detection area is measured in the error-free test material and introduced as a threshold 7 after standardization. All color values at which the threshold 7 exceeds the threshold 8 of the reject part are interpreted as belonging to the test material and thus do not lead to an error detection.

After learning, the product is automatically checked.

During the test, which can take days, systematic drift-like changes in the product can be expected. These changes result in system performance degrading over time. To avoid this, the classification system is doubled. One system takes over the test task, while the other system measures the current color value distribution of the product. The measurement of the current color value distribution is monitored by the checking classifier so that no color values of the rejects are recorded during this measurement. After a representative number of measured values has been recorded, the learning classifier with the newly measured distribution is activated for the test task, while the classifier which has been set to test so far takes over the learning task.

This adjustment is only possible if a detected color point classified as conspicuous does not always lead to a committee decision. If a detected color point would always lead to a reject decision, the learning classifier could not adopt new color values, since the newly learned color value distribution is rejected when a reject decision is made. Since color points detected in the system are only classified as rejects if they form a larger coherent area, the measured frequency can can also be adjusted for detected color values. Conversely, with this adjustment, the system can detect color values belonging to the committee that were present in the color value distribution of the product in an earlier measurement and are no longer included in the currently measured distribution.

When the signal is recorded, the test object is illuminated, for example, by two lamps from the direction of the line scan camera. The optical axis of the line scan camera lies between the two lamps. With this arrangement, the design of the background is of great importance because the background should not expand the color value distribution of the error-free product if possible. An extension would lower the detection performance.

This requirement cannot be met if the test material is taken up lying on the conveyor belt. The tape is not of a uniform color due to dirt and wear. In addition, shadows form on the conveyor belt, which leads to a significant expansion of the color value distribution when measuring the error-free test material. For this reason, the test material is observed in flight.

In a first embodiment variant, the background has the color of the test material, which has the advantage that the contrast between the background and the test material is low and therefore the color value distribution of the test material is not significantly expanded by edge effects at the transition from the background to the test material. This variant provides the best results in terms of color and spatial resolution.

The disadvantage of contamination is avoided by designing the background as a rotating roller, which immediately throws away deposits. The shadow of the test material on the background becomes diffuse and, depending on the bulk density, harmless if the rotating roller is installed at a suitable distance from the test material. When the bulk density of the test material is high an excessive darkening of the background is avoided by additional lighting of the background. Alternatively, the background can be a cylindrical emitter that emits the color of the test material and is surrounded by a transparent rotating roller that throws away the deposits.

In a second embodiment, the background is a dark hole, which has the advantage that the test material can be segmented from the background and there is no impairment due to dirt and shadows. In the case of segmentation of the test material, for example, the form for separating good parts and rejects can be used.

To create the dark hole, the largest possible container with low-reflection walls is built. The line scan camera looks into this container through a slit. The width of the slot is adapted to the aperture and focal length of the camera lens and to the distance to the focus plane.

During the image acquisition, the light of each pixel is broken down into the three colors red (R), green (G) and blue (B). Depending on the selected scanning principle and the adjustment of the camera, the color components are not ideally measured at the same location, but rather at different locations. In conventional color cameras, the color sensors are even located side by side, so that the color sensors see different spatial areas of the measurement object with respect to a pixel. With reference to FIG. 4, the color sensors (R, G, B) are arranged horizontally, while the measurement object moves past this horizontal line from top to bottom. In this example, the background generates the signal levels R = 0, G = 0 and B = 0 for the respective color sensors, while the measurement object produces the signal levels R = 100, G = 50 and B = 20. In Fig. 4, only the sensor triple Xn, Yn measures the correct color of the measurement object. For all other triples, color values are measured that contain at least one color value that is darker than the corresponding color value of the test material. So measure to Example the triple Xn, Yn-1 the levels R = 50, G = 25 and B = 10. To avoid these disturbances, pixels whose color values are darker by an adjustable factor than those of the corresponding neighboring point are suppressed by the signal levels are saved and a comparison of the horizontal and vertical neighboring points is carried out.

Claims (10)

Method for the optical sorting of bulk material with rejects, such as agricultural products, pharmaceuticals, ores, etc. in a color sorting machine by conveying it over a conveyor belt and moving past an observation head with a light source and a product signal receiver arranged in the vicinity of the light source, whereby the reflected light of the pixels of the test material is broken down into several color components by various color filters of adjacent detection elements of a line of the receiver and the test material is sorted on the basis of the color values corresponding to the measured intensities of the respective color components, characterized in that the color values of the product in several selected sub-areas of the Color space are examined in that a classifier in each sub-area determines contiguous areas of pixels with color values falling in the respective sub-area and according to predetermined criteria a classification based on the geometry and size of these detection areas. A method according to claim 1, characterized in - That the test material is measured without rejects in a pre-learning process and its color value frequency distribution is determined depending on the color; - That in a re-learning process in a first step when using test material without reject parts, a color value range for good test material is determined by in each case a threshold based on experience being placed over the frequency distribution of the color values, the intersection between the threshold and the curve of the frequency distribution showing Result in limits of the test material color value range; - that in the re-learning process when using test material without rejects, color-dependent foreign matter-related measured values, which would be erroneously classified as rejects according to the limits of the test goods color value range of the first step of the re-learning process, are determined and the size of the local concentration of these measured values is determined; and - That in the re-learning process when using test material without reject parts when a predetermined size of this local concentration of foreign matter-suspected measured values is exceeded, the threshold value decision of the first step of the re-learning process is changed in a color-dependent manner so that good test material is decided for these measured values. Method according to one of the preceding claims, characterized in that when testing the test material mixed with rejects, the classifiers working in parallel only analyze sub-areas of the color space in which rejects are suspected. Method according to Claim 3, characterized in that in the subregions of the color space in which reject parts are suspected, the color value distribution is learned by showing reject parts. Method according to one of the preceding claims, characterized in that the test material to which the rejects are added is checked using a first classification system, while the current frequency distribution of the color values of the test material is measured using a second classification system to adapt to systematic drift-like changes in the test material, the measurement is monitored by the testing first classifier so that no reject values are recorded during the measurement. Method according to claim 5, characterized in that the two classification systems alternate in their function. Method according to one of the preceding claims, characterized in that large gradients and these generally disturbed color values are not taken into account in the measurement by comparing color signals from adjacent pixels. Method according to one of the preceding claims, characterized in that the sorting of the goods takes place in flight when the goods translate, for example, from one belt to another belt. Method according to one of the preceding claims, characterized in that the measurement background is a dark hole. Method according to one of claims 1 to 8, characterized in that the measurement background is designed as a cylindrical emitter with a rotating transparent roller surrounding the emitter, the emitter emitting light in a color matched to the test material.

Images