WO2017216124A1

WO2017216124A1 - Sorting device and corresponding sorting method

Info

Publication number: WO2017216124A1
Application number: PCT/EP2017/064329
Authority: WO
Inventors: Thomas Längle; Wolfgang Karl; Georg Maier; Michael Bromberger; Mario Kicherer; Thomas Becker
Original assignee: Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V.; Karlsruher Institut Für Technologie (Kit)
Priority date: 2016-06-14
Filing date: 2017-06-13
Publication date: 2017-12-21
Also published as: EP3468727A1; EP3468727B1; DE102016210482A1

Abstract

The invention relates to a sorting device with which objects (0) of a material flow (M) are sorted, comprising an image-capturing unit (1) with which the material flow (M) is optically captured and image data (4) of same (M) is generated, an evaluation unit (2) with which objects (0) in the material flow (M) are identified and classified, and a sorting unit (3) with which classified objects (0) of the material flow (M) are sorted, characterised in that one or more configuration parameters (5) characterising the material flow (M) in terms of the configuration of the objects (0) therein is determined with the evaluation unit (2) from the generated image data (4), and in that one or more evaluation parameters (6, 8) controlling the identification and classification of the objects (0) via the evaluation unit (2) is/are used on the basis of the determined configuration parameter/s (5).

Description

Sorting device and corresponding sorting method

The present invention relates to a sorting apparatus (hereinafter also referred to as an optical sorting system) according to the preamble of claim 1 (and a corresponding method).

In the optical sorting of bulk solids in material streams, which can be well separated by optical features in the visible spectrum such as color, shape and / or texture, or even near infrared properties, plant operators are interested in maximizing material throughput. This results in a better economy. A higher material throughput can be achieved by higher bulk densities, larger sorting widths and / or higher material speeds.

For a most accurate discharge and thus the best possible sorting result, it is necessary to arrange the sensors and the separation technology spatially as close to each other. This allows the location of the Objects in the material flow at the level of the separation mechanism predicted better and it can be correspondingly precise, for example by pneumatic valves, discharged. In addition, bycatch (that is, the unwanted rejection of objects) near an object to be rejected can be kept as small as possible.

For many known features, e.g. geometric descriptors, the calculation time depends directly on the object. In the case of very high bulk densities or occupation densities in the material flow, object agglomerates (object clusters) often additionally arise from the point of view of image processing, which algorithmically can only be separated with difficulty (in particular: segmented). As a result, even when sorting homogeneous products or objects, greatly different computational costs can arise for the classification.

The above-mentioned aspects present evaluation procedures with great challenges, since they are always accompanied by the demand for real-time evaluations. If real-time barriers are violated for certain objects contained in the material flow, there is no timely decision by the evaluation system about those objects, i. a sorting decision for a corresponding object (in particular: should this object be rejected or not) can not be taken in time. This potentially (or indeed) leads to wrong sorting.

Sorting systems, in particular for sorting bulk goods, are known from the prior art from the following publications:

• DE 10 2009 007 481 AI,

• DE 10 2010 046 438 AI,

• DE 10 2011 103 253 AI as well

• DE 10 2012 001 868 AI.

See also "Prior Art of Sensor Assisted Sorting" by H. Wotruba, BHM, 153rd ed. (2008), No. 6, pages 221-224.

The object of the present invention is to improve the performance of optical sorting systems for sorting objects in material flows, in particular to improve the real-time capability of the systems, ie, the Increase probability that there is actually a sorting decision for each object in the material stream at the time of discharge. In addition, the present invention is intended to provide a more efficient processing of sensor data in order to enable a real-time capability even at short distances between the image recording unit (sensor system) and the sorting unit (unit for discharging objects) (and consequently to increase the sorting performance). In particular, the present invention is intended to provide new approaches to sensor data processing.

This object is achieved by an optical sorting system according to claim 1 and a sorting method according to claim 14. Advantageously realizable variants can be found in the dependent claims.

The present invention is based on the following basic considerations: In conventional systems, there is the problem that for some objects when reaching the separation mechanism (sorting unit) there is no sorting decision, since the required calculations of the evaluation system (evaluation unit) have not yet been completed. A subset of the prior art known plants does not handle this condition at all, resulting in forcible loss of sorting performance and / or sorting quality. To solve these problems, the following approaches seem to be available:

• Reduction of material supply: If there are few objects for evaluation, more time per object can be used for calculations. However, this would have a negative impact on the profitability of the plant.

• Increasing the distance between the sensory acquisition and the separation: This allows the evaluation unit (at constant material speed) to have greater latency for image evaluation. However, this would entail the disadvantage that the position of objects on reaching the sorting unit (separation mechanism) can be forecast worse. This could be counteracted by larger blow-out windows, but this would increase bycatch (unwanted rejection of objects close to an object to be knocked out). • Implementation of the image evaluation in hardware: Although this would achieve very high processing speeds, this would have the disadvantage that the corresponding realization would have to be tailored to a specific sorting task. There would then be a lack of flexibility with regard to the underlying sorting task (in many systems, however, different sorting tasks have to be solved).

All of these solutions have in common that they either support high processing speeds or increase the tolerable latency. From these fundamental considerations, the present invention therefore concludes that an adaptive behavior of the sorting system is necessary to ensure or improve the real-time capability of the system. In other words, a basic idea of the present invention is to suitably respond to the required calculation times which fluctuate depending on the sensor data or the specific conditions in the material flow and / or on its objects.

The basic realization of this basic idea describes the optical sorting system according to claim 1.

The system according to the invention can be realized as a belt sorting system, but slide sorting systems are also conceivable. The optical detection of the material flow in this case usually means the recording of a plurality of individual images of the material flow or of cutouts thereof ever

Time unit. For example, while the sorting system is working, video images (fast frame sequences) of a neutral background over which the material stream is transported can be recorded and evaluated by the evaluation unit in real time (with regard to the objects recorded in the material flow in front of said background).

In this case, a color line or color area camera system (array), for example on a CCD or CMOS basis, can be used as imaging sensor system (image acquisition unit). The individual images or video images of the image recording unit can be evaluated by image processing algorithms by the evaluation unit to the six-dimensional pose (ie the three-dimensional position and the three-dimensional orientation) or at least to determine the three-dimensional position of each individual object of the material flow at defined times.

The setting of the evaluation parameter (s) is preferably carried out by or in the evaluation unit, in particular by a microcontroller or a computer unit thereof.

The sorting of "classified" objects does not preclude non-classified objects from being sorted (for example, for safety reasons, they are always rejected as "bad objects" or rejected). The sorting may, in particular, be a mechanical separation of the classified objects. In the simplest case, the objects are classified into two classes, namely "good objects" and "bad objects". The objects of the two classes can be collected during sorting in separate containers. For this purpose (as is generally known to the person skilled in the art, cf. also the aforementioned prior art), rapid air valves can be used to blow out the bad objects.

First advantageously realizable features can be found in claim 2.

In other words, a sorting decision can be made with the said minimum probability for any object in the material flow by the evaluation unit, so that the sorting unit can react to this object in accordance with the sorting decision made. If the (minimum) probability equals 100%, the evaluation unit makes a decision of identifying and classifying with absolute certainty for all objects in the material flow, ie a sorting decision. In this case, (especially at very high belt speeds of the conveyor belt of a belt sorting system and / or in the case of very high occupation densities of the objects in the material flow), the sorting decision may also be incorrect. It may also happen that a certain, classified object is still not rejected due to an erroneous action of the sorting unit on the flow of material, despite a (correct) sorting decision of the evaluation unit for "discharging." It may also happen that a true, classified object is discarded, even though the (correct) sort decision for this object is "not purged".

Further advantageously realizable features can be found in claim 3.

Thus, the accuracy GK of one or more process steps / e (for example, the segmentation process step or the process step of applying a decision tree-based classifier) can be set. The process step (s) is / are then carried out with the (in each case) set GK. The setting of a low GK can mean, for example, that the image data or the data resulting from the image data are only processed as coarse-screened as input data of the process step in order to minimize the number of calculations to be carried out with computer assistance. In this case, the process step performed with this accuracy GK is ended in any case.

However, in the case of iterative calculations or recursive calculations, this can also mean that after a defined (for example, low) number of repetitions of the iteration loop (for iterative calculations) or after reaching a defined (eg low) recursion depth (for recursive calculations), ie after fulfilling a defined termination criterion, a termination of the calculations of the process step takes place. As a consequence of such presetting or setting of different accuracies, varying calculation times and / or repetition numbers (in the case of iterative calculations) or recursion depths (in the case of recursive calculations) result. According to the invention, in particular, the accuracy of stepwise refining algorithms (in particular: recursively, iteratively and / or incrementally refining algorithms) can be used. This need not be the case, the algorithms can not be repetitive and / or analytical.

Further advantageously realizable features can be found in claim 4. Thus, the calculation time BZ of one or more process steps / e can be at least approximately predetermined or set. The process step (s) can then be carried out with the (respectively) set BZ or until the BZ has expired (termination criterion).

As a result of predetermining or setting different calculation times, varying accuracies and / or repetition numbers (for iterative calculations) or recursion depths (for recursive calculations) result.

Further advantageously realizable features can be found in claim 5. It is therefore possible to specify or set frequencies with which sequences of computation of process steps are repeated (with recursive algorithms, the frequency of the recursion depth corresponds to). Setting a lower frequency usually leads to a lower accuracy with which a process step is carried out or to a lower calculation time for the process step. After the calculation sequences have been repeated frequently (or the chosen recursion depth has been reached), the process step is ended. The termination criterion is therefore a frequency here. As a result of such predetermining or setting of different repetition frequencies (hereinafter also abbreviated to WH), varying accuracies GK and / or calculation times BZ result for the corresponding process steps.

According to the claims 3 to 5 according to the invention step-by-step accuracy-refining process steps can be implemented or realized. Any two or all three of the mentioned evaluation parameter types (GK, BZ and / or WH) can be used together by the evaluation unit for identifying and classifying the objects of the material flow. For example, in one process step (eg the segmentation in the image data) the GK can be set, while in another process step (for example, the application of a classifier based on decision trees) the GZ is set. Advantageously, the implementation of the algorithms used is carried out such that for all used evaluation parameters or types of evaluation parameters for each process step at any time during the execution of such a process step (or at the time of

Abortion of the calculations of such a process step) the conditions according to claim 2 are met.

In other words, preferably all calculations for all objects in the material flow are performed and completed in that time

(If necessary, for example, by premature termination of the calculations at still very low accuracy), which the objects each need to get to their detection by the image acquisition unit (for example on the conveyor belt) to the last possible place where they are still sorted out by the sorting unit can. According to the invention, it can thus be ensured that for each object or at least for 90%, 95% or 99% of the objects there is always a momentary (possibly still very rough) classification decision (which can then be followed by a corresponding sorting decision).

Further advantageously realizable features can be found in claim 6.

The identification and classification or the evaluation can therefore be carried out in several process steps (hereinafter also referred to as components) of the evaluation unit. For each process step (each component), one type, preferably exactly one type, of evaluation parameters according to claims 3 to 5 can be defined in each case. Further advantageously realizable features can be found in claim 7.

This means that the resulting calculation time (s), cf. In addition, the preceding dependent claims, so that even with changes in the material flow or at the same objects of the / the allocation parameters are determined so quickly and excluded the evaluation parameters are adapted so quickly

can / can, that even in the changed material flow or for the changed objects, the conditions of claim 2 can be met. This can be realized by a suitable hardware and / or software and / or by suitable, sufficiently low set accuracies.

Further advantageously realizable features can be found in claim 8. According to the first variant of this claim, evaluation parameters can thus be set which are used directly in the evaluation unit in order to directly, ie directly, control the identification and classification (or the process steps of the same) by the evaluation unit. According to the second variant of this claim, evaluation parameters can be set which control the image acquisition unit 1 and which thus indirectly or indirectly, ie as a consequence of this control, influence the identification and classification performed by the evaluation unit (or the process steps thereof). An example of such an evaluation parameter is the image resolution (or the number of pixels per unit area) in the case of

Image Capture Unit: This can be decreased (e.g., by combining multiple pixels). The image data then have a lower image resolution, the allocation parameters thus become less accurate / coarser and / or occur in smaller numbers (and thus can be determined more quickly). The latter then also simplifies or accelerates the identification and classification by or in the evaluation unit indirectly. The latter can also be effected by changing one or more evaluation parameters of the evaluation unit with respect to the setting on the basis of the evaluation parameter (s) of the image acquisition unit.

Further advantageously realizable features can be found in claim 9.

Such a sorting parameter can preferably be set by the sorting unit. One possible sorting parameter is a combination of several rere adjacent exhaust nozzles of the sorting unit to a nozzle cluster. For example, with such a nozzle cluster, large objects can be sorted out more safely or (with a low occupation density) larger blow-out areas around objects to be sorted out can be effected. This makes it possible to further increase the probability of actually sorting out items to be sorted out by the sorting unit.

Further advantageously realizable features can be found in claim 10.

The occupation density can be defined in particular as an average number of objects per unit area of the material flow (for example, as an average number of objects with which a unit area of the conveyor belt of a belt-type sorting system is occupied). For sliding-type or fall-type sorting systems, the occupancy density may be the mean number of objects per fallway or per fall-area unit.

The occupancy distribution can capture or describe whether the objects in the material flow are all singled out or with which probability any object is isolated on, for example, the conveyor belt (or whether objects still overlap or with what probability cluster formations are present).

Further advantageously realizable features can be found in claim 11.

If more than one of these process steps are performed, they are preferably performed in chronological order according to their order in the list of this claim (example: segmentation followed by context analysis followed by classification).

The context analysis as part of the identification may be a "Connected Component Analysis" as described, for example, in "Topological Algorithms for Digital Image Processing" by TY Kong, A. Rosenfeld, North Holland, Amsterdam, NL, 1996. The segmentation as part of the identification may be performed, for example, as described in "Digital Image Processing and Image Acquisition" by B. Jahne, Springer, Heidelberg, Germany, 2012. The identifying and classifying, in particular the identifying, may comprise further process steps such as an image pre-processing step (before the segmentation and before the context analysis) and / or a feature calculation step (after the segmentation and after the context analysis). See, for example, "Digital Image Processing and Image Acquisition" by B. Jahne, Springer, Heidelberg, Germany,

2012 and "Automatic Visual Inspection: Fundamentals, Methods and Practice of Image Acquisition and Image Evaluation", J. Beyerer, FP Leon, C. Frese, Springer-Verlag Berlin Heidelberg, Germany, 2012. After several identification process steps can be a process step of the classification and after a sorting decision process step takes place (the process step of the classification can already include the sorting decision). Further advantageously realizable features can be found in claim 12 and claim 13.

According to claim 13, the implementation of individual process steps can be done both in software and in hardware, but also only in software or only in hardware. Preferably, a central processing unit (server system) is used, with which all units of the sorting system are connected via bidirectional data lines. This central processing unit can computer-aided execute all necessary data processing measures, calculations and / or process steps.

An inventive sorting method can be found in claim 14.

The present invention thus describes a procedure in which a best possible sorting decision can fundamentally be made within the available computing time, in order to thereby enable a principle compliance with real-time barriers. The Improved valorization of the evaluation system (evaluation unit) can thus directly support qualitative advances in optical bulk material sorting, since time limits can be narrowed down. This is realized according to the invention by the use of algorithms which incrementally refine a sorting decision or alternatively align their calculations to a time budget allocated.

A number of implementations which differ in accuracy GK and calculation effort BZ can be used for subtasks of the evaluation system. During the individual process steps (in particular during the evaluation), a runtime system can select concrete implementations in order to maintain real-time barriers and to achieve the best possible result during the available time. The available time budget can be propagated to evaluation algorithms. The algorithms can be enhanced with intelligence so that they make the best possible use of the available time.

According to the invention interruptible sub-processes (the process steps) can be realized. The evaluation algorithm can be interrupted by a control system at any time and it can be queried the best process step result up to this time. By using the best possible decision up to this point in time, information about an object to be sorted can always be evaluated for a sorting decision. It is therefore not the case that an object for a classification or sorting is not considered at all when the time limit is exceeded. This ultimately results in an increase in sorting performance and sorting quality.

According to the invention, it can be ensured that a sorting decision is present for each object contained in the material flow when the separation mechanism (sorting unit) is reached. This decision is usually based on information collected about the object by the sensor (image acquisition unit), the quality of the evaluation differs depending on the time available. This means that a better decision can be made than is the case with conventional prior art systems. Moreover, in the present invention A lower latency between the sensory detection and the separation is made possible, whereby the spatial separation is minimized, an increase in the sorting quality is achieved, and a more compact sorting system is realized.

By the invention, an improved separation in the automated sorting of any bulk materials or material flows can be achieved. In addition, smaller blow-out windows can be used compared to the prior art. The invention can be used for the sorting of complex bulk materials which are classified on the basis of many complex features.

Optical sorting systems as in the present invention can be used if the materials or the objects differ on the basis of optical features, preferably in the visible or also in the near infrared range, and thus can be classified for sorting. The invention can be used for the raw material processing industry, in addition to a cost-effective production always a consistently high quality must be ensured. This includes, for example, the sorting of industrial minerals, for example, to reduce the iron content in a raw material. The invention can also be used in the food or beverage sector, where impurities in products (for example: dried peppers, dried grapes ...) must be eliminated. Another important area of application is the recycling of products (for example, scrap glass sorting).

The present invention can make a significant contribution to improved economy in optical sorting systems. As explained in more detail in the following exemplary embodiment, according to the invention with the help of the obtained sensor data (image data), a material occupancy (for example occupation density) as well as a distribution of the objects in the material flow can be determined by a monitoring unit. This can be achieved through learned models as well as through certain metrics. Based on this knowledge, a single control component can be used

Elements (components or process steps) of the processing chain. For example, with regard to the accuracy can be adjusted. Thus, a shorter calculation time of these members can be achieved. For example, with a high occupancy density and / or an unfavorable distribution in the occupancy (for example, when object clusters are present), the accuracy of at least one member or process step can be set to "coarser" or "lower", so that the real-time conditions in each Case (even if this reduces the accuracy of the classification or sorting decisions, or even if the error rate, ie the probability of a wrong sorting decision or

Classification decision for a currently viewed object is increasing). In accordance with the invention, this greatly increases the probability of compliance with all real-time conditions in the sorting system. This results in better sorting decisions, as a classification can be made at all by said increase in compliance for more objects in the material flow. (In extreme cases, in any case, i.e. for every single object in the

Material stream, a classification decision and thus a sorting decision like.)

According to the invention, the control component (evaluation unit) can set the evaluation parameters, in particular the accuracy, the calculation time and / or the repetition frequency, on the basis of relationships determined with known material flows (with known object occupancy, with known object types, object sizes, object weights, etc.). These relationships can dependencies between the material occupancy or the occupancy parameters on the one hand and the one to be carried out

On the other hand, reflect the settings of the evaluation parameters. Sorting quality and sorting performance result according to the set evaluation parameters. Figures 1 to 5 show an embodiment of an inventive optical sorting system (and a corresponding sorting method) as follows.

Showing:

Figure 1 shows an optical band sorter according to the invention. FIG. 2 shows the processing sequence in the sorting system according to FIG. 1 from the perspective of the digital data (in particular: the image data) and the material flow or the path of its objects.

3 shows the sequence of the sorting process according to FIGS. 1 and 2, in particular the process steps of the evaluation (identification and classification) in the evaluation unit of the system.

4 shows an example of an occupancy parameter determined in accordance with FIGS. 1 to 3 (here: material occupation density) and an evaluation parameter (here: accuracy GK) set using the same, which controls the identification and classification of the objects, that is to say the evaluation, by the evaluation unit.

FIG. 5 shows an example of the process step of the classification in the evaluation according to FIG. 3.

FIG. 1 shows an optical band sorter, which fundamentally follows the structure according to the prior art, wherein the features according to the invention are located in the evaluation unit 2 or the evaluation of the image data 4, respectively. In the belt sorter, a bulk material flow or material flow M is transported by means of a conveyor belt 11 in a manner known per se to a person skilled in the art past an image recording unit 1 to a sorting unit 3 arranged at a defined distance from the image recording unit 1. The bulk material flow M comprises a multiplicity of individual objects 0, which are to be classified or sorted here into only two classes, namely good objects (to be sorted into the collecting container 13b of the sorting unit 3) and bad objects (to be sorted into the further collecting container 13a the sorting unit 3). For this, the objects 0 of the material flow M must first be detected by the image acquisition unit, then classified by evaluation of the image data 4 recorded by this unit 1 in the evaluation unit 2 into good objects and bad objects, and finally sorted. The sorting into or splitting on the two containers 13 a, 13 b according to the sorting or

Klassifizierergebnisses 13 is carried out by the compressed air valves of the sorting unit 3, based on the evaluation results 10 of the evaluation unit 2 bad Remove objects from the material flow M by blowing out.

The blown bad objects fall into the container 13a of the bad objects, the good objects remain in the material flow M, so are not blown out, and fall into the container 13b for good objects. This describes the ideal state of the sorting result 13.

The image recording unit 1 comprises an imaging sensor system, here a CCD color line camera 1a, which detects the material flow M or the objects 0 thereof in the ejection area of the conveyor belt 11 against a background 12 or records a rapid imaging sequence of the material flow M against the background 12 , Illumination lb of the unit 1 illuminates the material flow M against the background 12 of the unit 1 in order to ensure optimum image recording conditions for the camera 1a.

The recorded image data or video data 4 are transmitted via a data line connection between camera la and evaluation unit 2 to the latter. The evaluation unit 2 then carries out the process steps described below in detail for identifying and classifying the objects 0 of the material flow M in the image data 4 and transmits them

Evaluation results 10 of the performed process steps 7a-7g (see Figure 3) via a data link to the sorting unit 3. The latter finally leads spaced from the discharge area of the conveyor belt 11 and of the receiving unit 1, the removal of bad objects, whereby the separation into Good objects (container 13b) and bad objects (container

13a) in the sorting result 13.

FIG. 2 shows the material flow (solid arrows) and the data flow (dashed arrows), in particular the data flow in the image data acquisition and evaluation, in the system according to FIG. 1. The material feed onto the conveyor belt 11 is first followed by a singling and settling of the objects O on the conveyor Conveyor before the transport state or state of the objects O shown in Figure 1 on the left in the image shown in the material flow M on the conveyor belt 11. Finally, in the ejection area, image acquisition or sensory detection of the objects O in the material flow M takes place by means of the image recording unit 1 or the camera 1a thereof. The image data 4 are sent to the Evaluation in the evaluation unit 2 transmitted, which performs the classification of the conclusion of the evaluation or makes the sorting decision for the individual objects 0. During the evaluation with the evaluation unit 2 (ie the execution of all necessary process steps 7, see FIG. 3, or calculation processes), the material flow M downstream of the discharge region of the conveyor belt 11 falls along the discharge parabola. The period of free fall of the objects 0 along the Abwurfparabel (until reaching the Ausschleusungsbereiches of the unit 3) is defined by the distance between the conveyor belt end on the one hand and the impact area of the compressed air valves of the sorting unit 3 on the other hand and corresponds to the latency or the time for a meeting of a sort decision for an object 0 is available.

According to the invention, for each object in the material flow M, cf. see also Figure 4, in this latency a classification and sorting decision like. This is done in the evaluation unit 2. The sorting decision (in good object or bad object) is finally implemented after completion of the latency for each object by the sorting unit 3 (separation). 3 shows a simplified view of the sequence in the sorting system according to FIGS. 1 and 2: After the material feed 30 to the material transport 31 on the conveyor belt 11, the image acquisition 32 takes place by means of the camera 1a of the image acquisition unit 1. Then the evaluation 33 starts by means of the evaluation unit 2 which connects the separation 34 through the sorting unit 3.

According to FIG. 3, the evaluation, that is to say the identification and classifying of the objects 0 by the evaluation unit 2, comprises a total of seven individual process steps 7. The input data of the evaluation are the image data 4. On the basis of the determined assignment parameter 5 (see FIG. 4: material occupation density) For each individual process step 7, exactly one evaluation parameter 6 controlling this process step in the evaluation unit is set. The necessary evaluation of the occupation density for calculating the evaluation parameters 6a-6g of the individual process steps 7a-7g of the identification and classifying by the evaluation unit 2 takes place in FIG

Computer system of the evaluation unit 2 (not shown). In the example shown in FIG. 3, the evaluation unit 2 for the first process step 7a of the image data preprocessing of the image data 4 sets the evaluation parameter GK 6a as the evaluation parameter controlling this preprocessing. The accuracy level can be eg "low", "medium" or "high" (see also FIG. 4).

In the process step 7b of the image cleanup subsequent to the step 7a, a fixed calculation time BZ is set by the evaluation unit 2 as the evaluation parameter 6b. The control of the step 7b of the evaluation is thus based on a BZ. The subsequent third process step 7c of the segmentation is again performed or controlled with an accuracy step GK set by means of the unit 2 as the evaluation parameter 6c. For the correlation analysis 7d, one of the three accuracy level values "low", "medium" or "high" is likewise assigned or set as the evaluation parameter 6d.

The temporally fifth process step of the feature calculation 7e is performed or controlled on the basis of a set repetition frequency WH as the evaluation parameter 6e. The feature calculation here comprises a calculation sequence to be performed recursively, the repetition frequency WH 6e being e.g. may take on a value between 3 and 7 (i.e., the recursion depth of the calculations may be chosen between 3 and 7, where 3 requires little computational effort, thus allowing a fast execution of step 7e and a value of 7 being high

Requires computational effort, so that the implementation of step 7e takes a long time, but can be done with high accuracy - the latter is thus useful or possible only with low occupancy density 5, if 0 sorting decisions are to be made for all objects).

Finally, the classification takes place as the sixth step 7f, an accuracy GK in turn being set as the evaluation parameter 6f for this process step 7f. The final sorting decision 7g also takes place on the basis of setting an accuracy value GK 6g.

The result of the seven process steps 7a - 7g with the respective Because the evaluation parameters 6a-6g set are the sorting result 13 from FIG. 1 for each individual object 0 of the material flow M.

FIG. 4 shows how, in the system of FIGS. 1 to 3, evaluation parameters 6 can be set based on occupancy parameters 5 previously determined in the image data 4. This is shown on the basis of a single assignment parameter 5, in this case the material occupation density, which is determined with the evaluation unit 2 from the image data recorded with the image recording unit 1 or the camera 1a (shown here individual images 4).

From the here exactly one allocation parameter material occupation density 5, which characterizes the material flow M with respect to the number of objects 0 per unit area, by means of the evaluation unit 2 exactly one evaluation parameter 6, here an accuracy (eg - simplified here - exactly one all individual process steps 7 controlling calculation accuracy which determines one of the values "high", "medium" or "low") This accuracy 6 is an evaluation parameter of the evaluation unit 2 which is set at the evaluation unit 2 in order to identify and classify the objects 0 by As shown in FIG. 4, in the case of a low material occupation density 5, a high accuracy 6 is set as the evaluation parameter for identifying and classifying or for all individual process steps thereof, so that the evaluation unit 2 performs all identification and classification process steps 7 with performs the accuracy "high" without causing a violation of the real-time conditions in the evaluation (cf. FIG. 4, middle image line).

If the material occupation density 5 in the material flow M increases to "high", so that the real-time conditions would no longer be maintained while maintaining the high accuracy of the evaluation process steps, the accuracy 6 is set to "low" by the evaluation unit 2 for all process steps 7 , The identification and classification is carried out from this point in time with low accuracy or controlled so that the real-time conditions can be met again (see Figure 4, lower image line). In addition, FIG. 4 shows that evaluation parameters 8 of the image recording unit 1 can also be set on the basis of the material occupation density 5. Thus, for example, as the only evaluation parameter 8 of the image recording unit 1 whose image resolution can be set. At low material occupation density 5, the camera la can image data 4 with high

Image Resolution 8 record, whereas at a significant increase in the material occupancy density 5, as shown in the lower line of Figure 4, the image resolution 8 can be significantly reduced when recording the image data 4. The thus adjustable evaluation parameter 8 of the image acquisition unit 1 (image resolution) can then indirectly also influence the determination of the occupancy parameter (s) 5 and thus also the setting of the evaluation parameter (s) 6 of the evaluation unit 2.

Finally, FIG. 5 shows an example of a possible implementation of one of the process steps or subtasks from FIG. 3, namely the classification

7f. The example shows an implementation of an interruptible

Classifier for making the best possible classification decision (or the following sorting decision). The basis for the classifier is a decision tree. In each inner node of the decision tree, a feature of a found object is compared against a threshold value, and according to the comparison result, the path is followed to the right or to the left, as shown in FIG. The leaves of the decision tree store the resulting classes. According to FIGS. 1 to 4, there are only two classes here, namely the class of the good objects and the class of the bad objects.

According to the invention, various methods known per se to those skilled in the art for automatically learning a corresponding decision tree on the basis of a given amount of training can be used. See, for example, "Machine Learning" by T.M. Mitchell, McGraw-Hill,

Boston, USA, 1997.

According to the invention, these methods can be extended to the effect that not only in the leaves of the decision tree is a class membership stored, but also that in each inner decision tree node a probability for the different classes is added. (based on the training amount). For example, if you distinguish two classes, namely a class A for good objects and class B for bad objects, where there are equal numbers of examples in the training set for both classes, then the probabilities for the class membership for A and B are in the top node f ₈ at 50% each.

Here, f describes the feature vector and the eighth feature included is described by f ₈ accordingly. It should be noted that a single feature in several nodes of the decision tree can be used as a test criterion.

An object to classify 0 falls after the comparison with f ₈ in the first, top node in the left or in the right subtree. The next node, for example right, is now compared with f ₄ . The probabilities can be redistributed. For example, it is conceivable that in the latter node the probabilities for A at 30% and for B with

70% are deposited. If the process step of the classification is interrupted at this point (or if the calculations have reached this point), then it can certainly be said that, according to current knowledge, a membership in class B is more probable than in class A.

According to the invention, the process step of the context analysis 7d (see FIG. 3) can be realized, for example, by a "Connected Component Analysis." See, for example, MB Dillencourt et al "A General Approach to Connected Component Labeling for Arbitrary Image Representations". Journal of the ACM 39 (2), 1992.

In this case, several breakpoints can be granted by displaying the image e.g. is subsampled with a shifted raster. For example, in a first pass, only every fourth pixel can be considered: for all pixels not considered, the same value can then be assumed as for the pixel last considered. If enough computation time is available, the grid can be moved: this way, the information about the image is refined with each pass or repetition. It is an iterative process, after each

Can provide valid information, but this information is cessive improved.

Interruptible sub-processes or process steps can also be realized, for example, as follows. The area of an object is determined in pixels. For this purpose, the pixels are simply counted in an algorithm. Such a process can be interrupted at any time and the previously counted pixels can be assumed to be an area. Alternatively, the area can be estimated from other data, if already available. For example, if an axis-aligned bounding box is known, the area of this box can be used as an estimate instead of the true area (this can provide advantages through fewer memory accesses).

In general, in the context of the invention, iterative methods or iteratively performed process steps can be interrupted during each loop pass. In the event of a break during a loop it is possible to use the result of the most recently completed loop pass.

Claims

FRAUNHOFER-GESELLSCHAFT...eV 179PCT 0617 patent claims

1. Sorting device with which objects (0) of a material stream (M) are sorted, comprising an image recording unit (1) with which the material stream (M) is optically recorded and with which image data (4) of the same (M) are generated Evaluation unit (2), with which objects (0) in the material stream (M) are identified and classified, and a sorting unit (3), with which classified objects (0) in the material stream (M) are sorted, characterized in that with the evaluation unit (2) one or more occupancy parameters (5), which characterize the material stream (M) with regard to its occupancy with the objects (0), are determined from the generated image data (4) and that on the basis of the certain occupancy parameter(s) (5) one or more evaluation parameters (6, 8) controlling the identification and classification of the objects (0) by the evaluation unit (2) is/are set.

2. Sorting device according to the preceding claim, characterized in that the evaluation parameter(s) (6, 8) is/are set so that for each object (0) in the material stream (M) with a probability of >90%, preferably > 95%, preferably >99%, particularly preferably 100%, both identification and classification are carried out by the evaluation unit (2).

3. Sorting device according to one of the preceding claims, characterized in that one or more accuracy(s) GK with which (a) process step(s) of identifying and classifying is/are carried out by the evaluation unit (2) is/are set as the evaluation parameter (6).

4. Sorting device according to one of the preceding claims, characterized in that the evaluation parameter (6) is one or more calculation time(s) BZ, with which (a) process step(s) of identifying and classifying is carried out by the evaluation unit (2). /be, is/are discontinued.

5. Sorting device according to one of the preceding claims, characterized in that the evaluation parameter (6) is one or more repetition frequency (s) WH, with which (an) iterative, recursive or incremental calculation sequence (s) of (a ) process step(s) of identification and classification to be carried out by the evaluation unit (2) is/are to be repeated, is/are set.

6. Sorting device according to one of the preceding claims, characterized in that the identification and classification of the objects (0) in the material stream (M) by the evaluation unit (2) is carried out in several process steps (7a, 7b, ...) to be carried out in succession or in parallel. ) is carried out, with one or more evaluation parameters (6a, 6b, ...) being set for each of the process steps (7a, 7b, ...).

7. Sorting device according to one of the preceding claims, characterized in that the determination of the occupancy parameter(s) (5) and/or the setting of the evaluation parameter(s) (6, 8) is carried out in real time.

8. Sorting device according to one of the preceding claims, characterized in that one or more evaluation parameters (6) of the evaluation unit (2) are/are set based on the specific occupancy parameter(s) (5), i.e. (an) evaluation parameter in which one or more process step(s) (7a, 7b, ...) of identification and classification to be carried out with the evaluation unit (2) is/are controlled by the evaluation unit (2), and/or that based on the specific occupancy parameter(s) (5), one or more evaluation parameters (8) of the image recording unit (1) is/are set, i.e. (an) evaluation parameter with which one or more processes/processes in the image recording unit ( 1) is/are controlled, which influence/influence the implementation of one or more of the process steps (7a, 7b, ...) of identification and classification by the evaluation unit (2).

9. Sorting device according to one of the preceding claims, characterized in that additionally (a) sorting parameter(s) controlling the sorting of classified objects (0) by the sorting unit (3) is set on the basis of one or more specific occupancy parameter(s). /become.

10. Sorting device according to one of the preceding claims, characterized in that an occupancy density and/or an occupancy distribution of the objects (0) in the material stream (M) is/are determined as the occupancy parameter (5).

11. Sorting device according to one of the preceding claims, characterized in that the identification and classification of the objects (0) by the evaluation unit (2) contains one, several or all of the following steps as a process step (s):

• a segmentation step,

• a step of context analysis, and/or

• a classification step using one or more training set-based decision tree(s).

12. Sorting device according to the preceding claim, characterized in that during the classification, probability(s) based on a training set(s) is/are stored in each node of the decision tree(s).

13. Sorting device according to one of the preceding claims, characterized by one or more preferably computer or microcontroller-supported computing unit (s), by means of which the recording, generating, identifying, classifying, sorting, determining, controlling and / or that Adjustment is/are carried out. Optical sorting method for sorting objects (0) of a material stream (M), characterized in that a sorting device according to one of the preceding claims is used to carry out the method.