Methods for training a computer-aided system, methods for analyzing filling material, and computer-aided system

Abstract

A method for training a computer-aided system (10) to analyze filling material (12) is shown, comprising the following steps: - Providing at least one sample comprising a down feather (14) and / or a feather (16) from land or waterfowl, and optically capturing the sample using a camera (50) to obtain image data of the sample, and / or providing image data of at least one sample comprising a down feather (14) and / or a feather (16) from land or waterfowl; - Analyzing the image data using computer-aided image recognition of the computer-aided system (10), whereby the sample is automatically classified and a classification result for the sample is output; - Check whether the classification result of the analyzed image data matches the at least one provided sample; and - Adjusting the computer-aided image recognition if the classification result differs from the provided sample. Furthermore, a method for analyzing filling material (12) using a computer-aided system (10) and a computer-aided system (10) are described.

Description

[0001] The invention relates to a method for training a computer-assisted system for analyzing filling material, in particular filling material for outdoor equipment, pillows and blankets, e.g., bedding. Furthermore, the invention relates to a method for analyzing filling material using a computer-assisted system and a computer-assisted system for analyzing filling material.

[0002] Down is one of the most popular filling materials for pillows and duvets. Down comes from the underpants of waterfowl, namely geese and ducks, and retains heat close to their bodies. Unlike feathers, down does not have a central quill, but rather secondary branches, also known as "filaments." These secondary branches are held together by a down core and are softer than the vane of a feather.

[0003] Down and feathers possess an astonishing range of functions. They are capable of insulating heat and absorbing moisture, for example. At the same time, they are breathable, which protects the body from cooling and annoying moisture. They are therefore suitable and used as filling material for bedding, such as blankets and pillows, and outdoor equipment. The thermal insulation properties of feathers are even surpassed by those of down. The fine branching of the down traps a comparatively large volume of air, which has excellent insulating properties, making down particularly suitable and in demand for body protection. In addition, down is lighter than feathers, so that the use of down can drastically reduce both the volume and weight. The higher the quality of the down, the less down is needed to achieve the same thermal insulation.

[0004] The EU standard EN 12934 has been in force in Europe since 2001. This is a technical standard that encourages manufacturers and suppliers to inform consumers about the composition of the filling used in products, for example, bedding, and to advertise this. According to this standard, for example, the labels of duvets and pillows must list the percentage ratio of down to feathers. A voluntary but widely used indication is the quality class, e.g., "Class 1." The quality class division is based on the proportion of other elements in the filling. The more broken feathers, torn down filaments (down feathers), poultry feathers, or residues such as quill pith and / or waste the filling contains, the lower the quality is considered. In addition, foreign matter, e.g.,A percentage of natural fibers other than feathers and down, and any type of man-made fibers or non-fibrous synthetic material in the filling means that, according to current labeling regulations, the filling may no longer be marketed as consisting of down and feathers. Another voluntary but widely used labeling requirement is the type of poultry used for the filling.

[0005] In order to guarantee the quality of the down used and to verify the information provided by manufacturers and suppliers, the Association of the German Feather Industry (VDFI) randomly purchases various bedding products on the market and checks whether the information on the composition of the filling material is correct, for example the content of down, feathers, other elements and the type of poultry stated. The testing is carried out in accordance with the European testing standard EN 12131. Currently, this is a time-consuming process that usually takes four to six hours and can only be carried out by specially trained and qualified persons. During this process, a sample of the filling material is taken and the individual components of the sample are separated from one another using tweezers and then individually examined and assessed, for example under a microscope, to ensure correct testing.Since the number of qualified people is constantly decreasing and due to the time-consuming analysis, the associated costs are very high.

[0006] Accordingly, it is an object of the invention to enable the analysis of the filling material, ie the recognition of the components and their assignment, in a more cost-effective manner without accepting any loss of quality in the assessment.

[0007] The object is achieved according to the invention by a method for training a computer-aided system for analyzing filling material, comprising the following steps: - Providing at least one sample comprising a down and / or a feather from land or water fowl, and optically capturing the sample by means of a camera to obtain image data of the sample, and / or providing image data of at least one sample comprising a down and / or a feather from land or water fowl; - analyzing the image data by means of computer-aided image recognition of the computer-aided system, wherein the sample is automatically classified by the computer-aided image recognition and a classification result for the sample is output by the computer-aided image recognition; - Checking, using the computer-assisted system, whether the classification result of the analyzed image data matches the at least one sample provided; and - Adjusting the computer-aided image recognition if the classification result deviates from the provided sample.

[0008] The computer-assisted system, in particular the computer-assisted image recognition of the computer-assisted system, is trained to classify or categorize a sample based on at least one image of the sample or based on image data of the sample, i.e. to determine the type of sample. This means that the computer-assisted system is first trained to recognize down or a feather or to differentiate between them. The invention is therefore based on the idea of ​​training a computer-assisted system in such a way that automatic classification of the filling material is possible, i.e. automatic recognition of the components contained in the filling material, since it is automatically recognized whether the components of the filling material are, for example, a feather or down. Through continuous training of the computer-assisted system, the classification of the system can be continually improved, i.e. the recognition of down and / or feathers.This can also be done during ongoing operations, using a trained computer-assisted system. Automating the classification process reduces costs and time. Furthermore, human error can be eliminated, which can even improve the quality of the assessment.

[0009] When analyzing the image data, (predefined) characteristics of the sample are used to classify the sample, i.e., to determine whether it is down and / or feather. These characteristics can include the shape of the feather, especially its tip, the presence of a quill, the presence of secondary branches, characteristics of the secondary branches, and / or the shape of the quill or the secondary branches.

[0010] Based on these characteristics, it is possible, among other things, to distinguish between feathers and down and / or to determine the type of poultry.

[0011] The image data can be processed to analyze areas of the optically captured sample where the features are present. For example, the image captured by the camera is first analyzed to identify the at least one feather and / or down in the captured image, thus isolating it from its surroundings.

[0012] The at least one feather or down can then be analyzed in detail to identify areas of the feather or down, for example the tip of the feather, a quill and / or secondary branches. Once the individual areas of the sample, i.e. the at least one feather or down, have been identified, for example the tip of the feather, a quill and / or secondary branches, these areas are analyzed in order to be able to be classified. For this purpose, data stored in a library can be used, for example, to compare this data with the image data. The classification results of the individual areas of the feather or down can then be combined to obtain a classification result for the feather or down.For example, the shape of the feather's tip, the presence of a quill, the presence of a vane, particularly consisting of branches, and / or the presence of secondary branches with relatively widely spaced nodes can lead to a classification result of "goose feather," meaning the sample is a feather rather than down, and the poultry species is identified as a goose. However, if the shape of the tip is different, e.g., more pointed, and / or the analysis of the node spacing on the branches and secondary branches of the vane shows that the nodes are close together, the classification result can be "duck feather," meaning the sample is a feather, and the poultry species is identified as a duck.If the analysis reveals the presence of a secondary feather and / or the analysis of the individual secondary branches with regard to the presence of “nodes” shows that the secondary branches do not have any nodes, the classification result can be “chicken feather”, i.e. the sample can be a feather and the poultry species can be identified as chicken.

[0013] However, if the sample is determined to be down, for example, because it lacks a quill, a vane, and / or a tip, individual filaments of the sample (down) can be analyzed. Based on the knot spacing on individual filaments, the poultry species of the down can be identified and assigned. If the knots are close together, the classification result may be "duck down," while if the knots are farther apart, the classification result may be "goose down." The knot spacing is only visible under high magnification, which is why a high-resolution camera is necessary.

[0014] The image captured of the sample can therefore be segmented or divided into a grid, for example, divided into individual pixels to isolate and classify individual areas of the sample. To do this, features are extracted from the image, i.e., the image data. In principle, segmentation can be performed in which an element of an image is localized down to the nearest pixel.

[0015] As an alternative to segmentation, the at least one feather or down can also be classified as a whole using computer-aided image recognition, i.e., identified as either down or feather. In this case, no segmentation takes place, as the sample is classified as a whole.

[0016] The sample preferably comprises a mixture of several different types of down and / or feathers, whereby a quantitative composition of the sample is output. Therefore, it is not a single feather or a single piece of down that is analyzed, but a mixture that includes at least down and / or feathers, e.g., from different poultry species. The mixture can also include other elements. The computer-assisted image recognition is therefore trained to analyze the proportions of the respective components of the mixture. To do this, it is first necessary to classify the individual components of the mixture, i.e., to recognize them as down, feather, etc., as already explained above. In a second step, the quantitative composition can then be determined.Based on the quantitative composition, it is possible to draw conclusions about the quality class of the mixture, so that the information provided by the manufacturers and / or suppliers of the filling material can be verified.

[0017] According to one variant, a portion of filling material in the sample is reported that comes from neither landfowl nor waterfowl. This allows the proportion of foreign matter in the sample to be determined. Foreign matter is considered to be substances and materials that come from neither landfowl nor waterfowl. This includes, in particular, a portion of synthetic filling material and / or cotton or other substances, such as dust and / or dirt.

[0018] Landfowl, such as chickens, do not have down, so only feathers from landfowl can be used as filling material. It is important to note that long feathers, i.e., feathers from the wing, may not be used in high-quality bedding. These long feathers can also be classified through appropriate training of the computer-assisted system. For this purpose, for example, the shape of the quill and / or the vane of the feather is analyzed, particularly the length of the vane or the distance from the end of the vane to the end of the quill.

[0019] According to another variant, the composition of the sample is determined. The composition depends on the proportion of down, the proportion of feathers, the proportion of broken feathers, the proportion of down fluff, the proportion of fluff, and / or the proportion of filling material that comes from neither landfowl nor waterfowl. The quality class depends on the proportion of "other elements" in the filling material and the "poultry species."

[0020] When analyzing the image data using computer-aided image recognition, a classified poultry species is preferably also output as part of the classification result. The poultry species output as the classification result are, in particular, geese, ducks, and chickens. The poultry species also influences the quality class. Goose down, in particular, is considered very good, while filling material with chicken feathers is assigned a lower quality class.

[0021] According to one variant, the classification result is checked automatically by providing the computer-assisted image recognition system with a predefined sequence of known samples. This way, the computer-assisted image recognition system already receives an evaluation result for each sample to be analyzed during training and can independently compare the respective classification result with the associated evaluation result. If the classification result output by the computer-assisted image recognition system during training does not match the evaluation result, the computer-assisted image recognition system automatically adjusts. This allows the training process to be automated to a certain extent, which accelerates the training of the computer-assisted system and reduces costs.If this has been done for a number of samples, especially different samples, it is ensured that the computer-aided image recognition is trained.

[0022] Training is therefore carried out using training datasets that contain the image data to be analyzed and the corresponding evaluation results. Based on the image data to be analyzed, computer-aided image recognition outputs the classification result, which is then compared with the evaluation result contained in the training dataset to determine any deviations. If an intolerable deviation occurs, computer-aided image recognition is adjusted or automatically adjusted.

[0023] Preferably, image recognition based on artificial intelligence is used to analyze the image data. This is particularly advantageous because artificial intelligence is capable of learning, allowing for continuous improvement in classification. The previously described steps—segmentation, detection, and classification—can be performed using image recognition based on artificial intelligence.

[0024] Artificial intelligence can, for example, be a machine learning method, in particular a convolutional neural network.

[0025] During training of the computer-aided system, the weighting factors of the computer-aided image recognition can be adjusted if the classification result deviates from the provided sample. The weighting of the individual connections between layers of artificial intelligence can therefore be changed during training. This allows the quality of the classification to be continuously improved, resulting in higher-quality and more reliable classification results. In other words, recognition is improved.

[0026] Furthermore, the object is achieved according to the invention by a method for analyzing filler material using a computer-assisted system. The method comprises the following steps: - providing at least one sample comprising a down and / or a feather from land or waterfowl; - Optically capturing the sample using a camera to obtain image data of the sample; - Analyzing the image data using computer-aided image recognition of the computer-aided system; - Automatic classification of the sample using computer-aided image recognition; and - Outputting a classification result for the sample using computer-aided image recognition.

[0027] By classifying using the computer-assisted system, the time required can be drastically reduced, so that the process now takes only 4 to 6 minutes instead of the previous 4 to 6 hours. Furthermore, costs are significantly reduced.

[0028] According to one variant, a classified poultry species and / or a quantitative composition of the sample is output, in particular wherein a proportion of filling material in the sample is output that comes neither from landfowl nor from waterfowl. Instead, this proportion consists, for example, of synthetic filling material and / or cotton or other substances, e.g. dust and / or dirt. If the proportion of foreign components or foreign substances is 0.5% or more, this means in Europe that the filling material may no longer be labeled as down filling or feather filling. As already mentioned with regard to the training procedure, the quality class can be determined from the quantitative composition, in particular the proportion of foreign components or foreign substances, and the identified poultry species, thus enabling an independent classification of the quality class.Since the computer-assisted system has been trained to output the quantitative composition of the sample as a classification result, the trained computer-assisted system is also capable of doing so.

[0029] The composition depends on the proportion of down, the proportion of feathers, the proportion of broken feathers, the proportion of down feathers, the proportion of feathers, and / or the proportion of filling material that comes from neither landfowl nor waterfowl. In addition to the quality class, this also allows one to assess whether the information on the label and, if applicable, the price of the product is justified.

[0030] Preferably, the computer-assisted system has been trained using the previously described method for training a computer-assisted system for classifying filling material. As already explained above, this ensures reliable and accurate classification.

[0031] Furthermore, the object is achieved according to the invention by a computer-aided system for analyzing filling material, wherein the computer-aided system comprises a camera module and an evaluation module.

[0032] The computer-assisted system is configured to perform the previously described process for analyzing filler material. In other words, the computer-assisted system enables reliable, automated, and accurate classification of the filler material, thus reducing both the time required and the associated costs.

[0033] The evaluation module can comprise a processor or be connected to a processor, with the computer-assisted image recognition being stored on the processor. Additionally, the evaluation module can also include weight recognition stored on the processor, so that the classification result also includes the quantitative weight distribution of the sample.

[0034] Computer-aided image recognition is also capable of capturing sample dimensions, for example, to determine the size or volume of the sample, especially of down or feathers. Once computer-aided image recognition has classified the sample, density values ​​can be used based on the classification result to determine the sample's weight based on the optically detected size or volume and the classification result.

[0035] According to one embodiment, the camera module comprises at least one high-resolution camera. The camera can be movable along a rail or be permanently mounted on a housing of the computer-assisted system. Of course, the camera module can also comprise multiple cameras that provide image data from different detection angles. This is particularly advantageous for determining the size or volume of the sample.

[0036] A swirling or mixing unit can be provided, which is designed to swirl the filling material to be classified, in particular to mix components of a mixture. This ensures that heavier components are not completely obscured by the lighter components and do not influence the classification result. This can further increase the degree of automation, as the sample or mixture to be examined does not have to be spread out manually beforehand.

[0037] It can be provided that the filling material to be classified is swirled or mixed once, with an image being taken via the camera module during or after this process, or alternatively that the camera module takes a series of images of the filling material to be classified during the swirling or mixing process.

[0038] Further features and advantages of the invention will become apparent from the following description and the drawings, to which reference is made. The schematic drawings show: - Fig. 1 is a schematic representation of a first embodiment of a computer-aided system according to the invention; - Fig. 2 a schematic representation of a second embodiment of the computer-aided system according to the invention; - Fig. 3 an overview of a method according to the invention for analyzing filling material; - Fig. 4 an overview of a method according to the invention for training a computer-aided system according to the invention; - Fig. 5 representations of feathers, where in Fig. 5A a duck feather, in Fig. 5B a goose feather, in Fig. 5C a chicken feather and in Fig. 5D a long spring is shown; - Fig. 6 illustrations of broken springs, where Fig. 6A a broken duck feather and in Fig. 6B shows a broken goose feather; - Fig. 7 representations of bent springs, where in Fig. 7A a bent duck feather and in Fig. 7B shows a bent goose feather; - Fig. 8 illustrations of damaged springs, where Fig. 8A a damaged duck feather and in Fig. 8B shows a damaged goose feather; - Fig. 9 representations of down, where in Fig. 9A and Fig. 9B different forms of down, in Fig. 9C an immature down and in Fig. 9D a down feather is shown; and - Fig. 10 representations of different side branches, where the side branch in Fig. 10A from a duck, in Fig. 10B from a goose and in Fig. 10C comes from a chicken.

[0039] Fig. 1 shows a computer-aided system 10 according to the invention for analyzing filling material 12, in particular filling material for outdoor equipment, pillows and blankets, e.g. bedding.

[0040] The filling material 12 shown is a mixture of down 14 and feathers 16. In addition, the filling material 12 shown includes so-called other elements, such as broken feathers, in particular broken goose feathers 18 and broken duck feathers 20, down fluff 22, feather fluff and / or filling material that originates neither from a landfowl nor from a waterfowl, for example synthetic filling material 24, cotton 26 and / or dust or dirt. The other elements can also include bent feathers and damaged feathers, in particular bent goose feathers 28, bent duck feathers 30, damaged goose feathers 32 and damaged duck feathers 34. In this respect, the other elements include both filling material that originates neither from a landfowl nor from a waterfowl, and filling material that originates from a landfowl or a waterfowl, but is, for example, broken, bent and / or damaged.

[0041] The individual components of the filling material 12 are partly also in Fig. 5 to 10 can be seen in detail.

[0042] The down 14 can be either goose down or duck down.

[0043] Similarly, feathers 16 are divided into goose feathers 40, duck feathers 42, and landfowl feathers. Landfowl feathers, for example, may be chicken feathers 44.

[0044] In order to be able to classify the filling material 12, the computer-assisted system 10 comprises a sample chamber 46, a camera module 48 with at least one camera 50 and an evaluation module 52.

[0045] For analysis, the filling material 12 is placed in the sample chamber 46 so that the filling material 12 can be optically captured by the at least one camera 50 of the camera module 48. Image data is generated, which is transmitted from the camera module 48 to the evaluation module 52 for further analysis, as explained in detail below.

[0046] The evaluation module 52 includes and / or has access to a processor on which computer-assisted image recognition is stored. Additionally, the evaluation module 52 can include additional programs, such as weight detection.

[0047] As in Fig. 1, the camera module 48 can comprise a carriage 54 to which the camera 50 is attached, wherein the carriage 54 is movable along a rail 56. The carriage 54 can thus be driven, so that the camera 50 is moved via the carriage 54 along the rail 56. This ensures that the entire sample of the filling material 12 can be captured by the camera module 48. This is particularly important if several down and / or feathers are to be analyzed simultaneously using the camera module 48, i.e. the sample to be analyzed is a mixture, and not each component of the mixture is analyzed individually, which would be more time-consuming.

[0048] In Fig. 2 shows an alternative embodiment of the computer-assisted system 10, in which the camera module 48 comprises a plurality of cameras 50 that are rigidly mounted on the walls of the sample chamber 46 of the computer-assisted system 10. Advantageously, the cameras 50 are adjusted to capture different areas of the sample chamber 46 in order to obtain the most complete possible image of the filling material 12. For adjustable alignment, the camera module 48 can include pivotable mounting means so that the angular orientation of the cameras 50 is adjustable.

[0049] In addition, the computer-assisted system 10 shown here comprises an (optional) swirling and / or mixing unit 58, which is configured to swirl or mix the filling material 12 to be classified. This ensures that no heaps of components of the filling material 12 form, in which heavier and / or smaller components are hidden by the lighter and larger components. The swirling and / or mixing unit 58 can be controlled such that a homogeneous distribution of the filling material 12 in the sample chamber 46 is achieved, as shown in Fig. 2. Using the multiple cameras 50, a three-dimensional image of the components of the filling material 12 can then be created, thereby generating corresponding image data that is transmitted to the evaluation module 52, in particular the computer-assisted image recognition.

[0050] The method used by the computer-aided system 10 for analyzing fill material 12 is shown schematically in Fig. 3, to which reference is made below.

[0051] In a first step S1, a sample of filling material 12 is provided, which comprises at least one down 14 and / or one feather 16 from land or water fowl. If a computer-assisted system 10 according to the embodiment in Fig. 1 is used, the filling material 12 should be spread out along the entire length so that there is as little or no overlap of the individual components of the filling material 12 of the sample.

[0052] The sample of filler material 12 may be a subset of a filler material 12 that was delivered as raw material for further processing or has already been used in a product. In other words, the sample of filler material 12 may be a random sample to be analyzed.

[0053] Next, in step S2, the sample of the filling material 12 is optically captured by the camera module 48, in particular the at least one camera 50 of the camera module 48. For this purpose, the camera module 48 is moved along the rail 56 via the carriage 54. During the process, the at least one camera 50 of the camera module 48 captures multiple images to obtain image data of the sample. Alternatively, the at least one camera 50 of the camera module 48 can also capture only one image, so that the image data only comprises the one image.

[0054] Does the computer-based system 10 correspond to the Fig. In the embodiment shown in Figure 2, the filling material 12 does not need to be distributed over the entire length of the computer-assisted system 10 intended for this purpose. Instead, the filling material 12 can be poured into the sample chamber 46. For optical detection by means of the camera module 48, the swirling and / or mixing unit 58 is switched on, so that the filling material 12 to be classified is swirled or mixed, and the individual constituents or components of the filling material 12 are mixed. This ensures that individual constituents are not completely covered.

[0055] During image capture, a single image can be generated by a single camera 50 of the camera module 48 to determine the image data, or an image can be generated by multiple cameras 50. Alternatively, it is also possible for at least one of the cameras 50 of the camera module 48 to take multiple images, based on which the image data is generated. Several or even all cameras 50 of the camera module 48 can also take multiple images, based on which the image data is generated.

[0056] It is also conceivable to use a computer-aided system 10 which has both a movable camera module 48 and a swirling and / or mixing unit 58.

[0057] In principle, the respective camera 50 can be a video camera that records a video, i.e., a sequence of individual images. The video or the sequence of individual images is then analyzed by the evaluation module 52. For this purpose, the video can first be broken down into its individual images, which can then be analyzed in their entirety, or only a subset of the individual images, e.g., every tenth individual image of the video.

[0058] Regardless of the specific embodiment of the computer-aided system 10, image data is generated which is used for further analysis.

[0059] These image data are then transmitted to the evaluation module 52 in a further step S3 and analyzed by the evaluation module 52 using the computer-aided image recognition of the computer-aided system 10.

[0060] Computer-aided image recognition, for example, is based on artificial intelligence in the form of a neural network, particularly a convolutional neural network. Of course, other well-known image recognition methods can also be used for computer-aided image recognition.

[0061] Based on the computer-aided image recognition analysis, the sample is automatically classified in a fourth step S4, and a classification result for the sample is subsequently output in a step S5. The classification result may include a quantitative composition of the sample, i.e., a percentage indication of the respective constituents or components of the entire sample of the filling material 12 that was analyzed. This may also be referred to as the composition of the sample of the filling material 12.

[0062] The classification result may also include a classified poultry species. Consequently, the computer-assisted system 10 can determine not only the proportions of down 14, feathers 16, and other elements, e.g., broken feathers, down fluff 22, feather fluff, synthetic filler 24, and / or cotton 26, but also the poultry species from which the down 14 and feathers 16 originate.

[0063] The classification result can therefore be, for example, that the sample of filling material 12 which has been analyzed contains 95.5% goose down, 4.3% goose feather and 0.2% other elements.

[0064] The other elements can, if desired, be further classified, for example as broken feathers, down fluff 22, feather fluff and synthetic filling material 24 and cotton 26. In this respect, the classification result can be output, for example, that the 0.2% further elements comprise 85% down fluff 22, 10% feather fluff, 4% broken feathers and 1% cotton 26.

[0065] In order to ensure such a classification reliably, the Fig. 4 outlined methods for training a computer-aided system 10 for classifying filling material 12 are used.

[0066] In a first training step S10, a sample is first provided which comprises a down 14 and / or a feather 16 from land or water fowl.

[0067] At the beginning of the training process, one or more down 14s of the same type or one or more feathers 16 of the same type can be provided as samples. As the training process continues, down 14s of different types, i.e., goose down and duck down, or feathers 16 of different types, i.e., goose feathers, duck feathers, and chicken feathers, can be provided as samples. Finally, during the training process, mixtures can be provided as samples that include both down 14, feathers 16, and other components or constituents, i.e., other elements, which ultimately corresponds to the later application.

[0068] As already explained in the method for classifying the filling material 12, the sample, if it is a mixture, is preferably distributed over the entire sample area of ​​the computer-assisted system 10 so that there is as little overlap as possible between the components of the sample or mixture.

[0069] Here, too, in a next training step S11, at least one, preferably several, images of the sample are captured by the at least one camera 50 of the camera module 48. The image data to be analyzed are obtained from these images.

[0070] If the training method uses a computer-aided system 10 according to the embodiment according to Fig. 1 is used, the camera 50 of the camera module 48 can be moved over the carriage 54 and the rail 56, provided that a mixture is used as a sample for training.

[0071] If the training method uses a computer-aided system 10 according to the embodiment according to Fig. 2 is used, the sample or mixture can be swirled or mixed by means of the swirling and / or mixing unit 58.

[0072] Independently of the computer-aided system 10 used, image data can be generated independently, which are used for training, as explained below.

[0073] Additionally, a training data set can be provided which contains an evaluation result, provided that the image data itself is generated based on a real sample of filling material 12, i.e. real down 14 and / or feathers 16.

[0074] Alternatively, a training data set can be provided for the entire training process, which contains image data from at least one sample as well as an expected evaluation result for the corresponding sample. In other words, the computer-assisted system 10 is fed exclusively with data in this training process.

[0075] The training procedure can also include a mixed data input, i.e. generating image data based on an optically acquired sample and using image data that is provided as a dataset.

[0076] In any case, there is image data that needs to be analyzed by computer-aided image recognition, namely either self-generated image data or image data contained in the training data set.

[0077] When analyzing the image data by the computer-aided image recognition of the computer-aided system 10 (see training step S12), the sample is automatically classified and a classification result for the sample is output.

[0078] Since the sample may contain a mixture of several different types of down 14 and / or feathers 16, a classified poultry species should also be output as part of the classification result. The poultry species to be classified are, in particular, geese, ducks, and chickens. Visual classification is possible for both down 14 and feathers 16, allowing the identification of goose and duck down, as well as goose, duck, and chicken feathers.

[0079] First, a distinction is made between a down 14 and a feather 16. Down 14 is often also called down feathers 15 and forms the undercoat of the plumage of waterfowl; see Fig. 9D.

[0080] A down 14 consists of tufts of light, fluffy filaments growing from a faintly indicated down core 60; see Fig. 9A. The filaments are, for example, down branches and / or downy branches. Down 14 is only referred to as down 14 if at least two down or downy branches are connected to one another at a single point. This is why down 14 looks similar to a snowflake in pictures.

[0081] The down and / or down feather branches correspond to goose branches 62 and duck branches 64. Chickens also have chicken branches 66, but these are part of the normal plumage and not part of a down 14. The corresponding differences in the branches are based on the Fig. 10A to 10C, which show the node spacing for a single filament, which will be discussed below.

[0082] Feathers 16, on the other hand, come from the horny covering of the poultry and have a short and soft vane 68. In contrast to down 14, feathers 16 have a well-developed quill 70, which consists of a coil 72 and a shaft 74. This is evident from the Fig. 5A to 5D clearly stand out.

[0083] Chicken springs 44 in particular are easy to identify, as they often have a secondary spring 76 which branches off from the same coil 72 of the normal spring 16 and is smaller than this; see Fig. 5C. Feathers 16 of waterfowl, on the other hand, look more similar, but a visual distinction is still possible. While goose feathers 40 look as if they were cut off, duck feathers 42 are pointed. In this respect, the feathers 16 of waterfowl differ at their respective tips, as can be seen from a comparison of the Fig. 5A and Fig. 5B. The various springs 16 are each in Fig. 5A, Fig. 5B and Fig. 5C shown.

[0084] Ideally, the filling material 12, especially in high-quality bedding, should not contain long feathers 80. As in Fig. As can be seen in Figure 5D, long feathers 80 have a longer and stiffer vane 68 than a feather 16 and originate from the stiff, coarse wing and tail feathers. Furthermore, the quill 70 of the long feathers 80 is straight, in particular straighter than those of the feathers 40, 42 intended for high-quality bedding, which are found in the Fig. 5A and Fig. 5B are shown.

[0085] Based on these characteristics, a distinction can be made between the individual poultry species and the feather type for feathers 16.

[0086] In the Fig. Figures 6 to 8 show broken, bent, and damaged goose and duck feathers 18, 20; 28, 30; 32, 34, which belong to the other elements. A broken feather has a broken quill 70 and more than 40% of the quill is missing, or it is a quill 70 without a vane 68. A bent feather, on the other hand, has a broken quill 70 that is not yet completely severed. A damaged feather has an intact quill 70 and thus retains its original shape, but the vane 68 is damaged or has missing branches or parts of the vane.

[0087] These are also unique features that can be determined with the help of the computer-assisted system 10, in particular computer-assisted image recognition, when analyzing the image data.

[0088] When it comes to down type 14, however, a distinction is initially made only between goose down and duck down. Chickens, which are considered land poultry, do not have down type 14.

[0089] The difference between goose and duck down can be seen in the filaments, especially in the secondary branches, as described above. The difference lies particularly in the distance between the nodes 78; see Fig. 10A to 10C. While the nodes 78 in goose secondary branches 62 are relatively far apart, the nodes 78 in duck secondary branches 64 are comparatively small. The individual secondary branches also differ in the shape of their nodes 78 and their size, as shown in the Fig. 10A and Fig. 10B can be seen.

[0090] For the sake of completeness, a chicken branch 66 is also included in Fig. 10C, which has more rounded nodes 78 compared to the slightly wedge-shaped goose and duck secondary branches 62, 64. Furthermore, the spacing of the nodes 78 is characteristic of a chicken secondary branch 66.

[0091] In the next training step S13, it is checked whether the classification result of the analyzed image data matches the at least one provided sample. This check determines whether the computer-assisted image recognition system has correctly analyzed the sample.

[0092] For this purpose, a known sample can be presented to the computer-assisted image recognition system in a predefined sequence, so that the evaluation module 52 receives an evaluation result for each sample, which is compared with the classification result. As explained above, the evaluation result can be part of the training data set.

[0093] If the training process has just begun, only samples of one type can be presented to the computer-assisted system 10, in particular the computer-assisted image recognition system, as already explained. In this respect, several hundred or several thousand duck down feathers, in particular their image data, can initially be analyzed, with the evaluation result being the same in each case. This can be done analogously with goose down, duck feathers, goose feathers, and chicken feathers. Different down types and / or feather types can then be mixed, whereby different evaluation results must then be taken into account. Later in the training process, the computer-assisted system 10, in particular the computer-assisted image recognition system, can then be presented with a mixture as a sample or several different mixtures as samples to be analyzed.

[0094] If the classification result output by the computer-aided image recognition based on its analysis deviates from the provided sample or the associated evaluation result, the evaluation module 52 adapts the computer-aided image recognition, as shown in training step S14.

[0095] In particular, the weighting factors of the artificial intelligence, i.e., the weightings between individual layers of artificial intelligence, are adjusted. This process is repeated until the classification result output by the computer-aided image recognition corresponds to the provided sample or the assigned evaluation result.

[0096] Preferably, as already explained above, the evaluation module 52 is first familiarized with a training data set, thus creating a foundation on which the actual training takes place. It is advisable to first build a data set that exclusively contains images of precisely identifiable down 14. The training data set can then be expanded to include image data of feathers 16 from waterfowl and landfowl, before additional image data for filling material that originates neither from landfowl nor from waterfowl are provided. Alternatively, it is of course also possible to first use a data set containing image data of feathers 16 from waterfowl and landfowl, followed by a data set containing images of precisely identifiable down 14. The order of the data sets is therefore irrelevant, as long as they contain image data of one and the same type.

[0097] The training of the computer-assisted system 10 may also include the trained computer-assisted system 10 having further analysis capabilities in addition to the pure differentiation between down 14 and feather 16 and the associated poultry species.

[0098] Based on the classification result, a quality class can also be determined, which can be compared with the quality class specified by the manufacturer or supplier. This allows for verification of whether the information provided is correct and whether the price is fair for the consumer. For this purpose, the composition of the sample of filling material 12 is determined, as explained above.

[0099] Ideally, immature down 82, such as in Fig.9C. Immature down 82 has, instead of the down core 60, a feather edge 84 from which the secondary branches extend. In other words, immature down 82 is not yet fully developed.

Claims

[1] A method for training a computer-aided system (10) for analyzing fill material (12), comprising the following steps - Providing at least one sample comprising a down (14) and / or a feather (16) from land or water fowl, and optically capturing the sample by means of a camera (50) to obtain image data of the sample, and / or providing image data of at least one sample comprising a down (14) and / or a feather (16) from land or water fowl; - analyzing the image data by means of computer-aided image recognition of the computer-aided system (10), whereby the sample is automatically classified and a classification result for the sample is output, - Check whether the classification result of the analyzed image data matches the at least one provided sample, and - Adjusting the computer-aided image recognition if the classification result deviates from the provided sample. [2] Method according to claim 1, characterized by that the sample comprises a mixture of several different types of down (14) and / or feathers (16), wherein a quantitative composition of the sample is output. [3] Method according to claim 1 or 2, characterized by that a proportion of filling material (12) is emitted in the sample which originates neither from a land fowl nor from a waterfowl, in particular a proportion of synthetic filling material (24) and / or cotton (26). [4] Method according to one of the preceding claims, characterized bythat a composition of the sample is output, the composition depending on the proportion of down (14), the proportion of feathers (16), the proportion of broken feathers, the proportion of down flight (22), the proportion of feather flight and / or the proportion of filling material which originates neither from a landfowl nor from a waterfowl. [5] Method according to one of the preceding claims, characterized by that when the image data is analyzed using computer-aided image recognition, a classified poultry species is output as part of the classification result. [6] Method according to one of the preceding claims, characterized bythat the checking of the classification result is carried out automatically by specifying a predefined sequence of known samples to the computer-assisted image recognition, so that the computer-assisted image recognition receives an evaluation result for each sample, and wherein the computer-assisted image recognition automatically compares the respective classification result with the associated evaluation result. [7] Method according to one of the preceding claims, characterized by that artificial intelligence-based image recognition is used to analyze the image data. [8] Method according to one of the preceding claims, characterized by that weighting factors of computer-aided image recognition are adjusted if the classification result deviates from the provided sample. [9] Method for analyzing filling material (12) by means of a computer-aided system (10), comprising the following steps - providing at least one sample comprising a down (14) and / or a feather (16) from land or water fowl, - optically capturing the sample by means of a camera (50) to obtain image data of the sample, - analyzing the image data by means of computer-aided image recognition of the computer-aided system (10), - Automatic classification of the sample, and - Output a classification result for the sample. [10] Method according to claim 9, characterized by that a classified poultry species and / or a quantitative composition of the sample is output, in particular wherein a proportion of filling material (12) in the sample is output which originates neither from a land fowl nor from a waterfowl, for example a proportion of synthetic filling material (24) and / or cotton (26). [11] Method according to one of claims 9 to 10, characterized bythat the computer-aided system (10) has been trained using a method according to one of claims 1 to 8. [12] Computer-assisted system (10) for analyzing filling material (12), wherein the computer-assisted system (10) comprises a camera module (48) and an evaluation module (52), and wherein the computer-assisted system (10) is configured to carry out the method according to one of the preceding claims. [13] Computer-aided system (10) according to claim 12, characterized by that a swirling or mixing unit (58) is provided which is designed to swirl the filling material (12) to be analyzed, in particular to mix components of a mixture.

Images