Method for training a decision model

Abstract

The invention relates to a method for evaluating optical lenses (1) which have been checked by means of an optical image recording device (10), wherein the method comprises at least the following steps: setting up a lens evaluation unit (14) with a machine learning module (19) on which a digital decision model (19a) is operated; setting up a decision database (17) into which a number of lenses (1) with specific quality criteria (18) are entered; evaluating the lens (1) with regard to the specific quality criteria (18) by at least one person (I, II); inputting the evaluation of the lens (1) by the at least one person (I , II) into the machine learning module (19); and comparing the specific quality criteria (18) from the decision database (17) and the evaluation of the lens (1) by the at least one person (I, II) in the machine learning module (19), so that the machine learning module (19) is trained on a continuously improved evaluation of the lenses (1) by the lens evaluation device (14) by comparing the evaluation of the lens (1) with the specific quality criteria (18).

Description

[0001] Kassel, December 10, 2025 Attorney's file SCH24203P1WO Official file number

[0002] NN

[0003] Applicant:

[0004] Schneider GmbH & Co. KG

[0005] Biegenstraße 8

[0006] 35112 Fronhausen

[0007] Representative:

[0008] WALTHER PATENT

[0009] Walther · Bayer · Faber

[0010] Heimradstraße 2

[0011] 34130 Kassel, Germany

[0012] METHOD FOR TRAINING A DECISION MODEL

[0013] The invention relates to a method for evaluating optical lenses that have been tested using an optical image acquisition device.

[0014] STATE OF THE ART

[0015] From EP 3 531 067 A1, a device for carrying out a method for measuring optical lenses is known, which uses a first radiation source in the form of a laser and a second radiation source in the form of a screen, so that both the position of the lens in space and the refractive power distribution of the lens over its surface can be determined. The surface refractive power distribution is also referred to as a power map. In order to detect defects in the lens after the processes of determining the lens position or the first surface of the lens in space and after determining the surface refractive power distribution of the lens, the lens must generally be fed to a further station to detect the defects and to treat the lens accordingly.For example, DE 3 943 909 A1 discloses a method for generating data about the lens that takes the defect into account, so that the lens is subsequently subjected to edge processing in which as many defects as possible are located in the edge area, which is removed when the mold body is edge-processed to form the spectacle lens.

[0016] This is achieved by identifying the position and size of any defects present in a semi-finished or finished lens after it has been provided. If corresponding edge data is recorded, a comparison can be made between this data set and the semi-finished or finished lens, each containing at least one defect. The semi-finished or finished lens is then assigned to a category. This category takes into account the type and position of the defect, so that, depending on the lens category, the edge of the semi-finished lens is designed to minimize the presence of the defect in the final lens.

[0017] WO 2019 / 238518 A2 discloses a method for material testing of optical lenses with an optical imaging system, the method comprising the following steps: at each of at least one pose of the lens in relation to the imaging system: controlling the optical imaging system to capture at least two images of the lens by means of at least one angle-variable illumination and angle-variable detection; for each of the at least one pose: processing the corresponding at least two images to generate a result image with digital contrast; and based on the at least one result image generated from the at least two images at the at least one pose: performing a material test of the lens.

[0018] The subsequently published DE 10 2024 126 710 A1 discloses a method for measuring optical lenses using a device, wherein the device comprises: a detection device arranged on a first side of the lens with respect to a plane of extension of the lens, one or more first radiation sources with which the position of the lens in space is determined, and a second radiation source with which a planar refractive power distribution of the lens over the surface is determined, wherein a planar extending, moving pattern is generated with the second or a further radiation source, which is detected through the lens by the detection device and by means of the computer unit, one or more defects on the surface of the lens are detected.The lens is detected, whereby the lens remains in an associated lens holder for determining the position of the lens in space - if necessary, determining the planar refractive power distribution of the lens over the surface and for detecting at least one defect.

[0019] Cosmetic inspections in contact lens manufacturing are mostly carried out manually. While there are usually a set of rules regarding, for example, the number, depth, or length of permissible damage to the lenses, such as scratches, these cannot realistically be adhered to in everyday production. Moreover, this is often not the only relevant criterion, as the overall impression is frequently important. The variability in the assessments can be high due to various "intentional" factors, such as different manufacturers with varying quality standards, or the fact that the same manufacturer may offer different quality levels with different price structures, for example, for a basic and a high-end category, or for different countries supplied by the manufacturer.There are also various “unwanted” factors, for example, that a standard is difficult to define because there is no objective measure for the impression of a lens; that is, human testers are always subjective due to the real measurement conditions.

[0020] If one wants to eliminate the human factor by using a measuring machine, the difficulty still lies in providing a system that can be adapted to a poorly defined standard without major expenses.

[0021] Defect detection during quality control should be as precise as possible; technically speaking, it should be possible to determine the defect class down to the pixel level. In the context of defect detection, this means, for example, that for each pixel in the lens image, it must be clear whether it belongs to, for instance, a scratch, a speck of dust, an inclusion, or a transparent lens element. However, a decision based solely on pixel-level information, and potentially including neighboring pixels, is prone to error. Different types of defects can be distinguished not only by pixel values ​​but also, for example, by their shape.

[0022] This can be achieved by using modern segmentation models, such as those based on a so-called UNet; these combine pixel-precise resolution with a large receptive field that encompasses the image area around a pixel, which can be used to determine the class of that pixel. UNets with a so-called ResNet backbone are particularly well-suited for this purpose, but other deep neural networks such as FCN, DeepLab, SegFormer, and MaskRCNN can also be used.

[0023] The difficulty for companies in assessing whether a lens is a good or bad part lies in the fact that currently no uniform standard can be defined and adhered to across shifts and production sites. In particular, it is not possible to train a new or even existing employee to comply with a specific standard, as manual assessment always involves a significant individual component.

[0024] In summary, with regard to the invention, it can be stated that manufacturers of spectacle lenses generally strive to achieve a certain quality level for the lenses they produce. In reality, however, spectacle lenses are currently only inspected manually for surface defects, for which only a short amount of time is usually allocated per lens. The human factor introduces a significant degree of uncertainty into the overall system, as the lack of time prevents the human inspector from precisely measuring, categorizing, and counting individual surface defects, but rather requires an overall assessment within a very short time. This results in natural variance when, for example, comparing inspector A with inspector B and / or with inspector C, but also when considering the decision-making behavior or...would compare the evaluation behavior of examiner A at different times of the day or on different days.

[0025] This system lacks an obvious, objective quality level, which makes setting up a measuring machine that performs the quality inspection described above particularly difficult, as it is not clear how the machine should be parameterized and which limits it should apply to classify lenses with specific defect patterns as good or bad. In particular, lenses may be evaluated very differently by different inspectors.

[0026] The problem lies in finding a threshold value, or a set of threshold values, that allows for objective decision-making. The only information available is that which human inspectors can provide based on the defects present on the lens. These inspectors merely implement the company's own quality rules imprecisely, which only provide an incomplete picture of the actual quality criterion in manufacturing, because these rules never consider the overall appearance of a complete lens, but only the presence of specific defects of a certain severity.

[0027] REVELATION OF THE INVENTION

[0028] It is therefore the object of the present invention to provide a method for evaluating optical lenses with which a limit value can be provided as accurately as possible to categorize a lens as a good part or as a bad part.

[0029] This problem is solved starting from a method for operating a method for evaluating optical lenses according to the preamble of claim 1 and in conjunction with the respective characterizing features. Advantageous embodiments of the invention are specified in the dependent claims.

[0030] The invention includes the technical teaching that the method comprises at least the following further steps: setting up a lens evaluation device with a machine learning module on which a digital decision model is operated, setting up a decision database into which a number of lenses with specific quality criteria are read, evaluating the lens with regard to the specific quality criteria by at least one person, inputting the evaluation of the lens by the at least one person into the machine learning module and comparing the specific quality criteria from the decision database and the evaluation of the lens by the at least one person in the machine learning module, so that the machine learning module is trained to continuously improve the evaluation of the lenses by the lens evaluation device by comparing the evaluation of the lens with the specific quality criteria.

[0031] It is crucial that the lens evaluation device is trained to assess the lens's specific quality, namely whether the lens is recognized as a good part or a bad part, for which a good / bad threshold is defined in the machine learning module as a result of the training.

[0032] For this purpose, the machine learning module can be trained with artificial intelligence and / or neural networks, so that the machine learning module is trained in conjunction with the decision database to output the good / bad threshold value, indicating whether the lens is a good part or a bad part.

[0033] It is also advantageous to perform the training in the machine learning module with artificial intelligence and / or neural networks using statistical methods designed to eliminate a subjective factor of evaluation by at least one person in the available data, so that the data, freed from the subjective factor, reflects an objective quality standard of the customer.The training of the decision-making process using the lens evaluation device can at least include the step of forming data pairs from the defects in the lens image and the expected evaluation of the lens, whereby the defects are fed to the machine learning module as input information, which calculates a decision and / or whereby a mathematical comparison function is defined between the calculated decision and the expected decision, the minimum of which corresponds to an optimal agreement between the second machine learning module and the expected decisions.

[0034] The decisions of several people can be unified into a higher-level decision using artificial intelligence and / or neural networks, with the higher-level decision being based on the specific quality criteria from the decision database.

[0035] Furthermore, it can be taken into account that lenses rated with greater variance by several people receive less or no weighting when training the decision model in the machine learning module, and lenses rated identically by several or all people receive more weighting. This allows for further optimization of the pass / fail threshold determination. Additionally, different weightings for individual reviewers based on their experience level are also conceivable.

[0036] Furthermore, a defect detection device can be set up to which test images via the lenses are passed from the image acquisition device, wherein the defect detection device detects defects on the lens and outputs a lens image comprising the at least one detected and classified defect to the lens evaluation device.

[0037] In conjunction with the defect detection device, a defect database can be set up in which different defects are stored, wherein the defect detection device has a first, upstream machine learning module that performs a segmentation, wherein the defect in the test image is recognized and classified based on the defects stored in the defect database, and with which the lens image is output encompassing the at least one recognized and classified defect.

[0038] Thus, prior to evaluating the lens with regard to its specific quality using the lens evaluation device, a defect detection device can be installed, with which the following steps are performed: feeding the test image from the image acquisition device to the defect detection device, detecting and classifying the defect on the lens using the defect detection device, outputting a lens image comprising the at least one detected and classified defect, setting up a lens evaluation device and feeding the lens image from the defect detection device to the lens evaluation device and evaluating the lens with regard to a customer's specific quality.

[0039] This allows a number of lenses from a customer to be initially recorded by the lens evaluation unit of the testing system. This serves to train the lens evaluation unit on a customer-specific quality level. Subsequently, one or more quality inspectors independently submit evaluations for the same lens. The resulting data set is used to train the lens evaluation unit, which learns the average evaluation criteria of the evaluator(s) and, after training, always makes the same objective decision.

[0040] Using such a trained lens evaluation algorithm, manual quality inspectors from various customers can subsequently be trained so that their decisions align as closely as possible with those of the inspection system according to the learned, objective quality standard. However, the lens evaluation can also be based on lens images that have been pre-classified pixel-by-pixel by the defect detection system.

[0041] The intended outcome is that the lens evaluation, based on the specific quality of the customer, will be incorporated into the lens evaluation system, determining whether the lens is classified as good or defective. Lenses that exhibit no defects whatsoever will, of course, be examined in the same manner and subsequently classified as good.

[0042] The detected and classified defect in the lens image can be stored as information in the lens image, particularly pixel-accurately, by means of the first, upstream machine learning module, preferably comprising a summary of the defect information about the type, location, intensity, number and / or size such as length and / or width and / or area of ​​the defect, wherein the summary of the defect information is passed to the lens evaluation device along with the lens image.

[0043] Preferably, at least one or exactly one defect detection device can be set up centrally, while several lens evaluation devices can be set up decentrally and / or in multiple locations, to which customer-specific decision databases are assigned.

[0044] Furthermore, based on the decision of the second machine learning module as to whether the lens is a good part or a bad part, inspectors can be trained in their decision-making behavior.

[0045] The training of the defect detection using the defect detection device can be carried out by comparing the expected lens image with a calculated lens image from the first machine learning module, in which a first comparison function between the expected lens image and the calculated lens image is calculated.

[0046] The detected and classified defect in the lens image is then stored as information within the lens image, particularly pixel-accurately, using the first machine learning module. This information preferably includes a summary of the defect details regarding the type, location, intensity, number, and / or size (e.g., length, width, and / or area). This summary of defect information can then be transferred to the lens evaluation unit along with the lens image.

[0047] The decision database is set up to store lens evaluations according to the customer's specific quality criteria, particularly on the customer's side, who uses the testing system with its specific quality criteria for lens evaluation. The lens evaluation unit includes a second machine learning module that assesses the digitally processed, at least one defect in the lens image based on the specific quality criteria represented in the decision database and outputs a decision as to whether the lens is to be rated as a good or bad part. It is possible for one or more defect detection units to be set up centrally, for example, at the manufacturer of the testing system, while multiple lens evaluation units are set up on the customer's side, thus multiple and / or decentralized, and each is assigned its own customer-specific decision database.Of course, it is also possible to set up all facilities and databases decentrally, and data maintenance, updates, and / or provision can be carried out via data exchange using data transmission, for example, via a cloud system. It is also conceivable to manage only the defect database for the defect detection device on the manufacturer's side, either through manufacturer-side storage and online data provision via data transmission, or through customer-side storage and manufacturer-side data maintenance via online data provision.

[0048] A further advantage is that the first and / or second machine learning module can each be trained using artificial intelligence and / or neural networks. This allows the machine learning modules to be trained in conjunction with the defect database or the decision database, respectively, to output the lens image or the decision as to whether the lens is a good or a defective part. Based on the decision of the second machine learning module as to whether the lens is a good or a defective part, human, i.e., manually operating, quality inspectors can be trained to make the most accurate decision possible.

[0049] If a machine learning module is used for defect detection, it must be trained, for example, in the case of a deep neural network, to recognize defects. For this purpose, the machine learning module is presented with test images, from which it calculates a lens image containing defects. This calculated lens image is then compared with an expected lens image, which was manually created to match the original test image, by defining a comparison function between the expected and calculated lens images. The machine learning module is characterized by the fact that it performs mathematical operations that depend on parameters to arrive at a decision. Training, therefore, means that these parameters are adjusted so that the calculated and expected lens images become more similar to each other.The machine learning module becomes better at recognizing and classifying the correct defects, which is usually formulated mathematically in such a way that this corresponds to a minimum of the comparison function.

[0050] The decision-making process includes at least one step: using the lens defects as input, to generate an evaluation of the lens. If a machine learning module is used for this purpose, it usually needs to be trained. This requires data pairs consisting of defects on a lens and expected evaluations based on these defects. A mathematical comparison function can then be defined between the calculated decision and the expected decision. The minimum of this function corresponds to an optimal match between the machine learning module's evaluation and the expected decisions. This comparison function can be used in a mathematical procedure to adjust the parameters of the machine learning module to approximate a minimum of the comparison function.

[0051] The problem preceding the invention is further solved by a testing system for testing optical lenses, comprising an optical image acquisition device for detecting at least one defect (see page 9 - segmentation) on the lens and for outputting a digitally processed test image, wherein, according to the invention, a defect detection device is provided with which the at least one defect on the lens can be detected and classified based on the test image, so that a lens image can be output by the defect detection device which includes at least one classified defect, and wherein, furthermore, a lens evaluation device is provided with which, based on the lens image, an evaluation of the lens can be carried out based on a specific quality of a customer.

[0052] Although separating the system into a defect detection unit and a lens evaluation unit offers the advantage of modularity—requiring only the replacement of the defect detection unit to introduce new defect types or correct errors in defect detection, and only the replacement of the lens evaluation unit to adapt to individual customer quality criteria—implementing both steps in a single unit would also be possible. In this case, greater modularity would be traded for more efficient computation and a simpler design.

[0053] In particular, the testing system is designed so that the defect detection unit is assigned a defect database and the lens evaluation unit a decision database. Since the defect database can be used across multiple customers, it typically contains significantly more images (at least several thousand) than the customer-specific decision databases (typically around 100 to 1000 records each). Both databases can also be continuously updated, i.e., enriched, during operation. Machine learning modules enable automated learning and comprise systems that recognize patterns and rules from datasets using either guided or self-learning methods. In machine learning, pattern recognition is achieved through statistical methods based on a training dataset. This contrasts with classical programming, where fixed algorithms and rules are defined.However, these must then be defined precisely and completely and be known to the developer, which can be very difficult for many questions.

[0054] Neural networks can form a machine learning model consisting of multiple interconnected artificial neurons. Each neuron processes input values ​​and outputs a value that is influenced by adjustable parameters. By layering multiple levels of neurons, deep neural networks are created, where the output of one neuron serves as the input for subsequent neurons. This enables the network to learn complex patterns.

[0055] The machine learning modules can also utilize so-called "convolutional neural networks," or CNNs for short. These networks are a special type of neural network that is particularly well-suited for processing image data and spatial information. CNNs are designed to recognize patterns and features in local image areas, making these systems especially effective for image and object recognition. The unique feature of a CNN lies in its "convolutional layers," which use filters—also called kernels—to "scan" small areas of the input image. These filters extract features such as edges, shapes, and textures, which are used as fundamental information for subsequent layers. By applying these layers multiple times, the network is enabled to recognize increasingly complex patterns as the learning process progresses.

[0056] In image processing, segmentation refers to the division of images into segments, which can be achieved by the defect detection system within the lens image. In the context of defect detection, this means that for each input image, a cropped and / or scaled output image is provided, in which each pixel is labeled with its predefined class. For example, in the context of defect detection, this means that for each pixel in the lens image, it is determined whether it belongs to, for instance, a scratch, a dust particle, an inclusion, or a transparent lens element. Such a task can be handled, for example, by a deep neural network, resulting in a segmentation model.

[0057] Finally, there are so-called UNets; these refer to a specific architecture, i.e., the structure and / or organization of the individual layers of a neural network, which can be used for segmentation tasks and can also be part of machine learning modules. In the first part, the resolution of the input image is gradually reduced, as with most CNNs, to enable the recognition of increasingly complex image structures. However, in the second part of the network, these structures are gradually enlarged again and combined with the less complex, but higher-resolution structures from the first part, in order to represent the complex structures precisely and at high resolution. This structure schematically resembles a U-shape. UNets represent one way to build segmentation systems with a neural network, but alongside other architectures, they are a modern "standard" model for this field.There is still considerable variation in the precise structure within the individual levels and cross-connections of the UNet. The invention therefore proposes a system with which an objective limit value can be determined, based in particular or exclusively on a decision database or a table in which the following is stored for one or more example lenses: Detected defects of the defect detection device, evaluation of the lens by tester A, evaluation of the lens by tester B, evaluation of the lens by tester C, evaluation of the lens by tester D, etc., for multiple testers or testers at different times.

[0058] To create this decision database, a set of lenses is first measured by the image acquisition and defect detection unit to identify defects. The same lenses are then given to different human inspectors for evaluation, without them knowing the machine's results or the results of their own inspections.

[0059] The resulting database reflects the specific quality criterion present in the work of the system's recipient, but is distorted by the human factor.

[0060] The decision module contains a machine learning module that receives the lens image with defects or summarized information about the defects along with the lens as input. However, within the scope of the invention, it is also conceivable to use a system that defines fixed rules for the presence of a maximum number of defects of a specific intensity and / or area in a zone on the lens and compares these rules with the detected defects. In both cases, these parameters must be trained by comparing the defects detected from the stored assignments with the assessments of inspectors A, B, C,... From this comparison, parameters for the machine learning module or rules for the rule-based system are derived. These rules are abstracted from the database and derived assuming noise in order to find an objective quality criterion.

[0061] Machine learning modules, such as deep neural networks, require training, preferably using statistical methods designed to handle and, where possible, eliminate noise in the existing data. We intend to leverage this property so that, based on the resulting database, we can abstract and / or define a set of parameters and / or rules that, after removing noise (i.e., subjective factors), reflects an objective quality standard for the customer.

[0062] Once the decision module is configured to reflect such an objective quality standard, it can also be used for employee training, for example: A new batch of lenses is evaluated using the testing system. Independently, one or more employees undergoing further training then inspect the lenses without knowing the results from the testing station evaluating the same lenses from the same batch. After the manual evaluation of each lens, the employee receives feedback on the identified defects and the objective assessment. This trains the employee to recognize the correct defects and to correctly classify them as OK / Not OK.

[0063] PREFERRED EXAMPLE OF THE INVENTION

[0064] Further measures improving the invention are described in more detail below, together with a description of a preferred embodiment of the invention, with reference to the figures. Figure 1 shows a schematic view of the process for evaluating optical lenses according to the invention.

[0065] Figure 2 shows a schematic representation of the inspection system, consisting of the image acquisition device, the defect detection device and the lens evaluation device.

[0066] Figure 3 shows a schematic representation of the structure of the defect database, as well as the training of the machine learning module of the defect detection device and

[0067] Figure 4 shows a schematic representation of the training of the lens evaluation device.

[0068] Figure 1 shows a schematic view of the process for evaluating optical lenses that have been tested using an optical image acquisition device. The core of the process is the setup of a lens evaluation device 14 with a machine learning module 19, on which a digital decision model 19a is run. This also involves setting up a decision database 17, into which a number of lenses 1 with specific quality criteria 18 are read.

[0069] Simultaneously, the lenses 1, which are exemplified by lenses A, B, C, D, E, F, are evaluated with regard to the specific quality criteria 18 by at least one person I, II, exemplified by testers I, II, III, IV. Subsequently, the evaluation of the lenses 1, i.e. lenses A, B, C, D, E, F, by at least one person I, II is entered into the machine learning module 19.Finally, a comparison is made between the specific quality criteria 18 from the decision database 17 and the evaluation of the lens 1 by at least one person I, II in the machine learning module 19, so that the machine learning module 19 is trained to continuously improve the evaluation of the lenses 1 by the lens evaluation device 14 by means of a comparison of the evaluation of the lens 1 with the specific quality criteria 18, whereby the training of the decision-making takes place via the training step in the decision model 19a via training step 110.

[0070] The image acquisition device 10 is followed by the defect detection device 12, which outputs the defects of lenses A to F. The lens evaluation device 14 is shown with dashed lines following the defect detection device 12, since, according to the invention, the training step 110 follows first, and no evaluation of the lens 1 takes place yet. In training step 110, the detected defects from the defect detection device 12 and the manual evaluations of inspectors I to IV are transferred to the lens evaluation device 14 for training the machine learning module 19, in order to ultimately determine whether a lens 1 is to be evaluated as a good part 1a or as a defective part 1b. Retrospective training of inspectors I to IV is also possible, as they receive the final evaluation and learn from it to better evaluate identical or similar defects in the future.The result is the elimination of noise, i.e., examiner-specific, and therefore subjective, factors.

[0071] Figure 2 schematically shows the structure of the inspection system 100 with the features of the present invention. The essential structural structure of the inspection system 100 comprises an image acquisition device 10, a defect detection device 12, and a lens evaluation device 14. The image acquisition device 10 can detect defects on the lens 1 and convert them into at least one digitally processed test image. Image acquisition devices 10 generally include a camera 21 with which an image can be captured, which is generated by means of a screen 22 and captured through the lens 1. The image acquisition device 10 has further functions as disclosed in WO 2023 041 659 A1.

[0072] The image acquisition device 10 thus enables the provision of images that are such that various defects can be detected and distinguished within them. If several stripe patterns are recorded with the camera 21, which were generated by means of the screen 22, they can then be processed to create several output images, which are referred to here as test images 11 and which serve as input images in the next step.

[0073] In the following step, the defect on lens 1 is detected and classified using the defect detection device 12. For this purpose, a defect database 15 is set up in which different defects are stored, wherein the defect detection device 12 has a first machine learning module 16, with which the defect in the test image 11 is detected and classified based on the defects stored in the defect database 15 and with which the lens image 13a comprising the at least one detected and classified defect is output as an accompanying data set.In other words, the at least one detected and classified defect in the lens image 13a is stored as information in the lens image 13a, particularly pixel-accurately, by means of the first machine learning module 16, whereby a summary of the defect information about the type, location, intensity, number and / or size such as length and / or width and / or area of ​​the defect can be generated, whereby the lens image and / or a summary of the defect information 13b is transferred to the lens evaluation device 14.

[0074] In a further step, the lens image and / or the summary of the defect information 13b is transferred from the defect detection unit 12 to the lens evaluation unit 14, where the lens 1 is evaluated with regard to a specific quality for a customer, who can be customer A or customer B and can define different specific quality criteria and their standards. Thus, at least one or exactly one defect detection unit 12 can be centrally located, while several lens evaluation units 14 can be decentralized and / or multiple, to which customer-specific decision databases 17 are assigned.

[0075] The lens evaluation device 14 bases its decisions on a decision database 17, in which defects detected on lenses 1 and their evaluation according to the respective specific quality criterion 18 of the customer currently being trained (generally, that of customer A or B) are stored for customer-specific evaluation. The lens evaluation device 14 has a second machine learning module 19, with which the digitally processed at least one defect in the lens image and / or in the summary of defect information 13b is evaluated based on the specific quality criterion 18 represented in the decision database 17, and a decision is issued as to whether the lens 1 is a good part 1a or a defective part 1b. For different quality criteria 18, customer A or B can specify customer-specific criteria 18, e.g., by means of...Several hundred lens-specific decision databases 17 are created, forming the basis for evaluation by a specific machine learning module 19 for each specific quality criterion. Although the image acquisition unit 10, the defect detection unit 12, and the lens evaluation unit 14 are referred to here as separate components, they are generally implemented in a single physical machine and separated only in the software implementation. This separation is not strictly necessary, however. The first machine learning module 16 could also be designed to process the camera images into the test image 11 and / or perform the tasks of the second machine learning module 19. In particular, depending on the architecture of the machine learning modules, separate training and subsequent merging of the modules is trivially possible.

[0076] The defect database 15 and the decision database 17 are generally only needed during the training of the model when using machine learning / AI. In production use, only the rules or parameters learned by the model based on these databases 15 and 17 are typically used.

[0077] Figure 3 shows a schematic representation of the structure of the defect database 15, as well as the training of the machine learning module 16 of the defect detection device 12. Several image acquisition devices 10 are shown as examples, each of which transmits test images to the defect detection device 12, for instance, from a number of different production sites. The defect detection device 12 and / or at least the defect database 15 with the defects stored therein can, for example, be located and / or hosted at the manufacturer of the inspection system 100. The defect database 15 is thus created for subsequent training, initially with the test images 11 from many lens images, without detecting defects or making final decisions. This database is supplemented by annotating the collected images, whereby all existing defects / anomalies are drawn into each test image 11 and sorted into categories.From this, data pairs are generated from stored test images 15a and expected lens images 15c.

[0078] The defect detection device 12 has a first machine learning module 16, which detects and classifies the defect in the test image 11 based on the defects stored in the defect database 15. The defect detection process, which involves training generalized rules about the manifestations of defects to continuously improve the function of the first machine learning module 16, is described below.

[0079] The stored test images 15a are presented as input to the neural network of the first machine learning module 16. This module delivers a calculated lens image 15b, which is compared with the expected lens image 15c from the defect database 15. The training of the machine learning module 16 is then formulated as a mathematical optimization problem by calculating a first comparison function V1 between the expected lens image 15c and the calculated lens image 15b. This function is minimized when the calculated result of the model from the data set in the defect database 15 exactly matches the expected result, the expected lens image 15c. The variables of this function are all parameters in the model, which are then minimized using an algorithm across all images in the defect database 15.

[0080] Ideally, the machine learning module 16 "learns" generalized rules about the appearance of defects, rather than simply comparing them to previously known defects. Figure 4 shows a schematic representation of the training process for the lens evaluation unit 14. First, for example, several hundred lenses, the lens array 1.1, are received by the inspection system 100. From the defect detection unit 12 of the inspection system 100, a lens image and / or a summary of the defect information 13b, along with the detected and classified data, is transferred to a decision database 17 for each data point. The decision database 17 is built up from at least one of the specific quality criteria 18, based on the several hundred lenses. The inspection system 100 then forwards the respective lens 1 to human inspectors for evaluation.The lens evaluation is recorded via an input interface 20 and, assigned to the defects detected by the testing system, stored in the decision database 17. The resulting data pairs are then passed to the lens evaluation unit 14 for training.

[0081] The goal of training the second machine learning module 19 is to map the specific quality criteria 18. Data pairs consisting of defects of a lens 1 and the expected evaluation of the lens 1 are extracted from the defect database 15. The defects are provided to the second machine learning module 19 as input information, which calculates a decision. A second mathematical comparison function V2 is defined between the calculated decision (i.e., whether it is a good part 1a or a defective part 1b) and the expected evaluation 24. The minimum of this function corresponds to an optimal agreement between the second machine learning module 19 and the expected decisions. The parameters of the second machine learning module 19 are the variables of the function. These are iteratively optimized during training to minimize the comparison function V2. The invention is not limited in its implementation to the preferred embodiment described above.Rather, a number of variations are conceivable, which utilize the presented solution even in fundamentally different designs. All features and / or advantages arising from the claims, the description, or the drawings, including design details or spatial arrangements, can be essential to the invention both individually and in various combinations. List of reference numerals:

[0082] 1 lens

[0083] 1a Good part

[0084] 1b Bad part

[0085] 1.1 Lentil group

[0086] 10 Image capture device

[0087] 11 Test image

[0088] 12 Defect detection device

[0089] 13a Lens image

[0090] 13b Lens image and / or summary of defect information

[0091] 14 Lens evaluation device

[0092] 15 Defect database

[0093] 15a Stored test image (from the database) 15b Calculated lens image

[0094] 15c Expected lens image (from the database) 16 First machine learning module

[0095] 17 Decision database

[0096] 18 specific quality criteria

[0097] 19 second machine learning module

[0098] 19a Decision model

[0099] 20 Input interface

[0100] 21 camera

[0101] 22-inch screen

[0102] 24 expected ratings

[0103] 100 examination system

[0104] 110 Training step A, B Customer

[0105] I, II examining person

[0106] V1 first comparison function V2 second comparison function

Claims

Claims:

1. Method for evaluating optical lenses (1) tested by means of an optical image acquisition device (10), the method comprising at least the following steps: - Setting up a lens evaluation device (14) with a machine learning module (19) on which a digital decision model (19a) is operated, - Setting up a decision database (17) into which a number of lenses (1) with specific quality criteria (18) are read, - Evaluation of the lens (1) with regard to the specific quality criteria (18) by at least one person (I, II), - Input of the evaluation of the lens (1) by the at least one person (I, II) into the machine learning module (19) and - Comparison of the specific quality criteria (18) from the decision database (17) comprising the evaluation of the lens (1) by at least one person (I, II) and the objective decision made by the machine learning module (19), so that the machine learning module (19) is trained to continuously improve the objective evaluation of the lenses (1) by the lens evaluation device (14) by comparing the objective evaluation of the lens (1) with the specific quality criteria (18).

2. Method for evaluating optical lenses (1) according to claim 1, characterized as described above, that the lens (1) is trained with regard to its specific quality by means of the lens evaluation device (14) to determine whether the lens (1) is recognized as a good part (1a) or as a bad part (1b), for which As a result of the training, a pass / fail threshold is set in the machine learning module (19).

3. Method for evaluating optical lenses (1) according to claim 1 or 2, characterized by that the machine learning module (19) is trained with artificial intelligence and / or neural networks, so that the machine learning module (19) is trained in conjunction with the decision database (17) to output the good / bad threshold whether the lens (1) is a good part (1a) or a bad part (1b).

4. Method for evaluating optical lenses (1) according to claim 3, characterized in that, that the training in the machine learning module (19) is carried out with the artificial intelligence and / or the neural networks using statistical methods designed to eliminate a subjective factor of evaluation by at least one person (I, II) in the available data, so that the data, freed from the subjective factor, reflect an objective quality standard of the customer (A, B).

5. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that the training of the decision-making process using the lens evaluation device (14) includes at least the step of forming data pairs from the defects in the lens image (13a or 13b) and the expected evaluation (24) of the lens (1), wherein the defects are supplied to the machine learning module (19) as input information are calculated, and / or a mathematical comparison function (V2) is defined between the calculated decision and the expected decision, the minimum of which corresponds to an optimal match between the second machine learning module (19) and the expected decisions.

6. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that the decisions of several persons (I, II) are unified into a higher-level decision using artificial intelligence and / or neural networks, so that the higher-level decision is compared with the specific quality criteria (18) from the decision database (17).

7. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that lenses (1) which are rated with greater variance by several persons (I, II) receive a lower or no weighting when training the decision model (19a) in the machine learning module (19) and lenses (1) which are rated equally by several or all persons (I, II) receive a higher weighting when training the decision model (19a) in the machine learning module (19).

8. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that a defect detection device (12) is set up, to which test images (11) are transferred via the lenses (1) from the image acquisition device (10), wherein the defect detection device (12) detects defects on the lens (1) and outputs a lens image (13a) comprising the at least one detected and classified defect to the lens evaluation device (14).

9. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that a defect database (15) is set up in which different defects are stored, wherein the defect detection device (12) has a first, upstream machine learning module (16) that performs a segmentation, wherein the defect in the test image (11) is recognized and classified based on the defects stored in the defect database (15) and with which the lens image (13a) comprising the at least one recognized and classified defect is output.

10. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that the detected and classified defect in the lens image (13a) is stored as information in the lens image (13a) by means of the first, upstream machine learning module (16), in particular pixel-accurately, preferably comprising a summary of the defect information about the type, location, intensity, number and / or size such as length and / or width and / or area of ​​the defect, wherein the summary of the defect information is combined with the The lens image (13b) is transferred to the lens evaluation device (14).

11. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that at least one or exactly one defect detection device (12) is set up centrally, with several lens evaluation devices (14) being set up decentrally and / or multiple times, to which customer-specific decision databases (17) are assigned.

12. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that based on the decision of the second machine learning module (19) as to whether the lens (1 ) is a good part (1a) or a bad part (1 b), testers (I, II) are trained.

13. Method for evaluating optical lenses (1) according to one of the preceding claims, characterized by that the training of the defect detection using the defect detection device (12) is carried out with a comparison between the expected lens image (15c) and a calculated lens image (15b) from the first machine learning module (16), in which a first comparison function (V1 ) between the expected lens image (15c) and the calculated lens image (15b) is calculated.

14. Device which implements a method for evaluating optical lenses (1) according to one of the preceding claims.

Images