Training method for entrance system classification

Abstract

A computer-implemented training method (100) for entrance system classification, the method (100) comprising: receiving (110), from a client device (310), a set of imagery (60) including at least two different portions of an entrance system (10); receiving (120), from the client device (310), user-inputted classification data (62), the user-inputted classification data (62) pertaining to the entrance system (10) depicted in the set of imagery (60); and training (130) a machine learning model (64) based on the user-inputted classification data (62) and the set of imagery (60).

Description

[0001] TRAINING METHOD FOR ENTRANCE SYSTEM CLASSIFICATION

[0002] TECHNICAL FIELD

[0003] The present invention generally relates to the field of entrance systems. More specifically, the present invention relates to a computer-implemented training method for entrance system classification. The present invention also relates to an associated backend computing service, a computer program product and a non-transitory computer readable storage medium.

[0004] BACKGROUND

[0005] Training a model to perform classification of entrance systems requires addressing technical challenges posed by the wide variety of door types and configurations. Entrance systems, such as automatic overhead sectional doors, sliding doors, swing doors, and revolving doors, each exhibit unique mechanical and operational characteristics. The lack of industry-wide standardization further complicates the task, necessitating a comprehensive understanding of the distinct specifications associated with each system type.

[0006] A robust machine learning model must be trained to accurately discern these differences to prevent the misclassification of entrance systems. Misclassification can result in the ordering of incorrect replacement parts, leading to delays, increased costs, and customer dissatisfaction. More critically, the installation of incompatible components can jeopardize the functionality, safety, and security of the systems, posing serious risks to users and surrounding property. Therefore, the training process must be designed to ensure reliability and precision in classifying diverse entrance systems.

[0007] It is in view of the above-identified concerns and others that the present inventors have realized that there is room for improvements in this field.

[0008] SUMMARY

[0009] The challenges outlined in the background section, such as the diverse configurations of entrance systems and the absence of industry-wide standardization, highlight the complexity involved in developing a reliable classification system for these systems. This complexity, coupled with the risk of ordering incorrect replacement parts, shows that a systematic approach to training machine learning models for entrance system classification is desired.

[0010] The present inventors have recognized that advancements in machine learning technology can be utilized to overcome these challenges by creating a specialized training process. By focusing on training a machine learning model with imagery of a present installation, in combination with user-inputted classification data indicating what certain parts may be classified as, it is possible to train the model for accurate future classifications. The machine learning model trained in this way reduces the risk of errors in part orders and contribute to enhancing the safety and functionality of entrance installations by ensuring component compatibility, ultimately leading to increased user satisfaction and reduced costs.

[0011] In a first aspect of this disclosure there is accordingly provided a computer- implemented training method for entrance system classification, the method comprising: receiving, from a client device, a set of imagery including at least two different portions of an entrance system; receiving, from the client device, user-inputted classification data, the user-inputted classification data pertaining to the entrance system depicted in the set of imagery; and training a machine learning model based on the user- inputted classification data and the set of imagery.

[0012] The first aspect provides advantages by utilizing a specialized training process that combines imagery of entrance systems with user-inputted classification data, enabling the machine learning model to reduce errors in part orders and enhance the safety and functionality of entrance installations, ultimately increasing user satisfaction and reducing costs.

[0013] In some embodiments, the user-inputted classification data comprises a classification indicator for at least some images in the set of imagery. A technical benefit may include enhancing the model’s accuracy by providing clear labels that guide its learning process.

[0014] In some embodiments, the user-inputted classification data includes one or more of an entrance system brand, an entrance system type, and a mechanical component type. A technical benefit may include allowing for more precise classification by incorporating specific and diverse contextual information. In some embodiments, the method further comprises obtaining an initial confirmation prompt from the client device including a request to carry out the training of the machine learning model. A technical benefit may include ensuring user validation and engagement, which enhances data accuracy and reliability.

[0015] In some embodiments, the method further comprises requesting, to the client device, additional imagery input in response to a detection of an image quality value in the set of imagery below a quality limit value. A technical benefit may include improving the overall quality of the training dataset, leading to better model performance.

[0016] In some embodiments, wherein the request includes an indication that the image quality value of at least some images of the set of imagery is below the quality limit value. A technical benefit may include providing targeted feedback to users, enabling them to improve image quality efficiently.

[0017] In some embodiments, wherein the request includes an indication which of one or more images of the set of imagery need recapturing. A technical benefit may include improving the recapture process by focusing efforts on specific images, saving time and resources.

[0018] In some embodiments, wherein the request includes an indication of one or more actions that should be carried out in order to recapture one or more additional images. A technical benefit may include guiding users to capture high-quality images, enhancing the training data's effectiveness.

[0019] In some embodiments, wherein the quality limit value is one or more of an image resolution, an image exposure, an image focus, a noise level, a contrast ratio, a color balance, an image sharpness, and a signal-to-noise ratio. A technical benefit may include ensuring comprehensive image quality assessment, leading to improved training inputs.

[0020] In some embodiments, the method further comprises preprocessing the set of imagery before training the machine learning model. A technical benefit may include standardizing and optimizing the imagery, which facilitates more efficient and accurate model learning.

[0021] In some embodiments, wherein the set of imagery includes a front panel portion of a movable door member of the entrance system, a rear panel portion of the movable door member, and a hinge portion of a hinge mechanism of the movable door member. A technical benefit may include selecting some of the most important portions of an entrance system that can provide a comprehensive view installation, enhancing classification accuracy and reducing processing time.

[0022] In a second aspect, a computer-implemented method of operating a client device for causing training of a machine learning model for entrance system classification is provided, comprising: obtaining, using the client device, a set of imagery including at least two different portions of the entrance system; submitting user-inputted classification data, the user-inputted classification data pertaining to the entrance system depicted in the set of imagery; and transmitting the set of imagery and the user-inputted classification data to a backend computing service, the backend computing service comprising a processor being configured to train the machine learning model based on the user-inputted classification data and the set of imagery.

[0023] In a third aspect, a backend computing service is provided, comprising a processor configured to receive, from a client device, a set of imagery including at least two different portions of an entrance system; receive, from the client device, user- inputted classification data, the user-inputted classification data pertaining to the entrance system depicted in the set of imagery; and train a machine learning model based on the user-inputted classification data and the set of imagery.

[0024] In a fourth aspect, a computer program product is provided, comprising computer code for performing the method according to the first aspect when the computer program code is executed by a processor of a backend computing service.

[0025] In a fifth aspect, a non-transitory computer readable storage medium is provided, having stored thereon a computer program comprising computer program code for performing the method according to the first aspect when the computer program code is executed by a processor of a backend computing service.

[0026] In a sixth aspect, a computer program product is provided, comprising computer code for performing the method according to the second aspect when the computer program code is executed by a processor of a client device.

[0027] In a seventh aspect, a non-transitory computer readable storage medium is provided, having stored thereon a computer program comprising computer program code for performing the method according to the second aspect when the computer program code is executed by a processor of a client device.

[0028] Embodiments of the invention are defined by the appended dependent claims and are further explained in the detailed description section as well as in the drawings.

[0029] It should be emphasized that the term “comprises / comprising” when used in this specification is taken to specify the presence of stated features, integers, steps, or components, but does not preclude the presence or addition of one or more other features, integers, steps, components, or groups thereof.

[0030] All terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a / an / the [element, device, component, means, step, etc.]" are to be interpreted openly as referring to at least one instance of the element, device, component, means, step, etc., unless explicitly stated otherwise.

[0031] The steps of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated.

[0032] BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Objects, features and advantages of embodiments of the invention will appear from the following detailed description, reference being made to the accompanying drawings.

[0034] FIG. 1 is a schematic illustration of an entrance system subject of training a machine learning model.

[0035] FIG. 2 is a schematic flowchart diagram depicting exemplary steps of a computer-implemented training method for entrance system classification.

[0036] FIG. 3 is a schematic flowchart diagram depicting exemplary steps of a computer-implemented method of operating a client device for causing training of a machine learning model for entrance system classification.

[0037] FIG. 4 is a flowchart diagram illustrating a computer-implemented training method for entrance system classification.

[0038] FIG. 5 is a flowchart diagram illustrating a computer-implemented method of operating a client device for causing training of a machine learning model for entrance system classification. FIG. 6 is a schematic illustration of a non-transitory computer-readable storage medium in one exemplary embodiment, capable of storing a computer program product.

[0039] DETAILED DESCRIPTION OF EMBODIMENTS

[0040] Embodiments of the invention will now be described with reference to the accompanying drawings. The invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and may convey the scope of the invention to those skilled in the art. The terminology used in the detailed description of the particular embodiments illustrated in the accompanying drawings is not intended to be limiting of the invention. In the drawings, like numbers refer to like elements

[0041] In the scenario depicted in FIG. 1, an authorized user 30 arrives at a location where an entrance system 10 is installed. The authorized user 30, which may be an expert in entrance systems, recognizes one or more parts of this entrance system 10 installation. This recognition allows the authorized user 30 to provide user-inputted classification data. Moreover, the authorized user 30 captures imagery of relevant parts of the entrance system 10 using a client device 310. The captured imagery, combined with the user’s knowledge, is submitted to a backend computing service 300. This collaborative data collection serves as the foundation for training a machine learning model. The imagery provides a visual representation of the entrance system, while the user’s input offers context, enhancing the quality of the imagery.

[0042] By integrating both visual data and expert knowledge, a machine learning model can be trained to recognize entrance system components accurately. This process ensures that the machine learning model can classify future, potentially unknown, entrance systems with greater precision. The scenario depicted ensures that the machine learning model leams from real-world examples of installations, improving its ability to adapt to new and varied entrance systems in practical applications.

[0043] In this example the entrance system 10 is an overhead door system, although other exemplary entrance systems may involve revolving door systems, swing door systems, sliding door systems, or the like. The entrance system 10 includes a movable door member 12 having a door leaf 11. The door leaf 11 includes several front panel portions 11-1 and a rear panel portion 11-2. For visualization purposes, the front and rear panel portions 11-1, 11-2 are shown as two separate units, although it should be understood that these are parts of the same door leaf 11 but on opposite sides thereof. In addition, a hinge mechanism 18 is shown (zoomed in at a hinge portion 18-1) attached to the movable door member 12.

[0044] The authorized user 30 is operating the client device 310. The authorized user 30 may be a service personnel, technician, operator, or the like. The client device 310 may be a smartphone, smart tablet, PDA, laptop, smartwatch, AR glasses, or the like. The client device 310 includes an image capturing device capable of capturing one or more images of the surroundings. The image capturing device may be integrated with the client device 310, or alternatively provided external to the client device 310 and capable of communicating the image(s) with the client device 310 using any known communication standards. The client device 310 is an interactive device, meaning that the authorized user 30 can interact with the client device 310 for instructing it to capture one or more images of the surroundings. The client device 310 comprises a display screen configured to present visual information to the authorized user 30. The client device 310 may be configured to provide other feedback, such as sound, audio, haptics, etc., through suitable devices integrated with or external to the client device 310.

[0045] The client device 310 is configured to communicate with a backend computing service 300, here depicted as a cloud. The communication may be effected using any known communication technologies known in the art, including but not limited to short- range communication interfaces such as IEEE 802.11, IEEE 802.15, ZigBee, WirelessHART, WiFi, Bluetooth®, BLE, RFID, WLAN, MQTT loT, CoAP, DDS, NFC, AMQP, LoRaWAN, Z-Wave, Sigfox, Thread, EnOcean, mesh communication, other forms of proximity-based device-to-device radio communication signal such as LTE Direct, or long-range communication interfaces such as W-CDMA / HSPA, GSM, UTRAN, LTE or Starlink.

[0046] The backend computing service 300 may be implemented using any commonly known cloud-computing platform technologies, such as e.g. Amazon Web Services, Google Cloud Platform, Microsoft Azure, DigitalOcean, Oracle Cloud Infrastructure, IBM Bluemix or Alibaba Cloud. The backend computing service 300 may be included in a distributed cloud network that is widely and publically available, or alternatively limited to an enterprise. In alternative implementations, the backend computing service 300 may be locally managed as e.g. a centralized server unit.

[0047] The backend computing service 300 may be configured to maintain a storage device, being included with or external to the backend computing service 300. Connection to cloud storage means may be established using DBaaS (Database-as-a- service). For instance, cloud storage means may be deployed as a SQL data model such as MySQL, PostgreSQL or Oracle RDBMS. Alternatively, deployments based on NoSQL data models such as MongoDB, Amazon DynamoDB, Hadoop or Apache Cassandra may be used. DBaaS technologies are typically included as a service in the associated cloud-computing platform.

[0048] Server configurations other than cloud-based may be realized in other examples, for instance based on any type of client-server or peer-to-peer (P2P) architecture. Server configurations may thus involve any combination of e.g. web servers, database servers, email servers, web proxy servers, DNS servers, FTP servers, file servers, DHCP servers, to name a few.

[0049] The present inventors have realized that the objectives described herein can be achieved by way of providing a set of imagery including at least two different portions of the entrance system 10. A “set” of imagery shall be understood as including one or more images. Therefore, in some examples one single image in the set of imagery may include a plurality of portions of the entrance system 10. For example, a picture taken from afar may show both a hinge portion 18-1 and a front panel portion 11-1 of the movable door member 12. In other examples, the set of imagery includes two or more images, each image in the set of imagery including at least one portion of the entrance system 10. A “portion” shall be understood as an entire component, such as a hinge, or at least parts thereof, such as an upper / lower hinge plate including one or more sets of screw holes or openings.

[0050] In the particular example of FIG. 1, the set of imagery will include three images, each image showing a respective portion of the entrance system 10. More specifically, the set of imagery includes a first image of a front panel portion 11-1, a second image of a rear panel portion 11-2, and a third image of a hinge portion 18-1. While this is just an exemplary set of imagery, the present inventors have conducted studies that indicate that these three particular portions in some cases may be sufficient for classifying the entrance system 10 using approaches taught herein. Also, any number of images can be captured in any order. This is due to these three portions oftentimes being associated with unique appearances, or at least a distinctive appearance that is associated with fewer possible alternatives compared to other portions of the entrance system. In particular, these three portions in combination typically enable training of a machine learning model such that it will become capable of classifying most entrance systems currently on the market.

[0051] For the panel portions 11-1, 11-2, this distinctive appearance can relate to patterns, colors, textures, carvings / engravings, glass inserts, hardware, number of panels, shapes, material contrast, and the like.

[0052] For the hinge portion 18-1, this distinctive appearance can relate to designs, shapes, finishes, coatings, materials, number of holes / plates, type of holes / plates, hole arrangements, plate configurations, sharp edges / round edges, ball bearings, and the like.

[0053] In other examples, similar or alternative unique or at least distinctive appearances can be envisaged for any of the other units / components / modules / etc., for which imagery can be captured. Such other units / components / modules / etc., may include automatic door operators, linkages, transmission types, sensory modules, electronic control units, handles, knobs, locksets, thresholds, windows, lights, door surroundings, architectural moldings, tracks, ventilation openings, emergency exit features, weather stripping and seals, and the like, each having a set involving one of more distinctive features.

[0054] In FIG. 2, a client device 310 is shown, operating in frontend contexts. The client device 310 is configured to provide data to a machine learning model for training purposes. The process of providing this data according to the teachings herein shall be understood as a continued and / or guided human-machine interaction process. The user operating the client device 310, such as the authorized user 30 as discussed above, will be assisted in performing the technical task of finding accurate replacement options for parts of the entrance system.

[0055] This particular computer program is named “ModelTrainer”, and includes a plurality of steps leading to the training of the machine learning model. In this example the ModelTrainer program involves an initial step, followed by a classification data submission step, followed by a three-step continued and / or guided human-machine interaction process, followed by a computer backend training of the machine learning model. Other examples may include any number of steps in the guided human-machine interaction process, in any suitable order, as this may generally be based upon how many images that are included in the set of imagery. The classification may also be realized in another order, such as between any image capture or after the capture of a final image.

[0056] Before the guided interaction process is initiated, the authorized user 30 may prompt to confirm that the entrance system 10 indeed is recognized, and that the recognized entrance system 10 can be used as a subject for training the machine learning model. Hence, this initial confirmation prompt includes a request to carry out the training of the machine learning model. The initial confirmation prompt may be submitted via the client device 310 to the backend computing resource 300, optionally upon request by the backend computing resource 300.

[0057] In this example the authorized user 30 clicks a “YES” button which is a graphical design element, whereby the procedure continues to the step of submitting classification data. There may be other suitable ways of providing the initial confirmation prompt, such as using a voice command interface that allows the user to verbally confirm the desire of training, deploying a separate physical device with a dedicated button for initiating the process, utilizing a touchscreen interface where users can swipe or tap to provide confirmation, integrating a gesture-based system that recognizes specific hand movements to proceed, employing a mobile app notification that prompts users to confirm through their smartphones, or implementing a QR code system where scanning the code on-site triggers the confirmation prompt.

[0058] The next step is for the user to submit classification data. The classification data is user-inputted, meaning that the authorized user 30 controls which information that is to be submitted as context for the training of the machine learning model. This user- inputted classification data can include various aspects such as the entrance system brand and type, and component type.

[0059] The user-inputted classification data may be submitted via one or more of a graphical user interface, optionally with dropdown menus, a touchscreen with virtual keyboards, voice recognition software, a barcode or QR code scanner, form-based input with text fields, a mobile app with templates, stylus or pen input, an NFC reader, a camera interface for text capture, an external keyboard for manual entry, or the like.

[0060] For the entrance system brand, the user might input names like “Brand A”, “Brand B” or “Brand C”, which are well-known manufacturers of entrance systems. This information helps the model learn brand-specific characteristics, which can be important for accurate classification.

[0061] Regarding the entrance system type, the user could specify whether the system is an “automatic overhead sectional door”, a “sliding door”, a “swing door” or a “revolving door”. Each type has unique operational features and components, and this data aids the machine learning model in distinguishing between them during training.

[0062] For the component type, the user might provide details such as “Model X hinge”, “Type Y motor”, or “Z-series control panel”. By mapping these components to specific models and types, the machine learning model can leam to recognize intricate details that differentiate one component from another.

[0063] While this information is in this example submitted before the capture of the imagery, in optional embodiments this data can be entered immediately after capturing each image, allowing the user to map the classification data to specific images. For instance, after taking a picture of a hinge, the user might input that “this hinge is of Model X”, thereby directly associating the image with its classification data.

[0064] The submission of detailed classification data can provide several technical advantages. By associating specific data like brand, type, and component model with each image, the machine learning model gains a richer contextual understanding. This context reduces ambiguity, allowing the machine learning model to make more accurate classifications. Moreover, providing specific details narrows the range of potential classifications the machine learning model must consider. Computational resources may accordingly be used more effectively, reducing the need for trial-and-error methods to save processing power. In addition, thanks to the data mapping, the machine learning model can focus on learning the distinguishing features of specific components or brands, leading to higher accuracy in recognizing similar patterns in new data. This allows the machine learning model to generalize more effectively to classify unseen entrance systems, even when they exhibit slight variations.. Yet additionally, the structured input allows for streamlined data processing, as the machine learning model can bypass irrelevant options and concentrate on pertinent classifications, enhancing processing efficiency. The above leads to the reduction in options and focused learning to enable faster convergence during training, decreasing the time required to achieve desirable model performance.

[0065] Now that the classification data has been inputted by the authorized user 30, the process then continues to the first step of the guided interaction process. Here, the client device 310 instructs the authorized user to take a picture of a front panel portion 11-1 of the entrance system 10, more specifically of the door leaf 11 of the movable door member 12. The client device 310 may continuously guide the user into capturing a sufficiently accurate representation of the front panel portion 11-1. The client device 310 may identify any potential issue with the captured image, such as poor lighting, wrong focus, too zoomed in or zoomed out, blurry view, and the like. The client device 310 may request that the authorized user 30 captures one or more additional images until the quality is sufficient for subsequent analysis, or from different perspectives, or from different distances to the front panel portion 11-1, etc.

[0066] In response to the client device 310 accepting the captured image(ry) with respect to one or more of the above considerations, a second step of the guided interaction process involves capturing a picture of a rear panel portion 11-2. This may involve similar considerations as for the first step.

[0067] In response to the client device 310 accepting the captured image(ry) with respect to one or more of the above considerations, a third step of the guided interaction process involves capturing a picture of a hinge portion 18-1. This may involve similar considerations as for the first and / or second steps.

[0068] After the completion of the guided interaction process, the remaining steps are functionally carried out by the backend computing service 300. The machine learning model is accordingly trained based on the provided imagery and submitted user-inputted classification data.

[0069] In this case, the machine learning model is successfully trained based on the provided data. In other examples the backend computing resource 300 may request additional imagery input. This may be received after the reception of all images, or as intermittent steps following the reception of one image. The client device 310 may receive a prompt from the backend computing resource 300, indicating one or more of the following: (i) that an image quality value of one or more of the captured images is below a quality limit value, (ii) which image(s) need / needs to be recaptured, and (iv) what action(s) the authorized user 30 should carry out in order to recapture additional image(s).

[0070] The indication in item (i) allows the authorized user 30 to be aware of any potential issues in the provided set of imagery 60, and that the quality thereof is not above the quality limit value. Low-quality images can lead to poor model performance, as they might not capture necessary details of the entrance system 10. By ensuring that all images meet a certain standard, the training process can become more robust and reliable, ultimately leading to a more effective model.

[0071] The indication in item (ii) allows the authorized user 30 to focus on improving specific images rather than redoing the entire set. For instance, if only an image of the hinge is blurry, the user can concentrate on recapturing that particular image. This selective approach saves time and resources while ensuring that the relevant components of the entrance system 10 are accurately represented in the dataset.

[0072] The indication in item (iii) provides guidance on how the image(s) need to be recaptured. This guidance could include instructions on using different camera settings, employing supplementary lighting, using stabilization tools to reduce motion blur, adjusting lighting, changing angles, or improving focus. Such detailed directions help the user capture images that better meet the requirements of model performance. By offering specific recapture instructions, the backend computing resource 300 may ensures that the images are not only retaken but improved, facilitating more effective training.

[0073] The above indications may be given as a prompt to the client device 310 upon the backend computing resource 300 detecting that an image quality value of the obtained imagery is below a quality limit value.

[0074] The image quality value may encompass various metrics that evaluate the clarity and usability of an image for training the machine learning model. These metrics may include one or more of an image resolution, an image exposure, an image focus, a noise level, a contrast ratio, a color balance, an image sharpness, and a signal-to-noise ratio. The quality limit value serves as a threshold that defines the minimum acceptable standards for an image quality value. The quality limit value may be predetermined, based on typical requirements for effective training, or adaptable, adjusting dynamically to specific needs during training sessions.

[0075] The backend computing resource 300 may detect whether an image quality value of an image meets the quality limit value through applying one or more image processing techniques. For instance, it can evaluate resolution by counting pixels, assess focus by analyzing edge sharpness, evaluate exposure to identify overexposure or underexposure, analyze noise levels to detect random color or brightness variations, and measure contrast by examining the dynamic range between dark and light areas. The focus might be assessed through edge detection, while exposure could be measured through histogram analysis. Other considerations may be taken into account for the other values. If any quality limit value falls below the quality limit value, the client device 310 is prompted based on any of the indications (i) to (iii) discussed above, ensuring all images used fortraining meet the required quality standards.

[0076] FIG. 3 shows the client device 310 operating in frontend contexts. The client device 310 transmits a set of imagery 60 and user-inputted classification data 62 to the backend computing service 300, operating in backend contexts.

[0077] The backend computing service 300 receives this data and provides it as training input to a machine learning model 64. A preprocessing module 64-1 may be configured to carry out preprocessing of at least some images in the set of imagery 60, with the aim of enhancing the quality and consistency of the images, ensuring they are suitable for training. The preprocessing may involve tasks such as resizing images to a standard dimension, adjusting contrast and brightness, reducing noise, and normalizing pixel values to fall within a specific range. These steps assist in standardizing the training input, allowing the machine learning model 64 to focus on learning relevant features without being affected by variations in image quality. Once preprocessing is complete, the refined data is fed into the machine learning model 64.

[0078] An encoding module 64-2 may be configured to encode the user-inputted classification data 62 into a machine-readable format, such as numeric integers or one- hot vectors, allowing the machine learning model 64 to interpret them during training.

[0079] A training module 63-3 is configured to train the machine learning model 64, incorporating both a feature extraction module 64-4 and a context mapping module 64- 5. The feature extraction module 64-4 is responsible for identifying and extracting visual patterns from the set of imagery 60, such as edges, textures, and shapes, which are critical for recognizing entrance system components.

[0080] The context mapping module 64-5 associates these extracted features with the semantic meaning conveyed by the user-inputted classification data 62. This data includes classification indicators that annotate the imagery with labels, providing contextual information about the entrance system 10. For example, an image of a sliding door might be labeled as “sliding door” for the type, “brand A” for the manufacturer, and “model X hinge” for a specific hinge. These labels can be structured either hierarchically or as independent classes.

[0081] The training module 63-3, which can be configured as a convolutional neural network (CNN), processes the images through convolutional layers to detect spatial patterns. The model’s predictions are compared to true labels using a loss function, such as categorical cross-entropy, to quantify prediction errors. Backpropagation adjusts neuron weights, and an optimization algorithm like stochastic gradient descent updates these weights to minimize error over successive iterations.

[0082] Finally, a postprocessing module 64-6 may be configured to refine the outputs of the training module 64-3 to enhance their usability and accuracy. After the training module 64-3 completes its task, the postprocessing module 64-6 interprets the raw model outputs into meaningful classifications or predictions that are easier for users to understand. It may apply additional rules or heuristics to correct any apparent errors in the predictions of the machine learning model 64. Furthermore, the postprocessing module 64-6 might assign confidence scores to each prediction, helping users assess the reliability of the results. The postprocessing module 64-6 may also format the outputs into standardized reports or visualizations, such as charts or tables, for easier analysis and decision-making. Additionally, the postprocessing module 64-6 facilitates integration with other systems or databases, ensuring that results are seamlessly incorporated into broader workflows or applications.

[0083] In summary, the machine learning model 64 ensures that image features of the set of imagery 60 are associated with context based on the user-inputted classification data 62. The trained machine learning model 64 can in subsequent operations be used to identify and categorize unseen entrance systems, helping streamline classification tasks and improve decision-making in practical applications.

[0084] FIG. 4 shows a computer-implemented method 100 for entrance system classification. The method 100 is implemented by the backend computing service 300. The method 100 comprises a step 110 of receiving, from the client device 310, a set of imagery 60 including at least two different portions of the entrance system 10. The method 100 further comprises a step 120 of receiving, from the client device 310, user- inputted classification data 62, the user-inputted classification data 62 pertaining to the entrance system 10 depicted in the set of imagery 60. The method 100 further comprises a step 130 of training a machine learning model 64 based on the user-inputted classification data 62 and the set of imagery 60.

[0085] FIG. 5 shows a computer-implemented method 200 of operating the client device 310 for causing training of a machine learning model 64 for entrance system classification. The method 200 comprises a step 210 of obtaining, using the client device 310, a set of imagery 60 including at least two different portions of the entrance system 10. The method 200 comprises a step 220 of submitting user-inputted classification data 62, the user-inputted classification data 62 pertaining to the entrance system 10 depicted in the set of imagery 60. The method 200 comprises a step 230 of transmitting the set of imagery 60 and the user-inputted classification data 62 to a backend computing service 300. The backend computing service 300 comprises a processor being configured to train the machine learning model 64 based on the user- inputted classification data 62 and the set of imagery 60.

[0086] FIG. 6 is a schematic illustration of a non-transitory computer-readable storage medium 400 in one exemplary embodiment, capable of storing a computer program product 410, comprising computer program code for performing either the method 100 or the method 200 as described herein. The non-transitory computer-readable storage medium 400 in the disclosed embodiment is a memory stick, such as a Universal Serial Bus (USB) stick; the non-transitory computer-readable storage medium 400 may however be embodied in various other ways instead, as is well known per se to the skilled person. The USB stick 400 comprises a housing 430 having an interface, such as a connector 440, and a memory chip 420. In the disclosed embodiment, the memory chip 420 is a flash memory, i.e. a non-volatile data storage that can be electrically erased and re-programmed.

[0087] The memory chip 420 stores the computer program product 410 which is programmed with computer program code (instructions) that when loaded into and executed by a processing device, such as a CPU, will perform either the method 100 or the method 200 as described herein, and optionally any of their respective embodiments or examples as described herein. The USB stick 400 is arranged to be connected to and read by a reading device for loading the instructions into the processing device. The processing device may, for instance, be comprised in the client device 310 or in the backend computing service 300.

[0088] It should be noted that a computer-readable medium can also be other mediums such as compact discs, digital video discs, hard drives or other memory technologies commonly used. The computer program code (instructions) can also be downloaded from the computer-readable medium via a wireless interface to be loaded into the processing device.

[0089] The invention has been described above in detail with reference to embodiments thereof. However, as is readily understood by those skilled in the art, other embodiments are equally possible within the scope of the present invention, as defined by the appended claims.

Claims

CLAIMS1. A computer-implemented training method (100) for entrance system classification, the method (100) comprising: receiving (110), from a client device (310), a set of imagery (60) including at least two different portions of an entrance system (10); receiving (120), from the client device (310), user-inputted classification data (62), the user-inputted classification data (62) pertaining to the entrance system (10) depicted in the set of imagery (60); and training (130) a machine learning model (64) based on the user-inputted classification data (62) and the set of imagery (60).

2. The method (100) of claim 1, wherein the user-inputted classification data comprises a classification indicator for at least some images in the set of imagery (60).

3. The method (100) of any of claims 1-2, wherein the user-inputted classification data (62) includes one or more of an entrance system brand, an entrance system type, and a mechanical component type.

4. The method (100) of any preceding claim, further comprising obtaining an initial confirmation prompt from the client device (310) including a request to carry out the training of the machine learning model (64).

5. The method (100) of any preceding claim, further comprising requesting, to the client device (310), additional imagery input in response to a detection of an image quality value in the set of imagery (60) below a quality limit value.

6. The method (100) of claim 5, wherein the request includes an indication that the image quality value of at least some images of the set of imagery (60) is below the quality limit value.

7. The method (100) of any of claims 5-6, wherein the request includes an indication which of one or more images of the set of imagery (60) that need recapturing.

8. The method (100) of any of claims 5-7, wherein the request includes an indication of one or more actions that should be carried out in order to recapture one or more additional images.

9. The method (100) of any of claims 5-8, wherein the quality limit value is one or more of an image resolution, an image exposure, an image focus, a noise level, a contrast ratio, a color balance, an image sharpness, and a signal-to-noise ratio.

10. The method (100) of any preceding claim, further comprising preprocessing the set of imagery (60) before training the machine learning model (64).

11. The method (100) of any preceding claim, wherein the set of imagery (60) includes a front panel portion (11-1) of a movable door member (12) of the entrance system (10), a rear panel portion (11-2) of the movable door member (12), and a hinge portion (18-1) of a hinge mechanism (18) of the movable door member (16).

12. A computer-implemented method (200) of operating a client device (310) for causing training of a machine learning model (64) for entrance system classification, comprising: obtaining (210), using the client device (310), a set of imagery (60) including at least two different portions of the entrance system (10); submitting (220) user-inputted classification data (62), the user-inputted classification data (62) pertaining to the entrance system (10) depicted in the set of imagery (60); and transmitting (230) the set of imagery (60) and the user-inputted classification data (62) to a backend computing service (300), the backend computing service (300) comprising a processor being configured to train the machine learning model (64) based on the user-inputted classification data (62) and the set of imagery (60).

13. A backend computing service (300) comprising a processor configured to: receive, from a client device (310), a set of imagery (60) including at least two different portions of an entrance system (10); receive, from the client device (310), user-inputted classification data (62), the user-inputted classification data (62) pertaining to the entrance system (10) depicted in the set of imagery (60); and train a machine learning model (64) based on the user-inputted classification data (62) and the set of imagery (60).

14. A computer program product comprising computer code for performing the method (100) according to claim 1 when the computer program code is executed by a processor of a backend computing service (300).

15. A non-transitory computer readable storage medium having stored thereon a computer program comprising computer program code for performing the method (100) according to claim 1 when the computer program code is executed by a processor of a backend computing service (300).

16. A computer program product comprising computer code for performing the method (100) according to claim 12 when the computer program code is executed by a processor of a client device (310).

17. A non-transitory computer readable storage medium having stored thereon a computer program comprising computer program code for performing the method (100) according to claim 12 when the computer program code is executed by a processor of a client device (310).

Images