Visual sorting method
By setting recognition patterns on the raw material strip and building a deep learning model, the problem of low accuracy in visual inspection of die-cut parts was solved, and efficient identification of uncommon defects and output of qualified parts were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN LLMACHINECO LTD
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-26
AI Technical Summary
In the existing technology, the visual inspection accuracy of die-cut parts is low, making it difficult to identify uncommon defects, resulting in the output of unqualified parts.
By setting recognition patterns on the raw material strip, image data is acquired using a vision module to form datasets of good and defective products. A deep learning model is then built to expand the defective product dataset and improve the accuracy of visual inspection of die-cut parts.
It improves the accuracy of visual inspection of die-cut parts, reduces the risk of identifying uncommon defects, and ensures that the die-cutting production line outputs qualified parts.
Smart Images
Figure CN122076731A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of die-cutting production technology, and in particular to a visual sorting method. Background Technology
[0002] Die-cut components are found in portable products such as mobile phones, headphones, and smartwatches. As technology advances, the compactness of these products continues to increase, leading to higher demands on the dimensional accuracy and surface quality of these die-cut components. Die-cutting production lines produce these components. Visual inspection of these components is typically performed on the production line to identify those that fail to meet dimensional or surface quality standards, thus reducing the risk of outputting defective parts. However, current technology struggles to identify uncommon defects in die-cut components, resulting in low accuracy of visual inspection and consequently, the output of defective parts from the die-cutting production line. Summary of the Invention
[0003] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a visual sorting method that can improve the accuracy of visual inspection of die-cut parts.
[0004] This invention provides a visual sorting method, comprising: providing a raw material strip; acquiring the material information of the raw material strip and determining the optical properties of the raw material strip based on the material information; setting a recognition pattern on the raw material strip, the graphic features of the recognition pattern being determined by the optical properties, the graphic features including shape and color; die-cutting the raw material strip to form a die-cut component, and giving the die-cut component the recognition pattern; providing a vision module to acquire image data of the die-cut component; collecting image data, determining the position of the recognition pattern in the image data, and labeling the image data to form a good product dataset and a defective product dataset; expanding the defective product dataset using a generative adversarial network; constructing a deep learning model using the good product dataset and the defective product dataset, and using the deep learning model to recognize the image data to determine whether the die-cut component is qualified.
[0005] The visual sorting method provided by the embodiments of the present invention has at least the following beneficial effects:
[0006] Selecting graphic features for recognition based on the optical properties of the raw material strip makes the relationship between the recognized graphic in the image data acquired by the vision module and the graphic of the die-cut part clearer. This can improve the quality of the defective dataset expanded by generative adversarial networks. By constructing a deep learning model using the expanded defective dataset and recognizing image data through the deep learning model, the risk of difficulty in recognizing uncommon defects on the die-cut parts can be reduced, thereby improving the accuracy of visual inspection of the die-cut parts.
[0007] In one embodiment of this implementation, an identification pattern is provided on the raw material strip, including: Part of the raw material strip is cut off, and through holes are formed on the raw material strip. The through holes form an identification pattern.
[0008] In one embodiment of this implementation, the die-cutting raw material strip includes: A background element is provided, the color of which is determined by optical properties. A through-hole is positioned opposite the background element so that the color of the background element fills the identification pattern.
[0009] In one embodiment of this implementation, the die-cutting raw material strip includes: The raw material is die-cut into multiple die-cut parts, and each die-cut part has a recognition pattern. The image data of the multiple die-cut parts is acquired by a vision module, and a recognition module is provided to identify the relative position and relative angle of the recognition patterns on two adjacent die-cut parts in order to obtain the position deviation data of the die-cut parts.
[0010] In one embodiment of this implementation, a method for constructing a deep learning model using a dataset of good products and a dataset of defective products is provided, including: The assembly device and mating parts are provided. Position deviation data is sent to the assembly device to assemble the die-cut parts and mating parts, and the assembly device adjusts the relative positions of the die-cut parts and mating parts according to the position deviation data.
[0011] In one embodiment of this implementation, a visual module is provided, including: The vision module includes a recognition camera and a recognition lens, and selects the resolution of the recognition camera and the type of recognition lens based on the graphic features.
[0012] In one embodiment of this implementation, a visual module is provided, including: A lighting module is provided, and the lighting parameters of the lighting module are selected according to the graphic features. The lighting parameters include the light source brightness, light source shape and light source color. The lighting module illuminates the die-cut parts. An evaluation scheme is formulated. The accuracy of the recognition module in recognizing image data is evaluated through the evaluation scheme, and the exposure time and lighting parameters of the recognition camera are adjusted.
[0013] In one embodiment of this implementation, the die-cutting raw material strip includes: A base film is provided, and raw materials are die-cut to form die-cut parts on the base film. The base film is transparent, and image data of the die-cut parts on the base film is obtained through a vision module.
[0014] In one embodiment of this implementation, a visual module is provided, including: The vision module includes two recognition cameras, which acquire image data of the die-cut parts from opposite sides of the base film.
[0015] In one embodiment of this implementation, a deep learning model is constructed using a dataset of good products and a dataset of defective products, including: The raw material is die-cut into multiple die-cut parts, and these parts are arranged in a queue. A transfer device is provided to remove unqualified die-cut parts from the queue, creating vacancies in the queue, and then qualified die-cut parts are moved into the vacancies.
[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments, wherein: Figure 1 This is a schematic diagram of a visual sorting method according to one embodiment of the present invention. Detailed Implementation
[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0019] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.
[0020] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.
[0021] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.
[0022] In the description of this invention, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0023] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
[0024] Please see Figure 1 , Figure 1 This is a schematic diagram of a visual sorting method according to one embodiment of the present invention. The present invention provides a visual sorting method, which includes: Step S100: Provide raw material strip, obtain material information of raw material strip, and determine optical properties of raw material strip based on material information.
[0025] Step S200: Set an identification pattern on the raw material strip. The graphic features of the identification pattern are determined by optical properties, including shape and color.
[0026] Specifically, identification patterns are formed on the raw material belt by die-cutting using methods such as inkjet printing or embossing. The shape of the identification image can be round, cross-shaped, or L-shaped. The raw material belt extends along the conveying direction, and multiple die-cut patterns are set on the raw material belt and arranged at intervals along the conveying direction.
[0027] Step S300: Die-cut the raw material strip to form a die-cut part, and give the die-cut part an identification pattern.
[0028] Specifically, a base film is provided, and the portion of the raw material strip corresponding to the identification pattern is die-cut to die-cut multiple portions of the raw material strip corresponding to the identification pattern onto the base film, so as to form multiple die-cut parts arranged along the conveying direction on the base film.
[0029] Step S400: Provide a vision module to acquire image data of the die-cut parts.
[0030] Specifically, the base film is conveyed along the conveying direction, and multiple die-cutting components on the base film pass through the vision module in sequence, so that the vision module acquires image data of multiple die-cutting components in sequence.
[0031] Step S500: Collect image data, determine the position of the recognized graphics in the image data, and label the image data to form a good product dataset and a defective product dataset. Expand the defective product dataset by generating an adversarial network.
[0032] Specifically, the outlines of the recognition graphics and die-cut parts in the image data are manually identified, and the relative positions of the recognition graphics and die-cut parts are judged to determine whether the die-cut parts corresponding to the image data are defective. The image data corresponding to the defective parts are labeled manually or semi-automatically to distinguish multiple image data into good product datasets and defective product datasets. Defective product image data is synthesized by generative adversarial network, and the synthesized defective product image data is summarized into the defective product dataset to expand the defective product dataset.
[0033] Step S600: Construct a deep learning model using the good product dataset and the defective product dataset, and use the deep learning model to identify image data in order to determine whether the die-cut parts are qualified.
[0034] Specifically, the deep learning model is trained using both good and defective product datasets. Synthetic and original image data from the defective product dataset are used together in the training, increasing the number of uncommon defect images in the training dataset and thus improving the model's accuracy in identifying uncommon defects in die-cut parts. Image data acquired by the vision module is input into the deep learning model, which then performs real-time identification of multiple die-cut parts passing through the vision module to determine their quality on the base film.
[0035] The visual sorting method of the present invention selects the graphic features of the identification pattern based on the optical characteristics of the raw material strip, which makes the relationship between the identification pattern in the image data acquired by the vision module and the graphic of the die-cut part clearer. This can improve the quality of the defective product dataset expanded by generative adversarial network. The deep learning model is constructed by the expanded defective product dataset, and the image data is identified by the deep learning model. This can reduce the risk of difficulty in identifying uncommon defects on the die-cut parts, thereby improving the accuracy of visual inspection of the die-cut parts.
[0036] Please see Figure 1 In one embodiment of this implementation, an identification pattern is provided on the raw material strip, including; Part of the raw material strip is cut off, and through holes are formed on the raw material strip. The through holes form an identification pattern.
[0037] Specifically, the shape of the through hole is the shape of the recognition image. The through hole can be cross-shaped, L-shaped, or round. The raw material strip is punched with a die to form the recognition image on the raw material strip.
[0038] Understandably, on the one hand, the surface of the raw material strip may be relatively smooth. Under such conditions, when forming recognition images on the surface of the raw material strip through methods such as embossing and inkjet printing, specular reflection may occur, making the recognition patterns in the image data acquired by the vision module difficult to identify. This can lead to inaccurate annotation of the image data, thereby reducing the quality of the image data synthesized by the generative adversarial network. Cutting through holes into the surface of the raw material strip reduces the risk of specular reflection causing the recognition patterns in the image data to be difficult to identify. On the other hand, the dimensional and positional accuracy of recognition patterns formed by inkjet printing and embossing is relatively low. Using through holes formed by cutting as recognition patterns can reduce the impact of dimensional and positional deviations of the recognition patterns themselves on the dimensional judgment of the die-cut parts. This is beneficial to improving the quality of image data in the defective product dataset and also to improving the accuracy of deep learning models in recognizing image data.
[0039] Please see Figure 1 In one embodiment of this implementation, the die-cutting raw material strip includes: A background element is provided, the color of which is determined by optical properties. A through-hole is positioned opposite the background element so that the color of the background element fills the identification pattern.
[0040] Specifically, the background component can be a background board or a background cloth, with the background component facing the through hole on the die-cutting component. Image data is obtained from the side of the die-cutting component away from the background component through the vision module, so that the color of the background component in the image data fills the through hole.
[0041] It is understandable that by setting a background element and filling the background element with the color of the recognition graphic, the contrast between the recognition graphic and other parts of the die-cut component in the image data can be increased, thereby improving the clarity of the recognition graphic in the image data and thus improving the accuracy of deep learning models in recognizing image data.
[0042] Please see Figure 1 In one embodiment of this implementation, the die-cutting raw material strip includes: The raw material is die-cut into multiple die-cut parts, and each die-cut part has a recognition pattern. The image data of the multiple die-cut parts is acquired by a vision module, and a recognition module is provided to identify the relative position and relative angle of the recognition patterns on two adjacent die-cut parts in order to obtain the position deviation data of the die-cut parts.
[0043] Specifically, multiple sections of the raw material strip distributed along the conveying direction are punched using a die to form through holes spaced apart along the conveying direction on the raw material strip. The multiple sections with through holes on the raw material strip are then die-cut to form multiple die-cut parts with through holes on the base film. The multiple die-cut parts are arranged along the conveying direction on the base film. Image data of two adjacent die-cut parts are acquired by a vision module, and the relative position of the recognition pattern of the two die-cut parts in the image data is identified by a recognition module to obtain the deviation data of the relative position and relative angle of the two adjacent die-cut parts.
[0044] It is understandable that the raw material strip and the base film may have a certain light transmittance, which may cause the outline of the die-cut parts in the image data to be relatively blurry. By recognizing the recognition pattern on the die-cut parts of the recognition module, the relative position and relative angle deviation data of two adjacent die-cut parts can be obtained, which can improve the accuracy of the obtained deviation data.
[0045] Please see Figure 1 In one embodiment of this implementation, a method for constructing a deep learning model using a dataset of good products and a dataset of defective products is provided, including: The assembly device and mating parts are provided. Position deviation data is sent to the assembly device to assemble the die-cut parts and mating parts, and the assembly device adjusts the relative positions of the die-cut parts and mating parts according to the position deviation data.
[0046] Specifically, the assembly device includes a robotic arm, with the mating parts in sheet form. The die-cut parts are conveyed through a bottom film so that multiple die-cut parts pass through the assembly device in sequence. Position deviation data is sent to the assembly device through an identification module.
[0047] Understandably, under certain conditions, it is necessary to assemble die-cut parts and mating parts. The assembly device can transfer the mating part to a preset position and place it over the die-cut part at that position to complete the assembly. If there is a deviation in the relative position and angle between two adjacent die-cut parts, when the die-cut part is conveyed through the assembly device, there will be positional and angular deviations between the die-cut part and the preset position, which may result in lower assembly accuracy between the die-cut part and the mating part. Adjusting the relative position of the die-cut part and the mating part based on the positional deviation data using the assembly device helps improve the assembly accuracy of the die-cut part and the mating part.
[0048] Please see Figure 1 In one embodiment of this implementation, a visual module is provided, including: The vision module includes a recognition camera and a recognition lens, and selects the resolution of the recognition camera and the type of recognition lens based on the graphic features.
[0049] Specifically, graphic features include size. When the size of the graphic to be recognized is less than 1mm, a camera with a resolution of 5,000,000 pixels is selected as the recognition camera, and a telecentric lens is selected as the recognition lens. It is understandable that selecting a camera with a resolution of 5,000,000 pixels as the recognition camera can reduce costs while ensuring the accuracy of image data recognition. Selecting a telecentric lens as the recognition lens can reduce the risk of reduced image data quality acquired by the vision module due to perspective distortion, thereby improving the accuracy of the deep learning model in recognizing image data.
[0050] Please see Figure 1 In one embodiment of this implementation, a visual module is provided, including: A lighting module is provided, and the lighting parameters of the lighting module are selected according to the graphic features. The lighting parameters include the light source brightness, light source shape and light source color. The lighting module illuminates the die-cut parts. An evaluation scheme is formulated. The accuracy of the recognition module in recognizing image data is evaluated through the evaluation scheme, and the exposure time and lighting parameters of the recognition camera are adjusted.
[0051] Specifically, when the identified pattern is a cross or L-shape, a white ring light is selected as the light source; when the identified pattern is a circular hole, a monochromatic ring light is selected as the light source. Radial independent functional and performance tests are conducted on the vision module and the deep learning model. An ARR (Automatic Recognition) and GRR (Gross Recognition) inspection system is developed to verify the smoothness and timing coordination of the die-cut product formation and recognition process, thereby obtaining inspection analysis results. Based on the inspection analysis structure, the image processing parameters, feature extraction logic, and decision threshold of the deep learning model are adjusted. Furthermore, the conveying speed of the base film to the die-cut parts and the exposure time of the recognition camera are adjusted based on the inspection analysis results.
[0052] Understandably, selecting the lighting parameters of the lighting module based on graphic features can improve the clarity of the die-cut parts and the recognized graphics in the image data. By developing an evaluation scheme, the recognition effect of the die-cut parts can be verified. Adjusting the deep learning model and the vision module based on the evaluation results can improve the accuracy of the die-cut parts recognition.
[0053] Please see Figure 1 In one embodiment of this implementation, the die-cutting raw material strip includes: A base film is provided, and raw materials are die-cut to form die-cut parts on the base film. The base film is transparent, and image data of the die-cut parts on the base film is obtained through a vision module.
[0054] Specifically, the die-cutting component is conveyed to the background component via the base film, and the vision module acquires image data from the side of the die-cutting component away from the base film.
[0055] Understandably, die-cutting raw materials into die-cut components on a translucent base film can reduce the risk of light reflected from the background components being obstructed by the base film and reaching the vision module through the through-holes in the die-cut components, which is beneficial to improving the clarity of the recognized graphics in the image data.
[0056] Please see Figure 1 In one embodiment of this implementation, a visual module is provided, including: The vision module includes two recognition cameras, which acquire image data of the die-cut parts from opposite sides of the base film.
[0057] Specifically, two recognition cameras are arranged at intervals, and the die-cut parts are transported to the interval between the two recognition cameras through the bottom film, so that the two recognition cameras can acquire image data of the die-cut parts from opposite sides of the bottom film.
[0058] Understandably, the bottom film is transparent so that the two recognition cameras can acquire image data from opposite sides of the die-cut component, enabling the recognition of the die-cut component through image data from multiple angles, which helps to improve the accuracy of die-cut component recognition.
[0059] Please see Figure 1 In one embodiment of this implementation, a deep learning model is constructed using a dataset of good products and a dataset of defective products, including: The raw material is die-cut into multiple die-cut parts, and these parts are arranged in a queue. A transfer device is provided to remove unqualified die-cut parts from the queue, creating vacancies in the queue, and then qualified die-cut parts are moved into the vacancies.
[0060] Specifically, two raw material strips and two base films are provided. The two raw material strips are die-cut onto the two base films to form two queues of die-cut parts that correspond one-to-one on the two base films. The transfer device includes a robot arm that identifies the die-cut parts on the two base films using a deep learning model. The transfer device removes the defective die-cut parts from the queues on the two base films to create vacancies in the corresponding positions in the queues. The transfer device then transfers the qualified die-cut parts from one base film to the vacancies in the queues on the other base film.
[0061] Understandably, removing defective die-cut parts from the base film using a transfer device can reduce the risk of the die-cutting production line outputting defective die-cut parts. By forming two queues of die-cut parts through die-cutting and transferring the die-cut parts in one queue to the other queue, the die-cutting production line can output queues of qualified die-cut parts that conform to the expected arrangement.
[0062] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention. Furthermore, the embodiments of the present invention and the features thereof can be combined with each other unless otherwise specified.
Claims
1. A visual sorting method, characterized in that, include: Provide a raw material strip, obtain the material information of the raw material strip, and determine the optical properties of the raw material strip based on the material information; An identification pattern is set on the raw material strip, and the graphic features of the identification pattern are determined by the optical properties, including shape and color; The raw material strip is die-cut to form the die-cut part, and the die-cut part has the identification pattern. A vision module is provided to acquire image data of the die-cut component; Collect the image data, determine the position of the recognized graphic in the image data, and label the image data to form a good product dataset and a defective product dataset. Expand the defective product dataset by using a generative adversarial network. A deep learning model is constructed using the good product dataset and the defective product dataset, and the image data is identified using the deep learning model to determine whether the die-cut parts are qualified.
2. The visual sorting method according to claim 1, characterized in that, The step of setting the identification pattern on the raw material strip includes: A portion of the raw material strip is cut off, and through holes are formed on the raw material strip, the through holes constituting the identification pattern.
3. The visual sorting method according to claim 2, characterized in that, The die-cutting of the raw material strip includes: A background element is provided, the color of which is determined by the optical properties, such that the through-hole is opposite to the background element, so that the color of the background element fills the identification pattern.
4. The visual sorting method according to claim 1, characterized in that, The die-cutting of the raw material strip includes: The raw material is die-cut into multiple die-cut parts, and each die-cut part has the recognition pattern. The vision module acquires image data of the multiple die-cut parts, and a recognition module is provided to identify the relative position and relative angle of the recognition patterns on two adjacent die-cut parts to obtain the position deviation data of the die-cut parts.
5. The visual sorting method according to claim 4, characterized in that, The provision of constructing a deep learning model using the good product dataset and the defective product dataset includes: An assembly device and a mating component are provided. The positional deviation data is sent to the assembly device to assemble the die-cutting component and the mating component, and the assembly device adjusts the relative positions of the die-cutting component and the mating component according to the positional deviation data.
6. The visual sorting method according to claim 1, characterized in that, The provided visual module includes: The vision module includes a recognition camera and a recognition lens, and the resolution of the recognition camera and the type of the recognition lens are selected according to the graphic features.
7. The visual sorting method according to claim 4, characterized in that, The provided visual module includes: A lighting module is provided, and lighting parameters of the lighting module are selected according to the graphic features. The lighting parameters include light source brightness, light source shape, and light source color. The lighting module illuminates the die-cutting part. An evaluation scheme is formulated, and the accuracy of the recognition module in recognizing the image data is evaluated through the evaluation scheme. The exposure time of the recognition camera and the lighting parameters are adjusted.
8. The visual sorting method according to claim 1, characterized in that, The die-cutting of the raw material strip includes: A base film is provided, and the raw material is die-cut to form the die-cut component on the base film. The base film is light-transmitting, and the image data of the die-cut component on the base film is acquired through the vision module.
9. The visual sorting method according to claim 8, characterized in that, The provided visual module includes: The vision module includes two recognition cameras, which acquire image data of the die-cut component from opposite sides of the film surface of the base film.
10. The visual sorting method according to claim 1, characterized in that, The construction of a deep learning model using the good product dataset and the defective product dataset includes: The raw material is die-cut into multiple die-cut parts, and the multiple die-cut parts are arranged into a queue. A transfer device is provided to remove unqualified die-cut parts from the queue so that the corresponding position in the queue is empty, and qualified die-cut parts are moved into the empty position.