Coaxial telecentric imaging tinplate can defect identification method and system

Abstract

The invention provides a coaxial telecentric imaging tinplate can defect identification method and system, and relates to the technical field of defect identification, the system is composed of a coaxial telecentric camera assembly, a polarization multispectral light source assembly and an image processing and control unit, and can body images with constant magnification are obtained through a telecentric lens and coaxial illumination; a switchable cross-polarization visible light and near-infrared light source is used for synchronously collecting multi-modal images, and separation and difference are carried out on different polarization and wave band images to obtain a specular reflection component image, a diffuse reflection component image and a multi-spectral difference image. The image is divided into a character area, a pattern area and a bottom color area according to the format of the can body, comprehensive judgment is conducted through character comparison, a lightweight neural network and defect area connectivity analysis, automatic recognition of defects such as scratches, through printing and smudginess is achieved, and the accuracy and stability of appearance detection of the tin can on a high-speed production line can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of defect recognition, in particular to a coaxial telecentric imaging tin can defect recognition method and system. BACKGROUND

[0002] Tin cans are commonly used as packaging materials for food cans. The surface of the tin cans is plated with tin, which is not easy to rust, and has good sealing performance and metal luster. However, the printing quality, scratches, dirt and other defects on the surface of the tin cans during production will directly affect the appearance and quality of the products, and need to be effectively detected. Traditionally, the detection of can body labels, text printing and expiration date coding and the like relies on manual visual inspection, which is time-consuming and labor-intensive and the accuracy is difficult to guarantee. In recent years, in order to improve efficiency, automatic detection systems based on machine vision have appeared, for example, multiple cameras are used to obtain can body images from different angles, combined with polarization filters to reduce the interference of metal reflection on the imager, to detect defects such as missing labels, anti-fake seals and date codes. Although this multi-camera solution improves detection speed, it also has problems such as complex hardware, high calibration difficulty and high cost. In addition, the use of only white light illumination and ordinary imaging often makes it difficult to find some small and low-contrast defects, such as superficial scratches, fine oil stains or partial missing printing. SUMMARY

[0003] The coaxial telecentric imaging tin can defect recognition method comprises the following steps: A coaxial telecentric imaging device composed of a telecentric lens and an industrial camera is used to obtain images of the tin can, the telecentric lens ensures constant imaging magnification and extremely low distortion, and the coaxial imaging device includes a coaxial light source along the optical axis to uniformly illuminate the surface of the can body; A plurality of switchable polarized illumination light sources are provided, including a first linear polarized light source and a second linear polarized light source with perpendicular polarization directions, and at least one light source with a wavelength band different from the visible light wavelength band, wherein each light source is aligned with the optical axis of the coaxial imaging device; The polarized light sources are controlled to flash in a predetermined time sequence, and the industrial camera is triggered synchronously to capture multiple images when the same can body is in the imaging field of view, including a first image captured under the illumination of the first linear polarized light source, a second image captured under the illumination of the second linear polarized light source, and a third image captured under the illumination of a different wavelength band light source; The first image and the second image obtained are subjected to polarization component separation processing to extract a specular reflection component image and a diffuse reflection component image of the surface of the can body; the third image is compared and analyzed with a visible light image to obtain an image representing the multispectral reflection difference of the can body; The can body image is divided into at least a character area, a pattern area and a solid color background area, and corresponding feature images are selected for defect recognition in different areas: for the solid color background area, the specular reflection component image is used to detect high light or dark spots caused by scratches or stains; for the character area, the integrity of the characters in the diffuse reflection component image is analyzed and compared with the standard character information to identify missing printing or damage; for the pattern area, a pre-trained lightweight neural network model is used to determine ink missing, color error and other defects based on the multispectral difference image; The identification results of each area are combined to determine whether the tin can has defects, and a control signal is output to reject or alarm when a defect is detected.

[0004] Further, the telecentric lens is a material side telecentric structure, and the working field of view covers the area to be detected of the tin can. The coaxial light source is coupled into the telecentric lens light path through a semi-transparent lens, thereby achieving vertical and uniform illumination of the can body, so that the imaging sizes of different height or position parts of the can body are consistent.

[0005] Further, the first linear polarized light source and the second linear polarized light source are visible light LED light sources, the central wavelength range is 400-700 nm, and the included angle between the polarization axes of the polarizing plates in front of the two light sources is 90°. In the step 3), the first image and the second image correspond to the imaging directions parallel and perpendicular to the camera detection polarization direction, respectively. The first image contains a strong specular reflection component, and the specular glare in the second image is suppressed.

[0006] Further, the different waveband light source includes a near-infrared light source with a central wavelength of 700-1000 nm. By comparing and analyzing the gray scale difference between the third image (near-infrared image) and the visible light diffuse reflection component image, the change in reflectivity of the tin can printing ink and the metal base at different wavebands is identified to detect the invisible missing printing, fading or covering foreign matter.

[0007] Further, the polarization component separation processing includes taking the minimum value or fusing according to a predetermined ratio for corresponding pixels of the first image and the second image to generate the diffuse reflection component image; and calculating the difference or ratio of the first image and the second image to obtain the specular reflection component image, which is used to highlight the specular highlight area of the can body surface.

[0008] Further, the lightweight neural network model is a convolutional neural network, the input of which includes the diffuse reflection component image, the specular reflection component image and the multispectral difference image corresponding to the pattern area. After feature extraction and classification, it outputs whether there is a defect in the area. The parameter size of the neural network model is compressed to meet the requirement of high-speed real-time detection.

[0009] The multi-modal polarization system for tin can defect recognition under coaxial telecentric imaging, characterized in that the system comprises: A coaxial telecentric imaging device includes a telecentric lens, an industrial camera, and a coaxial light source coupling assembly. The object-side field of view of the telecentric lens covers the detection area of ​​the tin can, and the imaging surface is connected to the industrial camera. The coaxial light source coupling assembly is used to introduce illumination light along the optical axis into the telecentric lens to illuminate the can. The polarization and multispectral illumination module includes at least two sets of linearly polarized light sources with polarization directions perpendicular to each other and a set of near-infrared light sources. Each set of light sources is arranged to provide coaxial illumination with the optical axis of the telecentric lens and is equipped with an independent drive circuit to enable rapid on or off. The synchronization controller is electrically connected to the industrial camera and each light source. After receiving an external trigger signal, it controls each group of light sources in the light source module to emit light in sequence according to a preset timing sequence, and synchronously triggers the industrial camera to expose and acquire multiple corresponding images. The image processing unit receives multimodal image data output by the industrial camera. The built-in program is used to perform pixel-level operations on orthogonal polarization images to separate specular reflection and diffuse reflection components, perform differential enhancement processing on images of different bands, and perform region division and defect analysis on tank images. The decision and output unit generates a defect decision signal based on the analysis results of the image processing unit. When a defect above a predetermined threshold is identified in the tin can, the unit outputs a rejection control signal or an alarm signal through the interface for the production line to perform subsequent processing.

[0010] Furthermore, the telecentric lens in the coaxial telecentric imaging device has an object-side telecentric optical path and a large depth of field. The coaxial light source coupling component is a semi-reflective semi-transparent lens, which is installed in the optical path of the telecentric lens to reflect the light from the polarization and multispectral illumination module into the lens optical axis. The industrial camera uses a global shutter chip to avoid image motion blur during high-speed movement.

[0011] Furthermore, the polarization and multispectral illumination module includes a ring-shaped visible light LED light source and a ring-shaped near-infrared LED light source concentric with it. A replaceable polarizer assembly is attached in front of the visible light LED light source to provide linearly polarized illumination with a polarization direction of 0° or 90°. The module is equipped with multiple light source driving channels, which can switch the illumination state of different light sources within 1 millisecond to achieve near-synchronous acquisition of multimodal images.

[0012] Furthermore, the image processing unit includes an industrial computer or embedded processor and stores software algorithms or neural network models for defect identification. The algorithm includes instructions for calculating the minimum pixel value of two polarized images to generate the diffuse reflection component map, calculating the difference between the two polarized images to generate the specular reflection component map, and instructions for segmenting the image according to a preset region location. The neural network model receives multi-channel image data of a specific region and outputs the probability of defect presence. The image processing unit summarizes the judgment results for each region and outputs them to the decision and output unit.

[0013] Compared with the prior art, the present invention has the following advantages: This invention significantly improves the performance of automatic defect detection in tin cans by cleverly combining coaxial telecentric imaging and multimodal polarization detection. Telecentric imaging ensures image scale consistency and low distortion, providing a foundation for accurate comparison; polarization light source switching captures specular and diffuse reflection information, eliminating glare interference from metal surfaces and highlighting details of defects such as scratches; multi-band imaging expands the detection dimensions, making ink defects that are difficult to detect with the naked eye visible under specific bands; regionalized difference enhancement and lightweight AI algorithms optimize the discrimination strategy for different content, reducing false positives and false negatives. The entire system has a reasonable structure, sufficient creativity and industrial practical value, and can be widely applied to production lines for food canning, high-end metal packaging, etc., to achieve high-speed, high-precision online quality inspection. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the coaxial telecentric imaging multimodal polarization defect detection system of the present invention; Figure 2 This is a flowchart of the tin can defect identification method of the present invention. Detailed Implementation

[0015] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.

[0016] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.

[0017] Example A method for identifying defects in tin cans using coaxial telecentric imaging includes the following steps: A coaxial telecentric imaging device consisting of a telecentric lens and an industrial camera is used to acquire images of tin cans. The telecentric lens ensures a constant magnification and extremely low distortion. The coaxial imaging device includes a coaxial light source that illuminates along the optical axis to uniformly illuminate the surface of the can. Multiple sets of switchable polarized illumination sources are provided, including a first linearly polarized source and a second linearly polarized source with polarization directions perpendicular to each other, and at least one light source with a wavelength different from the visible light band, wherein each light source is aligned with the optical axis of the coaxial imaging device. The polarization light source is controlled to flash sequentially according to a predetermined time sequence, and the industrial camera is triggered synchronously to capture multiple images when the same tank is in the imaging field of view, including: a first image captured under the illumination of the first linearly polarized light source, a second image captured under the illumination of the second linearly polarized light source, and a third image captured under the illumination of light sources of different wavelengths. The first and second images are subjected to polarization component separation processing to extract the specular reflection component and diffuse reflection component of the tank surface; the third image is selected and compared with the visible light image to obtain an image characterizing the multispectral reflectance differences of the tank. The can image is divided into at least a text area, a pattern area, and a solid color background area. For each area, corresponding feature images are selected for defect identification: for the solid color background area, the specular reflection component image is used to detect highlights or dark spots caused by scratches or stains; for the text area, the integrity of the characters in the diffuse reflection component image is analyzed and compared with standard character information to identify missing or damaged characters; for the pattern area, a pre-trained lightweight neural network model is used in conjunction with the multispectral difference image to identify defects such as missing ink or miscolor. Based on the identification results of each area, a comprehensive judgment is made as to whether the tin can has defects, and a control signal is output to remove or alarm when a defect is detected.

[0018] Furthermore, the telecentric lens has an object-side telecentric structure, and its working field of view covers the area of ​​the tin can to be inspected. The coaxial light source is coupled into the optical path of the telecentric lens through a semi-transparent lens, thereby achieving vertical and uniform illumination of the can and making the imaging size of the can parts at different heights or positions consistent.

[0019] Furthermore, the first and second linearly polarized light sources are visible light LED light sources with a center wavelength range of 400–700 nm, and the polarization axis of the polarizers placed in front of them is 90°; in step 3, the first image and the second image correspond to imaging with polarization axes parallel and perpendicular to the camera detection polarization direction, respectively, wherein the first image contains a strong specular reflection component, and specular glare is suppressed in the second image.

[0020] Furthermore, the different wavelength light sources include near-infrared light sources with a center wavelength of 700–1000 nm; by comparing and analyzing the grayscale difference between the third image (near-infrared image) and the visible light diffuse reflectance component image, the changes in reflectance of tin can printing ink and metal substrate at different wavelengths can be identified to detect invisible defects such as printing omissions, fading, or foreign matter covering.

[0021] Furthermore, the polarization component separation process includes taking the minimum value of corresponding pixels in the first image and the second image or fusing them at a predetermined ratio to generate the diffuse reflection component map; and calculating the difference or ratio between the first image and the second image to obtain the specular reflection component map, which is used to highlight the specular highlight area on the surface of the tank.

[0022] Furthermore, the lightweight neural network model is a convolutional neural network, whose input includes the diffuse reflection component map, specular reflection component map, and multispectral difference map corresponding to the pattern area. After feature extraction and classification, it outputs whether there are defects in the area. The parameter scale of the neural network model is compressed to meet the requirements of high-speed real-time detection.

[0023] A multimodal polarization system for identifying defects in tin cans using coaxial telecentric imaging, characterized in that the system comprises: A coaxial telecentric imaging device includes a telecentric lens, an industrial camera, and a coaxial light source coupling assembly. The object-side field of view of the telecentric lens covers the detection area of ​​the tin can, and the imaging surface is connected to the industrial camera. The coaxial light source coupling assembly is used to introduce illumination light along the optical axis into the telecentric lens to illuminate the can. The polarization and multispectral illumination module includes at least two sets of linearly polarized light sources with polarization directions perpendicular to each other and a set of near-infrared light sources. Each set of light sources is arranged to provide coaxial illumination with the optical axis of the telecentric lens and is equipped with an independent drive circuit to enable rapid on or off. The synchronization controller is electrically connected to the industrial camera and each light source. After receiving an external trigger signal, it controls each group of light sources in the light source module to emit light in sequence according to a preset timing sequence, and synchronously triggers the industrial camera to expose and acquire multiple corresponding images. The image processing unit receives multimodal image data output by the industrial camera. The built-in program is used to perform pixel-level operations on orthogonal polarization images to separate specular reflection and diffuse reflection components, perform differential enhancement processing on images of different bands, and perform region division and defect analysis on tank images. The decision and output unit generates a defect decision signal based on the analysis results of the image processing unit. When a defect above a predetermined threshold is identified in the tin can, the unit outputs a rejection control signal or an alarm signal through the interface for the production line to perform subsequent processing.

[0024] Furthermore, the telecentric lens in the coaxial telecentric imaging device has an object-side telecentric optical path and a large depth of field. The coaxial light source coupling component is a semi-reflective semi-transparent lens, which is installed in the optical path of the telecentric lens to reflect the light from the polarization and multispectral illumination module into the lens optical axis. The industrial camera uses a global shutter chip to avoid image motion blur during high-speed movement.

[0025] Furthermore, the polarization and multispectral illumination module includes a ring-shaped visible light LED light source and a ring-shaped near-infrared LED light source concentric with it. A replaceable polarizer assembly is attached in front of the visible light LED light source to provide linearly polarized illumination with a polarization direction of 0° or 90°. The module is equipped with multiple light source driving channels, which can switch the illumination state of different light sources within 1 millisecond to achieve near-synchronous acquisition of multimodal images.

[0026] Furthermore, the image processing unit includes an industrial computer or embedded processor and stores software algorithms or neural network models for defect identification. The algorithm includes instructions for calculating the minimum pixel value of two polarized images to generate the diffuse reflection component map, calculating the difference between the two polarized images to generate the specular reflection component map, and instructions for segmenting the image according to a preset region location. The neural network model receives multi-channel image data of a specific region and outputs the probability of defect presence. The image processing unit summarizes the judgment results for each region and outputs them to the decision and output unit.

[0027] This invention provides a multimodal polarization method for identifying defects in tin cans using coaxial telecentric imaging, comprising the following steps: Step 1: Constant Magnification Imaging. The body and lid of the tin can are imaged using a coaxial telecentric lens and an industrial camera. The telecentric optical system ensures constant magnification within the field of view, resulting in distortion-free and geometrically stable images. This ensures consistent inspection standards across all areas, unaffected by minor fluctuations in the can's position. A coaxial integrated light source illuminates the can along the optical axis via a beam splitter, providing uniform illumination and highlighting surface details.

[0028] Step 2: Multi-polarization, multi-spectral illumination. Building upon coaxial illumination, multiple sets of switchable polarization-oriented light sources and light sources of different wavelengths are introduced. Preferably, two sets of linearly polarized light sources are provided, with their polarization directions orthogonal to each other (e.g., 0° and 90°); the spectrum includes visible light and near-infrared illumination. For example, a visible light LED with a polarizer serves as the initial polarization source, while another set has a perpendicular polarization direction; and a set of near-infrared LED light sources (also with polarizers). Each light source is independently triggered by a controller, allowing light of different polarization states and wavelengths to sequentially flash and illuminate the surface of the tank.

[0029] Step 3: Multimodal Image Acquisition. The industrial camera and the light source are synchronously triggered to sequentially acquire images under different illumination modes within a very short time. Specifically, image A is first captured under visible light parallel polarization illumination, then image B is captured under visible light vertical polarization illumination; subsequently, image C (with optional polarization) is captured under near-infrared illumination. Due to the extremely short exposure interval and telecentric imaging avoiding parallax, each image maintains strict correspondence. Multiple synchronous multimodal image data are thus obtained for the same tank, including visible light images and multi-band images under different polarization states.

[0030] Step 4: Polarization Information Separation and Defect Feature Enhancement. Pixel-level registration and multi-channel processing are performed on the acquired multimodal images. Utilizing the principle of polarization imaging, the two visible light images obtained from orthogonally polarized illumination are calculated to separate the specular reflection component and the diffuse reflection component. For example, by comparing the intensity differences of each pixel in the parallel polarization image A and the vertical polarization image B, strong specular reflection areas formed by linearly polarized light are filtered out, obtaining image D, which mainly represents diffuse reflection; simultaneously, the specular highlight component image E is obtained based on the difference between the two images to highlight surface reflective features. This separation makes specular abnormal highlights caused by scratches, etc., clearly presented in image E, while image D retains the uniformity information of the printed pattern and background color, which helps to detect diffuse reflection anomalies such as color / ink loss. In addition, comparing visible light and near-infrared images can reveal the contrast changes of certain defects under different spectra: for example, if ink is missing in a certain area of ​​the can, the metal substrate has a high reflectivity in the near-infrared, and this area will appear brighter in the infrared image C, thus forming a strong contrast with the surrounding well-printed areas. This multispectral difference information can also be extracted to enhance the visibility of specific defects. Through the above polarization and spectral separation, the present invention can generate multiple enhancement channels, such as diffuse reflection enhancement maps, specular reflection enhancement maps, and multispectral difference maps, providing useful clues for the detection of different defects.

[0031] Step 5: Image Region Segmentation. Based on the tank's design features or a pre-trained model, the tank image is divided into several functional regions, such as text coding areas, pattern printing areas, and solid background areas. Text areas typically include production dates or batch codes, pattern areas feature colored trademarks or decorations, and solid background areas consist of large areas of a single-color coating. This region segmentation can be accomplished using template matching or image segmentation algorithms, ensuring that each region has relatively consistent texture features and detection requirements. The purpose of region segmentation is to select the most suitable defect discrimination strategy for different regions, thereby improving detection accuracy.

[0032] Step 6: Defect Identification and Classification. For different regions and different enhancement channels, differential analysis or lightweight neural network models are used for defect identification respectively: For areas with a solid base color: The specular reflection enhancement image E and the diffuse uniformity image D are used to compare and analyze local brightness anomalies. For example, in areas with a solid base color coating, image D normally has a uniform tone. If a scratch (exposed metallic reflection) or dirt (altering diffuse reflection characteristics) appears in a certain area, it will cause grayscale anomalies in image D. Simultaneously, the scratch may appear as a bright or dark line in the specular enhancement image E. By comparing D and E, or calculating their difference, the characteristics of such defects can be significantly highlighted. Simple image thresholding or edge detection can mark the scratch outline. Subsequently, features such as area and shape are combined to determine the defect type; for example, thin, elongated linear highlights are identified as scratch defects.

[0033] Text encoding area: This area contains text information such as production dates printed or embossed. The main defects are missing characters, illegible text, or incorrect characters. OCR algorithms can be used to obtain the actual string in this area, which is then compared with the standard information to identify missing or incorrect characters. Simultaneously, differential enhancement is used: the diffuse reflection image D is compared with a pre-stored defect-free standard image; if any character strokes are missing, they will appear as bright spots in the differential image. Furthermore, a lightweight convolutional neural network (CNN) can be trained to classify small images of this area, outputting a judgment of "character complete" or "character unclear / missing". Polarization removes metallic glare interference and improves the image contrast of the character area, thereby enhancing the reliability of the OCR and the algorithm.

[0034] Pattern Printing Area: This area contains complex colored graphics or trademarks. Defects may manifest as localized ink loss, color mismatch, or contaminant coverage. Due to the complexity of the pattern content, this invention preferably employs a lightweight neural network model for discrimination. This model can be input with multi-channel information from the area, including diffuse reflection image (D), infrared image (C), and differential enhancement image. Through convolution, it extracts texture and color features and automatically learns the differences between normal and defective areas to output classification results. For example, areas with ink loss exhibit different reflectance characteristics in multispectral channels compared to their surroundings; the CNN can thus classify this as a defect. This model is small in scale and has a fast inference speed, allowing for real-time results during high-speed production line operation. Furthermore, to improve interpretability, simple image processing can be incorporated: for example, applying color difference to the pattern area to mark pixel areas with significant differences between the infrared and visible images, indicating potential printing omissions for further manual confirmation or model focus.

[0035] After the above steps, the system summarizes the detection results for each region. If any region has an anomaly exceeding a threshold, the tin can is determined to be a defective product, and a defect category (e.g., scratches, missing lettering, dirt, etc.) is output. Simultaneously, the system records the defect location and removes the defective can through production line control. It is worth noting that the rich information obtained using polarization and multispectral imaging improves the robustness of defect detection. For example, polarization imaging effectively suppresses glare on bright curved surfaces, improves image uniformity and signal-to-noise ratio, and helps to reliably extract defect textures. Compared to traditional solutions relying on a single image, this invention can significantly improve detection accuracy and anti-interference capabilities without significantly increasing detection time (acquiring multiple images in a single trigger, completed in milliseconds), making the system more suitable for high-speed production scenarios.

[0036] Example 1: System Hardware Structure (see...) Figure 1The system in this embodiment includes a coaxial telecentric imaging assembly consisting of an industrial camera 1 and a telecentric lens 2, used to acquire high-resolution images of the surface of tin cans. The telecentric lens 2 has a built-in semi-transparent lens that couples parallel light from the coaxial light source assembly 3 into the optical path. The light source assembly 3 includes multiple independently controllable sub-light sources: visible light polarized light sources 3a and 3b (e.g., ring LED light sources, with pre-polarized filters set to polarization directions of 0° and 90° respectively), and a near-infrared light source 3c (with a polarizer or ring illumination). The controller 4 is electrically connected to the camera 1 and the light source 3, and can sequentially illuminate each sub-light source and synchronize camera exposure in a very short time according to a trigger signal. The tin can 5 being inspected is located in the center of the camera's field of view and is conveyed through the inspection area by a conveyor belt. The system also includes a main control computer 6, which has built-in image processing and discrimination software for receiving camera images and executing defect recognition algorithms. The output of the main control computer 6 is connected to the production line rejection device, and triggers the rejection cylinder when a defect is detected.

[0037] In this embodiment, the telecentric lens 2 employs an object-side telecentric design to ensure consistent magnification across all parts of the can surface, maintaining stable image size even with minor vibrations or diameter differences in the can. Light sources 3a and 3b are coaxially mounted, allowing light to perpendicularly illuminate the can surface and be collected by the lens, providing uniform illumination and highlighting surface details. Linear polarizers are used, with their polarization directions orthogonal to each other, resulting in one set of illumination polarizations parallel to the analyzer at the camera front (obtaining specular + diffuse reflection images), and another set perpendicular to it (primarily obtaining diffuse reflection information). The near-infrared light source 3c has a center wavelength of, for example, 850nm, to supplement detection by utilizing the difference in reflectivity between the tinplate substrate and the ink in the infrared band. The brightness of all light sources has been experimentally set to ensure that the grayscale of each channel image obtained by the camera is within the appropriate exposure range, without overexposure or underexposure.

[0038] Example 2: Multimodal Image Acquisition and Processing (see...) Figure 2 When tank 5 enters the center of camera 1's field of view via the conveyor belt, the position sensor sends a trigger signal, and controller 4 begins to drive the light source and camera to acquire images according to a predetermined sequence. First, the visible light polarized light source 3a is lit, and camera 1 exposes and acquires image A; then 3a is turned off, and the orthogonally polarized light source 3b is lit, and the camera exposes again to acquire image B; subsequently, the near-infrared light source 3c is triggered, and the camera acquires infrared image C. The entire process is completed within milliseconds, and can be completed while a single tank is stationary (or moving at low speed), achieving quasi-synchronous multi-channel imaging. The resulting images A and B are visible light images with different polarization states at the same viewing angle, and C is the infrared image at the corresponding viewing angle. The main control computer 6 first performs grayscale correction and alignment on the three images (since telecentric imaging has no parallax, A, B, and C are naturally aligned in space, and only brightness difference correction needs to be considered).

[0039] Then, computer 6 executes a polarization component separation algorithm: comparing the intensity of A and B pixel by pixel, for each pixel, the lower value of the two is taken as the gray level of the diffuse reflection component D (equivalent to cross-polarization filtering out most specular reflection); the difference between the two, or the larger value minus the smaller value, is taken to obtain the specular reflection enhancement component E. Thus, image D reflects the substrate image of the can without specular highlight interference, suitable for checking color uniformity and printing integrity; image E highlights which areas have strong directional reflection differences, i.e., possible scratches, highlights, etc. Next, pixel subtraction is performed between the visible light diffuse reflection image D and the infrared image C to obtain the spectral difference image F. For example, for a certain pattern area, if the infrared is much brighter than the visible light, then that area will show brightness in F, indicating that there may be substrate exposure (missing print) or differences in the covering material. Through the above processing, computer 6 obtains multiple derived images (D, E, F), which, together with the original images A, B, and C, constitute a multimodal image dataset for analysis.

[0040] Example 3: Region Differentiation and Defect Recognition Algorithm. Using can design information, computer 6 divides the image into predefined regions of interest: such as the date code area on the lid, the trademark pattern area on the can body, and the plain background area of ​​the can body. This can be achieved through image coordinate mapping (e.g., setting the position range of each region in the image based on the can size and printing layout). For each region, the most effective set of feature images and processing methods are selected: Date code text area: Computer 6 performs binarization or OCR recognition on the diffuse reflection image D, reads the actual characters, compares them with the standard template, and detects missing or misprinted characters. Combined with the mirror enhancement image E, if there are abnormal reflections or missing strokes at the character location (such as incomplete printing causing localized metallic sheen), a bright spot will appear at that stroke location in image E, serving as evidence of a defect. Once unclear or missing characters are found, they are recorded as a printing defect.

[0041] Trademark Pattern Area: This area has complex textures. The computer extracts a sub-image of this area from the D and F images and feeds it into a pre-trained convolutional neural network model for discrimination. This lightweight model (e.g., a few convolutional layers + fully connected layers) is trained with normal, defect-free patterns and various defect samples, and can output whether there are anomalies such as missing prints or miscolors in this area. Because the F image highlights the spectral contrast caused by ink loss, the model can more easily capture missing print features, improving the accuracy.

[0042] Solid Color Background Area: Computer 6 analyzes the specular enhancement image E. A normal, good coating background should be almost entirely black in image E (without specular highlights). Bright stripes indicate scratches revealing metallic luster; small, localized bright spots may indicate foreign matter adhering to the surface. These bright defects are located using morphological filtering and connected component detection, and the corresponding locations are checked in image D. If accompanied by abnormal diffuse reflection brightness (darkened or brightened spots), it can be further confirmed whether it is dirt or a scratch. Judgment logic, for example: thin, elongated bright lines with a darkened image D indicate scratches; irregular bright lines with a darkened / discolored image D indicate contamination spots.

[0043] In summary, the inspection results from each region are sent to the decision module. If any defect is detected as true, the system determines that tank 5 is unqualified and drives the mechanical rejection device to remove it from the production line via an output signal; at the same time, the defect type and image of the tank are recorded in the database for traceability.

[0044] Example 4: System Parameters and Performance. In practical applications, system parameters can be adjusted according to the production line speed and detection accuracy requirements. For example, camera 1 is selected as a global shutter high-speed camera with a resolution of over 2 megapixels and a frame rate of no less than 100fps to ensure rapid and continuous acquisition of multiple images upon triggering. The flash duration of light sources 3a / 3b can be set to several hundred microseconds to ensure exposure is completed during motion cessation. The extinction ratio of the polarizer needs to be sufficiently high (e.g., above 500:1) to effectively filter glare. The wavelength of the near-infrared light source can also be adjusted according to the characteristics of the ink material. For example, some colored inks have higher transparency at 940nm, so a 940nm LED is selected to highlight the missing print. The inference of the lightweight neural network can be completed on an industrial control computer CPU or GPU. If higher speed requirements are needed, it can also be ported to an FPGA / edge computing device.

[0045] Those skilled in the art will understand that the various embodiments disclosed above can be modified and altered in various ways without departing from the spirit of the invention. Therefore, the scope of protection of this invention should be defined by the appended claims.

[0046] It should be noted that not all steps and units in the above processes are necessary; some steps or units can be omitted as needed. The execution order of each step is not fixed and can be determined as required. The device structure described in the above embodiments can be a physical structure or a logical structure. That is, some units may be implemented by the same physical entity, or some units may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.

[0047] The specific embodiments described above are exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of the claims. The term "exemplary" as used throughout this specification means "serving as an example, instance, or illustration" and does not imply that it is "preferred" or "advantageous" compared to other embodiments. Specific details are included to provide an understanding of the described techniques. However, these techniques can be practiced without these specific details. In some instances, well-known structures and apparatuses are shown in block diagram form to avoid obscuring the concepts of the described embodiments.

[0048] The foregoing description of this disclosure is provided to enable any person skilled in the art to implement or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the widest scope of the principles and novel features disclosed herein.

Claims

1. A method for identifying defects in tin cans using coaxial telecentric imaging, characterized in that, Includes the following steps: A coaxial telecentric imaging device consisting of a telecentric lens and an industrial camera is used to acquire images of tin cans. The telecentric lens ensures a constant magnification and extremely low distortion. The coaxial imaging device includes a coaxial light source that illuminates along the optical axis to uniformly illuminate the surface of the can. Multiple sets of switchable polarized illumination sources are provided, including a first linearly polarized source and a second linearly polarized source with polarization directions perpendicular to each other, and at least one light source with a wavelength different from the visible light band, wherein each light source is aligned with the optical axis of the coaxial imaging device. The polarization light source is controlled to flash sequentially according to a predetermined time sequence, and the industrial camera is triggered synchronously to capture multiple images when the same tank is in the imaging field of view, including: a first image captured under the illumination of the first linearly polarized light source, a second image captured under the illumination of the second linearly polarized light source, and a third image captured under the illumination of light sources of different wavelengths. The first and second images are subjected to polarization component separation processing to extract the specular reflection component and diffuse reflection component of the tank surface; the third image is selected and compared with the visible light image to obtain an image characterizing the multispectral reflectance differences of the tank. The can image is divided into at least a text area, a pattern area, and a solid color background area. For each area, corresponding feature images are selected for defect identification: for the solid color background area, the specular reflection component image is used to detect highlights or dark spots caused by scratches or stains; for the text area, the integrity of the characters in the diffuse reflection component image is analyzed and compared with standard character information to identify missing or damaged characters; for the pattern area, a pre-trained lightweight neural network model is used in conjunction with the multispectral difference image to identify defects such as missing ink or miscolor. Based on the identification results of each area, a comprehensive judgment is made as to whether the tin can has defects, and a control signal is output to remove or alarm when a defect is detected.

2. The method for identifying defects in tin cans using coaxial telecentric imaging according to claim 1, characterized in that: The telecentric lens has an object-side telecentric structure, and its working field of view covers the area to be inspected on the tin can. The coaxial light source is coupled into the optical path of the telecentric lens through a semi-transparent lens, thereby achieving vertical and uniform illumination of the can and making the imaging size of the can parts at different heights or positions consistent.

3. The method for identifying defects in tin cans using coaxial telecentric imaging according to claim 1, characterized in that: The first and second linearly polarized light sources are visible light LED light sources with a center wavelength range of 400–700 nm. The polarization axis of the polarizers placed in front of them is 90°. In step 3), the first image and the second image correspond to the imaging with polarization axes parallel and perpendicular to the camera detection polarization direction, respectively. The first image contains a strong specular reflection component, and specular glare is suppressed in the second image.

4. The method for identifying defects in tin cans using coaxial telecentric imaging according to claim 1, characterized in that: The different wavelength light sources include near-infrared light sources with a center wavelength of 700–1000 nm; by comparing and analyzing the grayscale difference between the third image (near-infrared image) and the visible light diffuse reflectance component image, the changes in reflectance of tin can printing ink and metal substrate in different wavelength bands are identified, so as to detect invisible omissions, fading or foreign matter covering.

5. The method for identifying defects in tin cans using coaxial telecentric imaging according to claim 1, characterized in that: The polarization component separation process includes taking the minimum value of corresponding pixels in the first image and the second image or fusing them at a predetermined ratio to generate the diffuse reflection component map; and calculating the difference or ratio between the first image and the second image to obtain the specular reflection component map, which is used to highlight the specular highlight area on the surface of the tank.

6. The method for identifying defects in tin cans using coaxial telecentric imaging according to claim 1, characterized in that: The lightweight neural network model is a convolutional neural network. Its input includes the diffuse reflection component map, specular reflection component map, and multispectral difference map corresponding to the pattern area. After feature extraction and classification, it outputs whether there are defects in the area. The neural network model parameter scale is compressed to meet the requirements of high-speed real-time detection.

7. A multimodal polarization system for identifying defects in tin cans using coaxial telecentric imaging, used to implement the method described in any one of claims 1 to 6, characterized in that, The system includes: A coaxial telecentric imaging device includes a telecentric lens, an industrial camera, and a coaxial light source coupling assembly. The object-side field of view of the telecentric lens covers the detection area of ​​the tin can, and the imaging surface is connected to the industrial camera. The coaxial light source coupling assembly is used to introduce illumination light along the optical axis into the telecentric lens to illuminate the can. The polarization and multispectral illumination module includes at least two sets of linearly polarized light sources with polarization directions perpendicular to each other and a set of near-infrared light sources. Each set of light sources is arranged to provide coaxial illumination with the optical axis of the telecentric lens and is equipped with an independent drive circuit to enable rapid on or off. The synchronization controller is electrically connected to the industrial camera and each light source. After receiving an external trigger signal, it controls each group of light sources in the light source module to emit light in sequence according to a preset timing sequence, and synchronously triggers the industrial camera to expose and acquire multiple corresponding images. The image processing unit receives multimodal image data output by the industrial camera. The built-in program is used to perform pixel-level operations on orthogonal polarization images to separate specular reflection and diffuse reflection components, perform differential enhancement processing on images of different bands, and perform region division and defect analysis on tank images. The decision and output unit generates a defect decision signal based on the analysis results of the image processing unit. When a defect above a predetermined threshold is identified in the tin can, the unit outputs a rejection control signal or an alarm signal through the interface for the production line to perform subsequent processing.

8. The multimodal polarization system for identifying defects in tin cans under coaxial telecentric imaging as described in claim 7, characterized in that: The telecentric lens in the coaxial telecentric imaging device has an object-side telecentric optical path and a large depth of field. The coaxial light source coupling component is a semi-reflective semi-transparent lens, which is installed in the optical path of the telecentric lens to reflect the light from the polarization and multispectral illumination module into the lens optical axis. The industrial camera uses a global shutter chip to avoid image motion blur during high-speed movement.

9. The multimodal polarization system for identifying defects in tin cans under coaxial telecentric imaging as described in claim 7, characterized in that: The polarization and multispectral illumination module includes a ring-shaped visible light LED light source and a ring-shaped near-infrared LED light source concentric with it. A replaceable polarizer assembly is attached in front of the visible light LED light source to provide linearly polarized illumination with a polarization direction of 0° or 90°. The module is equipped with multiple light source driving channels, which can switch the illumination state of different light sources within 1 millisecond to achieve near-synchronous acquisition of multimodal images.

10. The multimodal polarization system for identifying defects in tin cans under coaxial telecentric imaging according to claim 7, characterized in that: The image processing unit includes an industrial computer or embedded processor and stores software algorithms or neural network models for defect identification. The algorithm includes instructions for calculating the minimum pixel value of two polarized images to generate the diffuse reflection component map, calculating the difference between the two polarized images to generate the specular reflection component map, and instructions for segmenting the image according to preset region locations. The neural network model receives multi-channel image data of a specific region and outputs the probability of defect presence. The image processing unit summarizes the judgment results for each region and outputs them to the decision and output unit.

Images