Cigarette end face visual inspection equipment and method thereof

Abstract

The invention discloses cigarette end face visual inspection equipment and a method thereof, and relates to the technical field of machine vision automatic inspection. Multi-view, multi-modal and three-dimensional environment data are acquired through an industrial camera, an additional modal acquisition unit, an environment perception sensor and a synchronous trigger controller, and are calibrated by a calibration data memory; the feature extraction and fusion module is used for extracting depth features, the physical simulation and guide module is used for simulating cigarette deformation, and a predicted deformation contour and a physical guide signal are output. After the instance proposal module determines a detection area, the GAN shape refining module generates self-adaptive refined instance representation, through the roundness calculation and uncertainty module and the deformation reason inference module, roundness errors and uncertainty are calculated, deformation reasons are judged, and finally a detection result is output by the result output module. High-precision and robust automatic cigarette end face defect detection is realized, and the production quality control efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine vision automated detection, and in particular to a device and method for visually detecting cigarette end faces. Background Art

[0002] The purpose of cigarette end-face inspection is to accurately assess the appearance and geometry of cigarette ends, promptly identifying defects or deformations that may occur during the production process, thereby improving product quality and consistency. By inspecting cigarette end faces, problems such as lack of roundness, irregular edges, and manufacturing process defects can be effectively identified. This helps to promptly remove defective products, optimize production processes, reduce defective rates, and provide data support for quality improvement. Furthermore, this inspection technology improves production efficiency and enables automated monitoring, ensuring product compliance with relevant standards and consumer expectations, thereby enhancing market competitiveness and brand reputation.

[0003] Existing online cigarette appearance inspection methods and devices collect three-dimensional information images of the cigarette holder end cavity, images of the cigarette holder end appearance, and images of the cigarette circumference. These images are then classified using an image processor and a deep learning model to distinguish good from bad cigarettes and eliminate defective cigarettes. However, due to single-modality data acquisition and limited inspection angles, these methods still have limitations in terms of detection efficiency, robustness, and comprehensiveness in detecting cigarette end defects. Summary of the Invention

[0004] The present invention provides a device and method for visually inspecting cigarette end faces, which realizes high-precision and robust automatic inspection of cigarette end face defects and improves production quality control efficiency.

[0005] The present invention provides a cigarette end face visual inspection device, including an industrial camera, a light source system and a control system. The control system includes:

[0006] Feature extraction and fusion module, which is used to extract and fuse deep features based on multi-view, multi-modal images acquired from industrial cameras and optional additional modality acquisition units to generate a unified fused feature representation;

[0007] A physical simulation and guidance module, which is used to perform quantitative physical simulation based on the estimation or perception of the surrounding environment and preset material properties to predict the deformation mode of an ideal round cigarette in the current environment, and output the predicted deformation profile and physical guidance signal;

[0008] The instance proposal module is used to propose the initial position and range of the cigarette instance based on the fusion feature representation and physical guidance signals, and generate instance proposals;

[0009] The GAN shape refinement module uses a generative adversarial network structure to receive the fused feature representation and instance proposals, and combines them with physical guidance signals to generate an adaptive and refined instance representation for each cigarette instance through adversarial training;

[0010] The roundness calculation and uncertainty module is used to extract the actual detection contour from the adaptive refined instance representation, calculate the roundness error of the cigarette end face, and quantitatively evaluate the uncertainty of the roundness error; the uncertainty is used to describe the accuracy of the roundness error;

[0011] The deformation cause inference module is used to compare the shape similarity between the actual detected contour and the predicted deformed contour, infer the cause of the roundness error, and output the deformation cause classification label and the corresponding confidence level;

[0012] The result output module is used to aggregate the roundness error, uncertainty, deformation cause classification label and confidence, and output them.

[0013] The device also includes: an environmental perception sensor, selected from at least one of a three-dimensional structured light scanner or a laser ranging unit, for acquiring spatial distribution information of adjacent cigarettes and providing the information to a physical simulation and guidance module within the control system; an additional modality acquisition unit, selected from at least one of a polarized light acquisition unit and a near-infrared / hyperspectral imaging unit, for acquiring supplementary modal information and providing the information to a feature extraction and fusion module within the control system; a synchronous trigger controller for ensuring synchronous operation of the industrial camera, the environmental perception sensor, and the additional modality acquisition unit; and a calibration data memory storing calibration parameters of each sensor.

[0014] The industrial camera includes at least three cameras, which are arranged at different spatial viewing angles.

[0015] The feature extraction and fusion module includes:

[0016] A feature extraction unit, configured to extract deep feature maps from images of various modalities and perspectives using a convolutional neural network;

[0017] The feature fusion unit is used to fuse deep feature maps to generate a unified fused feature representation by adopting one or more fusion strategies including cross-modal and cross-view attention mechanisms and geometric transformation alignment based on camera calibration parameters.

[0018] The physical simulation and guidance module includes:

[0019] a simulation execution unit, configured to quantitatively calculate the deformation field of an ideal circular cigarette and generate a predicted deformation profile based on neighboring environmental information, preset material mechanical properties, and boundary conditions, using one or a combination of a finite element method approximation, a boundary element method approximation, a spring mass model, and a neural network proxy model;

[0020] The guidance signal generating unit is used to generate physical guidance signals including stress concentration diagram, contact probability diagram or predicted displacement direction diagram according to the simulation calculation results.

[0021] The generator unit of the GAN shape refinement module is used to: derive an instance complexity metric based on the evaluation of input features or internal features. The instance complexity metric is used to measure the degree of instance adhesion, predict deformation amplitude, or shape irregularity;

[0022] According to the instance complexity index, one or a combination of the output including parameterized circle / ellipse parameters, spline curve defined by control points, local implicit field function parameters and edge probability map is adaptively selected as the adaptive refined instance representation.

[0023] The GAN shape refinement module also includes:

[0024] a discriminator unit, configured to receive the adaptively refined instance representation or its derived features and determine its similarity or authenticity with the real cigarette shape features;

[0025] The loss calculation unit is used to calculate at least an adversarial loss for promoting the generation of realistic shapes, a reconstruction loss for ensuring that the generated shapes are consistent with the target, and a physical consistency loss for penalizing significant deviations between the generated shapes and physical simulation predictions.

[0026] The circularity calculation and uncertainty module includes:

[0027] A contour processing unit is used to decode or sample a 3D edge contour point set from the adaptively refined instance representation, determine the best fitting plane using a robust plane fitting algorithm, and orthogonally project the 3D contour points onto the best fitting plane to obtain a 2D point set of the actual detected contour;

[0028] a shape fitting unit for applying a robust circle or ellipse fitting algorithm to the two-dimensional point set of the actual detected contour;

[0029] a metric calculation unit, configured to calculate a roundness error based on the fitting result and a two-dimensional point set of the actual detected contour, wherein the roundness error is selected from at least one metric selected from the group consisting of a radius standard deviation, a minimum / maximum radius ratio, and a ratio of an inner area of ​​the contour to an area of ​​the fitted circle / ellipse;

[0030] The uncertainty quantification unit is used to combine at least two pieces of information selected from the confidence level output by the GAN shape refinement module, the fitting residual of the shape fitting unit, and the consistency between the actual detected contour and the physical simulation prediction to calculate the uncertainty.

[0031] The deformation cause inference module includes:

[0032] a shape comparison unit, configured to quantify the shape similarity between the actual detected contour and the predicted deformed contour by using one or a combination selected from shape descriptor distance calculation, deformation field matching analysis, and contour curvature distribution correlation analysis;

[0033] The cause classification unit is used to determine the deformation cause classification label as external extrusion-dominated, manufacturing defect-dominated, or mixed based on shape similarity, the degree of physically predicted deformation, the actual roundness error, and the geometric characteristics of the surrounding environment through preset rule logic or pre-trained machine learning classifiers;

[0034] The confidence scoring unit is used to calculate the confidence based on the shape similarity and the strength of the discrimination basis of the cause classification unit.

[0035] A method for visually inspecting cigarette end faces, comprising:

[0036] Receive a multi-view image of the target cigarette end face containing at least RGB information, polarization information or near-infrared / hyperspectral information, and adjacent environment information from an environmental perception sensor from a multimodal data acquisition module;

[0037] Based on the image and information, perform feature extraction and fusion to generate a unified fusion feature representation;

[0038] Based on the surrounding environment information and preset material properties, quantitative physical simulation is performed to predict the deformation mode of an ideal round cigarette and output the predicted deformation profile and physical guidance signal;

[0039] Based on the fusion feature representation and physical guidance signals, the initial position and range of the cigarette instance are proposed to generate instance proposals;

[0040] Adopting a generative adversarial network structure, combined with fusion feature representation, instance proposals, and physical guidance signals, an adaptive and refined instance representation of each cigarette instance is generated through adversarial training.

[0041] The actual detection contour is extracted from the adaptively refined instance representation, the roundness error of the cigarette end face is calculated, and the uncertainty of the roundness error is quantitatively evaluated.

[0042] Compare the shape similarity between the actual detected contour and the predicted deformed contour, infer the cause of the roundness error as external extrusion, manufacturing defects, or mixed reasons, and output the deformation cause classification label and corresponding confidence level;

[0043] Aggregate roundness error, uncertainty, deformation cause classification label and confidence, and output the detection results.

[0044] Generating an adaptive refined instance representation includes adaptively selecting one or a combination of output parameterized circle / ellipse parameters, spline curves defined by control points, local implicit field function parameters, or edge probability maps according to an evaluated instance complexity index.

[0045] The technical solution of the embodiment of the present invention uses an industrial camera to simultaneously capture RGB images of the end face of a cigarette at multiple viewing angles, and then combines the polarized light or near-infrared / hyperspectral images obtained by the additional modality acquisition unit, and the three-dimensional spatial distribution information provided by the environmental perception sensor, so that the system can obtain rich information from multiple angles and multiple modalities; the synchronous trigger controller ensures that all sensor data are strictly synchronized in time, and the parameters in the calibration data memory ensure the consistency of geometric correction and color correction, thereby eliminating data distortion caused by differences between sensors and improving the integrity and accuracy of the overall data. The feature extraction and fusion module uses the pre-trained convolutional neural network in the feature extraction unit to extract deep features for images of different modalities and viewing angles, and then uses the cross-modal and cross-view attention mechanism and geometric alignment method through the feature fusion unit to effectively fuse the features of each channel to generate a unified fusion feature representation; this fusion method not only improves the robustness in low light, noise or complex backgrounds, but also provides high-quality and stable feature data for subsequent physical simulation and example proposals. The physical simulation and guidance module uses the simulation execution unit to quantitatively calculate the force and deformation of cigarettes in actual environments using the finite element method, boundary element method, or spring-mass model to obtain a predicted deformation contour. The guidance signal generation unit generates a physical guidance signal based on the simulation results, allowing the instance proposal module to fully utilize real physical information when proposing cigarette instances, significantly improving the accurate positioning of the detection area and the reliability of subsequent shape refinement. The generator unit in the GAN shape refinement module adaptively selects the output form by evaluating the instance complexity index and generates an adaptively refined instance representation. At the same time, the discriminator unit and loss calculation unit use an adversarial training mechanism to ensure that the generated instance representation is highly consistent with the characteristics of the real cigarette in terms of shape, which can both adapt to the complexity of the cigarette shape and maintain the rationality of the physical simulation results. The roundness calculation and uncertainty module uses the contour processing unit to extract the three-dimensional edge contour from the instance representation, and the shape fitting unit fits the optimal circle or ellipse shape. The metric calculation unit then calculates the roundness error of the cigarette end face. At the same time, the uncertainty quantification unit comprehensively utilizes the confidence output by the generator, the fitting residual, and the conformity information between the actual detection contour and the predicted deformation contour to quantify the uncertainty of the detection result, providing a reliable statistical basis for the detection result. In the deformation cause inference module, the shape comparison unit quantitatively evaluates the degree of cigarette deformation by comparing the shape similarity of the actual detection contour and the predicted deformation contour. The cause classification unit combines shape similarity, physical prediction results and actual roundness error, as well as environmental geometry information, and uses preset rules or machine learning classifiers to accurately determine the cause of deformation. The confidence scoring unit outputs the corresponding confidence, providing strong data support for subsequent process improvement and troubleshooting.The result output module integrates the data output by each key module and intuitively displays the test results through graphic overlays, data reports or real-time monitoring interfaces, facilitating real-time monitoring and automated quality control of the production line. The efficient collaboration and data transmission between modules enable the entire test process, from data collection and feature processing to final result output, to be automated and real-time, significantly improving production efficiency and test accuracy while reducing the defective rate.

[0046] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 This is a diagram of the architecture of a cigarette end face visual inspection device provided by an embodiment of the present invention;

[0049] Figure 2 A diagram showing the steps of a device for visually inspecting cigarette ends provided by an embodiment of the present invention;

[0050] Figure 3 This is a flow chart of a cigarette end face visual inspection device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0052] It should be noted that the terms "predicted," "actual," and the like in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0053] Example 1

[0054] Figure 1 This is a diagram of the architecture of a cigarette end face visual inspection device provided by an embodiment of the present invention; Figure 2 A diagram showing the steps of a device for visually inspecting cigarette ends provided by an embodiment of the present invention; Figure 3 A flowchart of a cigarette end face visual inspection device provided by an embodiment of the present invention is shown in FIG. Figure 1 、 Figure 2 、 Figure 3 As shown, this embodiment discloses a device for visual inspection of cigarette end faces, which primarily includes an industrial camera, a light source system, and a control system. The control system integrates multiple functional modules, including a feature extraction and fusion module 100, a physical simulation and guidance module 200, an example proposal module 300, a GAN shape refinement module 400, a roundness calculation and uncertainty module 500, a deformation cause inference module 600, and a result output module 700. These modules work together to implement the entire process from multi-view, multi-modal image acquisition to cigarette end face inspection result output.

[0055] Feature extraction unit 110 uses a pre-trained convolutional neural network (CNN) to process images from various modalities and perspectives captured by industrial cameras to extract depth feature maps. To improve robustness, this unit can employ a multi-level feature fusion strategy to extract both low-level texture features and high-level semantic features.

[0056] The feature fusion unit 120 is based on the cross-modal and cross-view attention mechanism, and combines the industrial camera calibration parameters for geometric transformation alignment. It adopts multiple fusion strategies (such as weighted fusion, feature splicing or adaptive fusion algorithm) to fuse the depth feature maps of each channel into a unified fusion feature representation.

[0057] The network structure can use residual network or DenseNet as the backbone to ensure the depth and breadth of feature extraction; during the fusion process, the attention mechanism is used to automatically adjust the weights of different perspectives and modalities to ensure that effective information can still be extracted in low light or noise interference conditions.

[0058] Based on ambient environmental information (e.g., temperature, humidity, external pressure) and pre-set material mechanical properties (e.g., elastic modulus, Poisson's ratio), simulation execution unit 210 employs finite element method (FEM) approximation, boundary element method (BEM), or a spring-mass model, or even incorporates a neural network proxy model, to quantitatively calculate the deformation field of an ideal circular cigarette under the current environment. The resulting deformation result is represented as a predicted deformation profile 201 in the form of coordinate points or a contour curve.

[0059] The guidance signal generating unit 220 generates a physical guidance signal 202 based on the calculation results of the simulation executing unit 210. The signal may include a stress concentration map, a contact probability map, or a predicted displacement direction map, which serves as an important reference for subsequent example proposal and refinement.

[0060] Boundary conditions and local material heterogeneity are introduced into the simulation algorithm to ensure the accuracy of the predicted deformation profile 201; the guidance signal 202 uses a heat map or probability distribution map to intuitively reflect the physical state and provide quantitative information for subsequent modules.

[0061] The instance proposal module 300 uses the unified fused feature representation output from the feature extraction and fusion module 100 and the physical guidance signal 202 output from the physical simulation and guidance module 200 to automatically propose the initial position and range of the cigarette instance.

[0062] The instance proposal algorithm can use a region proposal network (RPN) or a deep learning method based on candidate regions to preliminarily locate the area of ​​the cigarette end face in the image.

[0063] A multi-scale candidate region generation strategy is adopted to ensure accurate proposals even when cigarettes are of different sizes. At the same time, the candidate regions are optimized in combination with physical guidance signals 202 to improve positioning accuracy.

[0064] Generator unit 410 processes the input fused feature representation and instance proposal results. It first evaluates the input features or internal features to obtain an instance complexity index, which measures the degree of adhesion, predicted deformation amplitude, and shape irregularity of the cigarette instance. Based on this index, generator unit 410 adaptively selects output parameters, including parameters of parameterized circles / ellipses, spline curves defined by control points, parameters of local implicit field functions, or edge probability maps, to generate an adaptively refined instance representation 401 for each cigarette instance.

[0065] The discriminator unit 420 receives the generated adaptively refined instance representation 401 or its derived features and judges its similarity or authenticity with the real cigarette shape features as part of adversarial training.

[0066] The loss calculation unit 430 calculates the adversarial loss used to promote the generation of realistic shapes, the reconstruction loss to ensure that the generated shapes are consistent with the target, and the physical consistency loss to penalize the significant deviation of the generated shapes from the physical simulation predictions, thereby jointly optimizing the network performance.

[0067] The generator unit 410 adopts a multi-branch design, with different branches targeting parameterized models, spline curves, and implicit field models respectively, and finally fusion generates an adaptive refined instance representation 401; a gradient penalty term is introduced during adversarial training to stabilize the training process and ensure that the discriminator unit 420 and the generator unit 410 are balanced.

[0068] The contour processing unit 510 decodes or samples the adaptively refined instance representation 401 generated by the GAN shape refinement module 400 to obtain a 3D edge contour point set. A robust plane fitting algorithm is used to determine the best-fitting plane, and the 3D contour points are orthogonally projected onto the plane to obtain a 2D point set of the actual detected contour 501.

[0069] The shape fitting unit 520 applies a robust circle or ellipse fitting algorithm to the two-dimensional point set of the actual detected contour 501 to obtain a best fitting shape.

[0070] The metric calculation unit 530 calculates the roundness error 502 of the cigarette end face based on the fitting results and the two-dimensional point set. This error can be obtained through various metric methods such as radius standard deviation, minimum / maximum radius ratio, or ratio of the contour inner area to the area of ​​the fitted circle / ellipse.

[0071] The uncertainty quantification unit 540 combines the confidence output by the GAN shape refinement module 400, the fitting residual of the shape fitting unit 520, and the conformity information between the actual detected contour 501 and the predicted deformed contour 201 to calculate and output the roundness error uncertainty 503.

[0072] Robust statistical methods are used to evaluate fitting residuals to reduce the impact of outliers; uncertainty quantification integrates multiple indicators to ensure that the output results have a high degree of credibility.

[0073] The shape comparison unit 610 quantifies the shape similarity between the actual detected contour 501 and the predicted deformed contour 201 by shape descriptor distance calculation, deformation field matching analysis, or contour curvature distribution correlation analysis (or a combination thereof).

[0074] The cause classification unit 620 uses preset rule logic or pre-trained machine learning classifiers based on shape similarity, physically predicted deformation degree, actual roundness error 502 and geometric features of the adjacent environment to determine the cause of the roundness error and output a deformation cause classification label 601 (such as external extrusion-dominated, manufacturing defect-dominated or mixed causes).

[0075] The confidence scoring unit 630 calculates and outputs the corresponding confidence 602 based on the shape similarity and the strength of the judgment basis of the cause classification unit 620; the classification unit 620 can use a support vector machine (SVM), decision tree or neural network classifier to ensure multi-factor comprehensive judgment; the confidence scoring unit 630 combines statistical methods to quantify the credibility of the output results, making the detection results more valuable for reference.

[0076] The result output module 700 is responsible for aggregating the key detection results output by the aforementioned modules, including roundness error 502 , uncertainty 503 , deformation cause classification label 601 and confidence 602 .

[0077] The output results can be displayed in the form of image overlay, data report or real-time monitoring interface, which is convenient for subsequent quality control and process optimization.

[0078] The specific steps are as follows:

[0079] Data acquisition: The industrial camera collects multi-view, multi-modal images of the cigarette end face, and after being evenly illuminated by the light source system, the image is transmitted to the control system.

[0080] Feature extraction and fusion: The image data is first extracted by the feature extraction unit 110 using CNN to extract the deep feature map; then, the feature fusion unit 120 combines the cross-modal attention mechanism and geometric alignment method to fuse the features of each perspective into a unified fusion feature representation, which is output by the feature extraction and fusion module 100.

[0081] The features extracted from each modality and each view image are F i ∈R H×W×C (i=1,2,…N), the corresponding sensor calibration parameter is P i Where H is the vertical size of the image, that is, the number of pixel rows from top to bottom. W is the horizontal size of the image, that is, the number of pixel columns from left to right. C is the number of channels of the image, which represents the dimension of the color information in the image. The proposed multimodal fusion formula is:

[0082]

[0083] Among them, T(F i ,P i ) represents the feature reconstruction after spatial transformation and geometric alignment of features using correction parameters, αi is the weight calculated by the cross-modal attention mechanism.

[0084]

[0085] Here, φ(·) is a nonlinear mapping function composed of a multi-layer perceptron, which is used to capture the correlation between the modalities.

[0086] Physical simulation and guidance; the control system obtains the surrounding environment information and preset material properties, and the simulation execution unit 210 quantitatively calculates the deformation field of the ideal circular cigarette and outputs the predicted deformation profile 201; the guidance signal generation unit 220 generates a physical guidance signal 202 based on the simulation results to provide a reference for subsequent example proposals.

[0087] Based on environmental information and material properties, a formula for predicting the cigarette deformation field is designed. Assuming that at the plane coordinate (x, y), the environmental impact map is E(x, y), the material property mapping is M(x, y), the boundary constraint is B(x, y), and a small constant ∈ is introduced to avoid division by zero. The predicted deformation field D(x, y) is defined as:

[0088]

[0089] Where η, δ, and ξ are trainable parameters and ⊙ represents element-wise multiplication. The predicted deformation contour is extracted from D(x, y) through nonlinear mapping.

[0090] Instance proposal; the instance proposal module 300 automatically proposes the initial position and range of the cigarette instance based on the fused feature representation and the physical guidance signal 202, and determines the area to be detected.

[0091] GAN shape refinement: input the instance proposal results and fusion features into the GAN shape refinement module 400;

[0092] The generator unit 410 adaptively outputs an adaptively refined instance representation 401 according to the instance complexity index; the discriminator unit 420 judges the authenticity of the output result, and guides adversarial training through the loss calculation unit 430 to gradually improve the representation accuracy.

[0093] To adaptively generate refined instance representations, a new generator loss function is designed that integrates adversarial, reconstruction, and physical consistency constraints:

[0094] L G =λ1L adv +λ2L rec +λ3L phy +λ3‖C-C0‖ 2 ;

[0095] Among them, L advTo combat the loss, L rec To reconstruct the loss, L phy is the physical consistency loss, which is used to constrain the generated instance shape to be consistent with the physical simulation result, ‖C-C0‖ 2 is the penalty term based on the instance complexity index, C is the actual calculated complexity, C0 is the target complexity, and λ i is the weight hyperparameter of each loss term.

[0096] Roundness calculation and uncertainty assessment; the contour processing unit 510 in the roundness calculation and uncertainty module 500 extracts a three-dimensional edge contour point set from the refined instance representation 401, and orthogonally projects it to obtain the actual detection contour 501; the shape fitting unit 520 fits the optimal circle or ellipse shape, and the metric calculation unit 530 calculates the roundness error 502 based on this; at the same time, the uncertainty quantification unit 540 combines multiple indicators to calculate the uncertainty 503 of the roundness error.

[0097] The actual detection contour obtained from the refined instance representation is a two-dimensional point set The center of the fitted circle is c and the radius is r, then the roundness error E circ Calculated as:

[0098]

[0099] Among them, the uncertainty quantification formula combines the fitting residual, GAN confidence Υ GAN and physical consistency error E phy .

[0100] U=ω1E circ +ω2(1-Υ GAN )+ω3E phy ;

[0101] Among them, ω i is the adaptive weight, through ψ i (X) is a learnable function that takes the feature representation X as input.

[0102] Deformation cause inference; the shape comparison unit 610 in the deformation cause inference module 600 compares the actual detected contour 501 with the predicted deformation contour 201 to quantify the shape similarity; the cause classification unit 620 determines the deformation cause based on information such as shape similarity, physical prediction and actual roundness error 502, and outputs a deformation cause classification label 601; the confidence scoring unit 630 calculates the confidence 602 of the determination result.

[0103] In order to accurately determine the cause of cigarette deformation, a new determination formula that combines shape similarity and error measurement is introduced. Let shape similarity be S shape, the physical consistency error is E phy , and the roundness error is E circ , then the comprehensive score S cause Defined as:

[0104]

[0105] According to S cause The relationship between the adaptive thresholds T1 and T2 is as follows:

[0106]

[0107] Among them, κ i is the weight parameter, T1 and T2 are trainable threshold parameters.

[0108] Result output; Finally, the result output module 700 aggregates the roundness error 502, uncertainty 503, deformation cause classification label 601 and confidence 602, and outputs them in the form of graphics and data reports for subsequent quality inspection and process optimization.

[0109] By using CNN and cross-modal attention mechanism in the feature extraction and fusion module 100, robust feature extraction under multi-view and multi-modal conditions is achieved, significantly improving the detection accuracy.

[0110] In the physical simulation and guidance module 200, the cigarette deformation field is quantitatively calculated by the simulation execution unit 210, and the guidance signal generation unit 220 outputs the physical guidance signal 202, so that the instance proposal module 300 can accurately locate the cigarette instance, thereby improving the detection efficiency.

[0111] The GAN shape refinement module 400 adopts an adversarial training mechanism of a generator unit 410 and a discriminator unit 420 to effectively generate an adaptively refined instance representation 401, ensuring a high degree of consistency between the detection results and the true shape; the roundness calculation and uncertainty module 500 can accurately calculate the roundness error 502 through a refined contour processing unit 510 and a shape fitting unit 520, and combine multiple information to achieve uncertainty quantification 503, thereby improving the credibility of the detection results.

[0112] The deformation cause inference module 600 comprehensively utilizes the shape comparison unit 610 and the cause classification unit 620 to accurately determine the cause of cigarette deformation (such as external extrusion, manufacturing defects or mixed causes), and provides a reliable confidence level 602 through the confidence scoring unit 630, which facilitates subsequent defect analysis and process improvement.

[0113] Finally, the result output module 700 integrates and outputs various key indicators, providing comprehensive and accurate test data for cigarette production quality control, which helps to improve product consistency and reduce defective rates.

[0114] This solution realizes the full process of automated inspection from image acquisition, feature extraction, physical simulation, instance proposal, shape refinement, roundness calculation to deformation cause inference. It has the advantages of high precision, high robustness and good scalability, and can effectively adapt to the needs of cigarette end face inspection in complex production environments.

[0115] Example 2

[0116] Figure 1 This is a diagram of the architecture of a cigarette end face visual inspection device provided by an embodiment of the present invention; Figure 2 A diagram showing the steps of a device for visually inspecting cigarette ends provided by an embodiment of the present invention; Figure 3 A flowchart of a cigarette end face visual inspection device provided by an embodiment of the present invention is shown in FIG. Figure 1 、 Figure 2 、 Figure 3 As shown, the system's sensor configuration and data acquisition scheme have been further expanded. In addition to the integrated feature extraction and fusion module 100, physical simulation and guidance module 200, example proposal module 300, GAN shape refinement module 400, roundness calculation and uncertainty module 500, deformation cause inference module 600, and result output module 700, the following key hardware units have been added: The environmental perception sensor 15, which uses at least one of a 3D structured light scanner or a laser ranging unit, collects spatial distribution information of adjacent cigarettes and provides this data to the physical simulation and guidance module 200.

[0117] The additional modality acquisition unit is selected from at least one of the polarization light acquisition unit 12 or the near-infrared / hyperspectral imaging unit 13, and is used to obtain supplementary modality information and provide it to the feature extraction and fusion module 100 to make up for the limitations that may exist in a single modality.

[0118] The synchronous trigger controller 14 ensures that the industrial camera, the environmental perception sensor 15 and the additional modality acquisition unit are strictly synchronized during data acquisition, ensuring the temporal consistency of the multi-modal and multi-view data.

[0119] The calibration data memory 16 stores the calibration parameters of each sensor (including the industrial camera 11, the environmental perception sensor 15 and the additional modality acquisition unit) to support subsequent data alignment and geometric transformation correction.

[0120] Industrial camera 11, this embodiment uses at least three cameras 11, distributed at different spatial perspectives, to obtain rich multi-angle image information, providing comprehensive data support for subsequent feature extraction and physical simulation.

[0121] The data acquisition system is enhanced with at least three industrial cameras 11 arranged in different spatial positions. Through pre-calibration, the viewing angles of each camera are ensured to be complementary, providing comprehensive images of the cigarette end faces. The image resolution and frame rate can be adjusted according to the requirements of the production line to ensure real-time acquisition and high-precision detection.

[0122] The environmental perception sensor 15 uses a three-dimensional structured light scanner or a laser ranging unit: the sensor captures the three-dimensional spatial information of the cigarette and its surrounding environment by emitting structured light or laser beams, and outputs information including a depth map and spatial coordinates; the environmental perception sensor 15 transmits the collected spatial distribution data to the physical simulation and guidance module 200 in real time for correcting the simulation boundary conditions and estimating the impact of the surrounding environment.

[0123] An optional polarized light acquisition unit 12 is used to capture the subtle reflective characteristics of the cigarette surface; or a near-infrared / hyperspectral imaging unit 13 is used to identify material composition and defect information.

[0124] The supplementary image data collected by the additional modality acquisition unit and the RGB image collected by the industrial camera 11 are transmitted to the feature extraction and fusion module 100 for joint feature extraction after ensuring time consistency through the synchronization trigger controller 14.

[0125] The synchronous trigger controller 14 strictly aligns the data collection time of the industrial camera 11, the environmental perception sensor 15 and the additional modality acquisition unit through the hardware clock or trigger signal, ensuring that the data of each modality are collected at the same time point and eliminating the time error in dynamic scenes; it supports high-speed data transmission protocols to meet real-time requirements.

[0126] The calibration data storage device 16 stores the internal parameters and external calibration data of each sensor (industrial camera 11, environmental perception sensor 15, additional modality acquisition unit); Data correction: In the feature extraction and fusion module 100, the parameters in the calibration data storage device 16 are used to perform geometric alignment and color correction on the data collected by different sensors to ensure the consistency and accuracy of the multimodal data.

[0127] By adding the environmental perception sensor 15 and the additional modality acquisition unit, the information acquisition capability of the system is effectively expanded.

[0128] In the feature extraction and fusion module 100, the supplementary image from the additional modality acquisition unit is used and input into the feature extraction unit 110 together with the multi-view RGB image acquired by the industrial camera 11. The cross-modal attention mechanism is used in conjunction with the geometric transformation alignment based on the calibration data memory 16 correction parameters to further improve the quality of the fused feature representation generated by the feature fusion unit 120.

[0129] In addition to using the image data obtained by the industrial camera 11, the physical simulation and guidance module 200 uses the spatial distribution information of adjacent cigarettes provided by the environmental perception sensor 15 to further improve the simulation execution unit 210's modeling of the boundary conditions and external environmental factors for deformation field calculation, so that the predicted deformation profile 201 is more in line with the actual situation.

[0130] The specific steps are as follows:

[0131] During the data acquisition stage, the industrial camera 11, under the control of the synchronous trigger controller 14, simultaneously collects high-resolution image data of the cigarette end face from different perspectives; the additional modal acquisition unit synchronously collects polarized light or near-infrared / hyperspectral images to provide modal information complementary to the RGB image; the environmental perception sensor 15 simultaneously performs three-dimensional spatial scanning of the cigarettes and their surroundings in the production area to obtain depth data and spatial distribution information; the synchronous trigger controller 14 ensures that all sensors collect data at the same time and transmits the data to the control system through a high-speed data interface.

[0132] During the data preprocessing and correction stage, the calibration parameters of each sensor pre-stored in the calibration data memory 16 are loaded and used to perform color correction and geometric alignment on the collected image and spatial data; after preprocessing, the multimodal data enters the feature extraction and fusion module 100 for deep feature extraction and fusion processing.

[0133] During the physical simulation and guidance stage, the spatial distribution information output by the environmental perception sensor 15 is transmitted to the physical simulation and guidance module 200, and the environmental parameters in the simulation execution unit 210 are corrected to ensure accurate modeling of the deformation field of the cigarette in the real production environment; the guidance signal generation unit 220 generates a physical guidance signal 202 based on the corrected simulation results, providing more accurate guidance information for subsequent instance proposals.

[0134] In the subsequent processing stage, the remaining modules (instance proposal module 300, GAN shape refinement module 400, roundness calculation and uncertainty module 500, deformation cause inference module 600 and result output module 700) perform data processing, refined shape generation, roundness calculation, uncertainty quantification, deformation cause inference and result output according to the process in the above embodiment.

[0135] The supplementary image collected by the additional modality acquisition unit and the RGB image collected by the industrial camera 11 are strictly time-aligned by the synchronous trigger controller 14, and then jointly processed by the feature extraction and fusion module 100, so that the generated fusion feature representation is more comprehensive, significantly improving the integrity of the image information and the ability to capture details; the three-dimensional spatial distribution information obtained by the environmental perception sensor 15 provides accurate environmental data for the physical simulation and guidance module 200, so that the simulation execution unit 210 can fully consider the influence of the surrounding environment when calculating the cigarette deformation field, and predict the deformation contour 201 more accurately; the synchronous trigger controller 14 ensures that the industrial camera 11, the additional modality acquisition unit and the environmental perception sensor 15 collect data at the same time point, eliminating the errors caused by timing differences and improving the robustness of subsequent processing; the calibration parameters of each sensor stored in the calibration data storage 16 support the precise alignment and correction of multi-modal data, effectively reducing the system error caused by parameter mismatch between different sensors; combining multi-viewpoint, multi-modal and environmental information, the system can perform more accurate and comprehensive detection of cigarette end faces in complex production environments, improving the defect recognition rate and product quality control level.

[0136] In summary, by adding environmental perception sensors 15, additional modal acquisition units, synchronous trigger controllers 14 and calibration data storage devices 16, the data acquisition means are further enriched, the multimodal data fusion capability and physical simulation accuracy are enhanced, thereby achieving higher accuracy, higher robustness and better detection efficiency in the cigarette end face detection process, and providing reliable technical support for real-time monitoring and quality control of the production line.

[0137] Example 3

[0138] Figure 1 This is a diagram of the architecture of a cigarette end face visual inspection device provided by an embodiment of the present invention; Figure 2 A diagram showing the steps of a device for visually inspecting cigarette ends provided by an embodiment of the present invention; Figure 3 A flowchart of a cigarette end face visual inspection device provided by an embodiment of the present invention is shown in FIG. Figure 1 、 Figure 2 、 Figure 3 As shown in the figure, a complete set of cigarette end face visual inspection methods is constructed through multimodal data acquisition, deep feature extraction and fusion, quantitative physical simulation, instance proposal, GAN-based shape refinement, roundness calculation and uncertainty assessment, deformation cause inference and result output.

[0139] The following steps are used to detect cigarette end faces:

[0140] S1 data acquisition, at least three industrial cameras 11 are distributed at different spatial perspectives, and real-time RGB images of the cigarette end face are collected to ensure that complete image information is obtained from multiple angles; according to actual needs, a polarized light acquisition unit 12 is selected to collect the reflection characteristics of the cigarette surface, or a near-infrared / hyperspectral imaging unit 13 is used to obtain material composition and defect information; a three-dimensional structured light scanner or a laser ranging unit is used to obtain the three-dimensional spatial distribution information of the cigarette and the surrounding environment; ensure that the industrial camera 11, the additional modal acquisition unit and the environmental perception sensor 15 collect data synchronously at the same time point to eliminate timing errors; during the data acquisition process, the data collected by each sensor is simultaneously transmitted to the control system, and preliminary correction and geometric alignment are completed based on the calibration parameters of each sensor stored in the calibration data storage 16.

[0141] S2 feature extraction and fusion, the control system inputs the multimodal, multi-view image data obtained in S1 into the feature extraction and fusion module 100: the feature extraction unit 110 uses a pre-trained convolutional neural network (CNN) to process each modality image and extract deep feature maps respectively; the feature fusion unit 120 adopts a cross-modal and cross-view attention mechanism, combined with the geometric correction parameters provided in the calibration data memory 16, to fuse the feature maps from different sensors, generate a unified and information-rich feature representation, and provide high-quality input for subsequent steps.

[0142] S3 quantitative physical simulation, based on the spatial distribution information output by the environmental perception sensor 15 and the preset material properties, the physical simulation and guidance module 200 is started: the simulation execution unit 210 uses the finite element method (FEM), boundary element method (BEM) or spring mass model to simulate the stress and deformation of an ideal circular cigarette in the current environment, calculates the deformation field of the cigarette under different boundary conditions, and generates a predicted deformation contour 201; the guidance signal generation unit 220 outputs a physical guidance signal 202 in the form of a stress concentration map, a contact probability map or a predicted displacement direction map according to the simulation results, providing a reference basis for subsequent instance proposals and shape refinement.

[0143] S4 instance proposal: The instance proposal module 300 uses the fused feature representation in S2 and the physical guidance signal 202 in S3 to automatically determine the initial position and range of the cigarette instance through a region proposal algorithm (such as based on the candidate region network RPN) and delineate the area to be detected.

[0144] S5GAN shape refinement, after obtaining the preliminary proposal result, the system inputs the fused feature representation, instance proposal result and physical guidance signal 202 into the GAN shape refinement module 400: the generator unit 410 first evaluates the input deep features and regional information, and calculates the instance complexity index, which measures the degree of adhesion, predicted deformation amplitude and shape irregularity of the cigarette instance; according to the instance complexity index, the generator unit 410 adaptively selects the output form, and can output one or a combination of parameterized circle / ellipse parameters, spline curves defined by control points, local implicit field function parameters or edge probability maps to generate an adaptive refined instance representation 401; the discriminator unit 420 makes a authenticity judgment on the generated adaptive refined instance representation 401, and combines the loss calculation unit 430 to calculate the adversarial loss, reconstruction loss and physical consistency loss, and continuously optimizes the output of the generator unit 410 through adversarial training, so that the adaptive refined instance representation 401 is closer to the real cigarette shape.

[0145] S6 roundness calculation and uncertainty assessment: The adaptive refined instance representation 401 output from S5 is further processed using the roundness calculation and uncertainty module 500:

[0146] The contour processing unit 510 decodes or samples a three-dimensional edge contour point set from the instance representation 401, and uses a robust plane fitting algorithm to determine the best fitting plane, and orthogonally projects the three-dimensional point set to obtain a two-dimensional point set of the actual detection contour 501; the shape fitting unit 520 applies a robust circle or ellipse fitting algorithm to the two-dimensional point set to obtain the best fitting shape; the metric calculation unit 530 calculates the roundness error 502 of the cigarette end face (such as the radius standard deviation, the minimum / maximum radius ratio or the contour area ratio) based on the fitting results and the two-dimensional point set; the uncertainty quantification unit 540 combines the confidence, fitting residual and the conformity information of the actual detection contour 501 and the predicted deformed contour 201 output by the GAN shape refinement module 400 to calculate the uncertainty 503 of the roundness error.

[0147] S7 Deformation cause inference. To determine the cause of deformation of the cigarette end face, the system calls the deformation cause inference module 600: the shape comparison unit 610 uses methods such as shape descriptor distance, deformation field matching and contour curvature distribution to quantify the shape similarity between the actual detected contour 501 and the predicted deformed contour 201; the cause classification unit 620 uses preset rule logic or machine learning classifiers based on shape similarity, physical predicted deformation degree, actual roundness error 502 and adjacent environmental geometric information to determine the cause of the roundness error (such as external extrusion, manufacturing defects or mixed causes) and output the deformation cause classification label 601; the confidence scoring unit 630 calculates and outputs the confidence 602 of the determination result based on the shape similarity and the strength of the classification basis.

[0148] S8 outputs the test results. Finally, the result output module 700 integrates the roundness error 502, uncertainty 503, deformation cause classification label 601 and confidence 602 to form the test result data; the results can be intuitively displayed through graphic overlay display, data report or real-time monitoring interface for subsequent process optimization and quality control.

[0149] Through the collaborative work of the industrial camera 11, the additional modal acquisition unit and the environmental perception sensor 15, under the guarantee of the synchronous trigger controller 14, real-time acquisition of multi-perspective, multi-modal and three-dimensional environmental data is achieved, so that subsequent feature extraction and physical simulation have more comprehensive information support.

[0150] The feature extraction and fusion module 100 utilizes advanced CNN and attention mechanisms, combined with the correction parameters provided by the calibration data memory 16, to achieve accurate fusion of cross-modal and cross-view data. The generated fused feature representation provides high-quality input for subsequent steps.

[0151] The physical simulation and guidance module 200 uses the simulation execution unit 210 to quantitatively calculate the cigarette deformation field, and outputs a physical guidance signal 202 in combination with environmental information, so that the predicted deformation profile 201 is more consistent with the actual production environment.

[0152] The generator unit 410 in the GAN shape refinement module 400 adaptively selects the output form according to the instance complexity index, realizes the generation of adaptive refined instance representation 401, and ensures that the shape refinement process can adapt to shape complexity and meet physical consistency requirements.

[0153] The roundness calculation and uncertainty module 500 uses multi-level contour processing and shape fitting algorithms to accurately calculate the roundness error 502 and provides quantitative uncertainty 503 through the uncertainty quantification unit 540, thereby improving the reliability of the detection results.

[0154] The deformation cause inference module 600 can accurately determine the cause of external extrusion, manufacturing defects or mixing through comprehensive shape comparison and machine learning classification, and output the deformation cause classification label 601 and confidence level 602 to provide a basis for subsequent process improvements.

[0155] Data transmission and processing between modules are highly coordinated, and the test results are finally integrated through the result output module 700, providing efficient and accurate technical support for real-time monitoring of the production line, quality control and improvement of product consistency.

[0156] In summary, the coordinated processing of these steps enables a fully automated inspection process, from multimodal data acquisition, feature extraction, physical simulation, example proposal, shape refinement, roundness calculation and uncertainty assessment, to deformation cause inference and test result output. This method not only offers high precision and robustness, but also responds in real time in complex production environments, providing a comprehensive and efficient solution for cigarette end-face defect detection and quality control.

[0157] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved. This is not limited herein.

[0158] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A cigarette end face visual inspection device, comprising an industrial camera, a light source system and a control system, characterized in that: The control system includes: A feature extraction and fusion module, configured to extract and fuse depth features based on multi-view, multi-modal images acquired from the industrial camera and the optional additional modality acquisition unit to generate a unified fused feature representation; A physical simulation and guidance module, which is used to perform quantitative physical simulation based on the estimation or perception of the surrounding environment and preset material properties to predict the deformation mode of an ideal round cigarette in the current environment, and output the predicted deformation profile and physical guidance signal; an instance proposal module, configured to propose an initial position and range of a cigarette instance based on the fused feature representation and the physical guidance signal, and generate an instance proposal; A GAN shape refinement module, configured to employ a generative adversarial network structure, receive the fused feature representation and the instance proposal, and generate an adaptively refined instance representation of each cigarette instance through adversarial training in combination with the physical guidance signal; A roundness calculation and uncertainty module is used to extract the actual detection contour from the adaptive refined instance representation, calculate the roundness error of the cigarette end face, and quantitatively evaluate the uncertainty of the roundness error; wherein the uncertainty is used to describe the accuracy of the roundness error; a deformation cause inference module, configured to compare the shape similarity between the actual detected contour and the predicted deformed contour, infer the cause of the roundness error, and output a deformation cause classification label and a corresponding confidence level; The result output module is used to aggregate the roundness error, uncertainty, deformation cause classification label and confidence, and output them.

2. The cigarette end face visual inspection device according to claim 1, characterized in that: The device further includes: an environmental perception sensor, selected from at least one of a three-dimensional structured light scanner or a laser ranging unit, for acquiring spatial distribution information of adjacent cigarettes and providing the information to a physical simulation and guidance module within the control system; an additional modality acquisition unit, selected from at least one of a polarized light acquisition unit and a near-infrared / hyperspectral imaging unit, for acquiring supplementary modal information and providing the information to a feature extraction and fusion module within the control system; a synchronous trigger controller for ensuring synchronous operation of the industrial camera, the environmental perception sensor, and the additional modality acquisition unit; and a calibration data memory storing calibration parameters of each sensor. The industrial camera includes at least three cameras, and the cameras are arranged at different spatial viewing angles.

3. The cigarette end face visual inspection device according to claim 1, characterized in that: The feature extraction and fusion module includes: A feature extraction unit, configured to extract deep feature maps from images of various modalities and perspectives using a convolutional neural network; A feature fusion unit is used to fuse the deep feature maps to generate the unified fused feature representation by adopting one or more fusion strategies including cross-modal and cross-view attention mechanisms and geometric transformation alignment based on camera calibration parameters.

4. The cigarette end face visual inspection device according to claim 1, characterized in that: The physical simulation and guidance module includes: a simulation execution unit, configured to quantitatively calculate the deformation field of the ideal circular cigarette based on the adjacent environmental information, preset material mechanical properties, and boundary conditions, using one or a combination selected from the group consisting of a finite element method approximation, a boundary element method approximation, a spring mass model, and a neural network proxy model, and generate the predicted deformation profile; The guidance signal generating unit is used to generate the physical guidance signal including the stress concentration map, the contact probability map or the predicted displacement direction map according to the simulation calculation results.

5. The cigarette end face visual inspection device according to claim 1, characterized in that: The generator unit of the GAN shape refinement module is used to: obtain an instance complexity index based on the evaluation of input features or internal features, wherein the instance complexity index is used to measure the degree of adhesion of the instance, predict the deformation amplitude or shape irregularity; According to the instance complexity index, one or a combination of the parameters of the parameterized circle / ellipse, the spline curve defined by the control points, the parameters of the local implicit field function and the edge probability map is adaptively selected as the adaptive refined instance representation.

6. The cigarette end face visual inspection device according to claim 1, characterized in that: The GAN shape refinement module also includes: a discriminator unit, configured to receive the adaptively refined instance representation or its derived features and determine its similarity or authenticity with the shape features of a real cigarette; The loss calculation unit is used to calculate at least an adversarial loss for promoting the generation of realistic shapes, a reconstruction loss for ensuring that the generated shapes are consistent with the target, and a physical consistency loss for penalizing significant deviations between the generated shapes and physical simulation predictions.

7. The cigarette end face visual inspection device according to claim 1, characterized in that: The roundness calculation and uncertainty module includes: a contour processing unit, configured to decode or sample a three-dimensional edge contour point set from the adaptively refined instance representation, determine a best-fitting plane using a robust plane fitting algorithm, and orthogonally project the three-dimensional contour points onto the best-fitting plane to obtain a two-dimensional point set of the actual detected contour; a shape fitting unit for applying a robust circle or ellipse fitting algorithm to the two-dimensional point set of the actual detected contour; a metric calculation unit, configured to calculate the roundness error based on the fitting result and the two-dimensional point set of the actual detected contour, wherein the roundness error is selected from at least one metric selected from the group consisting of a radius standard deviation, a minimum / maximum radius ratio, and a ratio of an inner area of ​​the contour to an area of ​​the fitted circle / ellipse; An uncertainty quantification unit is used to calculate the uncertainty by combining at least two pieces of information selected from the confidence level output by the GAN shape refinement module, the fitting residual of the shape fitting unit, and the degree of conformity between the actual detected contour and the physical simulation prediction.

8. The cigarette end face visual inspection device according to claim 1, characterized in that: The deformation cause inference module includes: a shape comparison unit, configured to quantify the shape similarity between the actual detected contour and the predicted deformed contour by using one or a combination selected from shape descriptor distance calculation, deformation field matching analysis, and contour curvature distribution correlation analysis; a cause classification unit for determining, based on the shape similarity, the degree of physically predicted deformation, the actual roundness error, and the geometric characteristics of the adjacent environment, the deformation cause classification label as being dominated by external extrusion, dominated by manufacturing defects, or a mixed cause through a preset rule logic or a pre-trained machine learning classifier; The confidence scoring unit is used to calculate the confidence based on the shape similarity and the strength of the judgment basis of the deformation cause classification unit.

9. A method for visually inspecting cigarette end faces, characterized in that: include: Receive a multi-view image of the target cigarette end face containing at least RGB information, polarization information or near-infrared / hyperspectral information, and adjacent environment information from an environmental perception sensor from a multimodal data acquisition module; Based on the image and information, perform feature extraction and fusion to generate a unified fusion feature representation; Based on the adjacent environment information and the preset material properties, a quantitative physical simulation is performed to predict the deformation mode of the ideal round cigarette, and a predicted deformation profile and a physical guidance signal are output; Proposing an initial position and range of a cigarette instance based on the fused feature representation and the physical guidance signal, and generating an instance proposal; Adopting a generative adversarial network structure, combining the fused feature representation, instance proposal, and physical guidance signals, an adaptive and refined instance representation of each cigarette instance is generated through adversarial training. Extracting the actual detection contour from the adaptively refined instance representation, calculating the roundness error of the cigarette end face, and quantitatively evaluating the uncertainty of the roundness error; Comparing the shape similarity between the actual detected contour and the predicted deformed contour, inferring that the cause of the roundness error is external extrusion, manufacturing defects, or mixed causes, and outputting a deformation cause classification label and a corresponding confidence level; Aggregate the roundness error, uncertainty, deformation cause classification label and confidence, and output the detection result.

10. The method for visually inspecting cigarette ends according to claim 9, characterized in that: Generating the adaptive refined instance representation includes: According to the evaluated instance complexity index, one or a combination of the output parameterized circle / ellipse parameters, the spline curve defined by the control points, the local implicit field function parameters or the edge probability map is adaptively selected.

Images