Jelly foreign matter deep learning detection method based on multi-mode intelligent sensing data driving

Abstract

The invention provides a jelly foreign matter deep learning detection method based on multi-mode intelligent sensing data driving, and the method comprises the steps: obtaining an original image of jelly, and generating a cross-view light field matrix through a polarized light compensation algorithm and synchronous triggering time sequence control; a cross-view light field matrix is obtained, the attenuation path of each light ray is reversely tracked in the jelly volume direction, and the foreign matter shielding probability is calculated in combination with a sugar degree-absorbance curve in a material library; foreign matter existence probability thermodynamic diagrams are generated in the three physical levels respectively, and a three-dimensional centroid probability field is obtained; and dynamically compressing a three-dimensional centroid probability field analysis time window according to the passing speed of the current production line jelly, comparing misjudgment cases of the same kind of jelly in a historical library, and adjusting a judgment threshold value of a corresponding space level to obtain a real-time judgment instruction set with production line parameter adaptability. Through deep cooperation of multi-mode light field perception and cross-level probability modeling, the technical breakthrough of full-dimensional foreign matter detection of jelly products is realized.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a jelly foreign body deep learning detection method driven by multimodal intelligent sensing data. Background Art

[0002] Jelly foods often face challenges in detecting foreign matter in industrial production due to their transparent texture and strong concealment of foreign matter. Traditional detection methods (such as manual visual inspection and X-ray inspection) suffer from low efficiency, high missed detection rates, and insufficient sensitivity to small-sized foreign matter (such as plastic fragments and metal particles). They also struggle to cope with complex scenarios such as jelly surface reflections and internal bubble interference. With the development of deep learning technology, foreign matter detection methods based on single modality (such as RGB images) have gradually been applied to the food industry, but they still have the following limitations: single feature expression: relying solely on visual information makes it difficult to distinguish foreign matter from internal jelly textures or bubble interference; poor adaptability to dynamic environments: changes in lighting and the diversity of foreign matter materials lead to insufficient model generalization; low small target detection accuracy: traditional models have limited ability to locate tiny foreign matter (such as <0.5mm). Therefore, there is an urgent need for a new detection technology that integrates multimodal data to improve the robustness and accuracy of foreign matter recognition.

[0003] Prior art 1, application number: CN202510474067.7 discloses a food foreign body detection method and system based on deep learning, including an image acquisition module and a foreign body analysis module connected thereto; the image acquisition module is used to acquire RGB images and penetration images of food; the foreign body analysis module inputs the food image corresponding to the RGB image, the food image corresponding to the penetration image, the color features, and the shape features into a foreign body detection model. Although it solves the technical problem of using X-ray images and analyzing foreign bodies therein, the sensitivity of food X-ray images is too low for low-density foreign bodies, making it difficult to identify foreign bodies that meet the current high-quality requirements for food; using the color of the food as one of the judgment criteria and the shape features of the food as one of the judgment criteria, and judging them through the trained foreign body detection model can avoid a large missed detection rate when detecting foreign bodies in food with a single image; however, relying solely on the two-dimensional fusion of RGB + penetration images cannot achieve three-dimensional volume detection (0-8mm depth).

[0004] Prior art 2, application number CN202210458394, discloses an AI-based camera rolling shutter effect detection method and system, comprising: acquiring 3D point cloud data and an RGB image of the same viewing angle and range, and obtaining a depth image based on the 3D point cloud data; dividing the 3D point cloud data into multiple sub-regions, and determining candidate regions 1 based on the sub-regions; finding the depth region corresponding to candidate region 1 in the depth image, and screening candidate regions 2 in each depth region; dividing the depth image into M sub-block regions, and dividing the RGB image into the same sub-block regions; calculating a total rolling shutter effect quantitative index value based on the corresponding sub-block regions in the RGB image and the depth image; and finally obtaining the rolling shutter effect severity of the RGB image. Although accurate and quantitative rolling shutter effect indicators are obtained by comparative analysis of the 3D point cloud data and the RGB image, providing a reference for UAV mapping and making the image rolling shutter effect determination more accurate, the method is susceptible to interference from refraction by translucent gels, resulting in a high signal-to-noise ratio for deep (5-8mm) foreign objects.

[0005] Currently, existing technologies 1 and 2 have problems such as difficulty in three-dimensional penetration detection, difficulty in meeting dynamic interference suppression requirements, and lack of quantitative judgment standards. Therefore, the present invention provides a deep learning detection method for jelly foreign matter based on multimodal intelligent sensing data. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a jelly foreign body deep learning detection method based on multimodal intelligent sensing data, which comprises the following steps:

[0007] The cross-view light field matrix is ​​obtained, and the attenuation path of each ray is traced in reverse along the jelly volume. The probability of foreign object occlusion is calculated by combining the sugar content-absorbance curve in the material library. Heat maps of foreign object presence probability are generated at three physical levels: the surface layer (0mm-2mm), the middle layer (2mm-5mm), and the deep layer (5mm-8mm). This results in a three-dimensional centroid probability field that includes the spatial distribution probability of foreign objects and the confidence level of the material type.

[0008] According to the dynamic compression of the three-dimensional centroid probability field analysis time window of the jelly on the current production line through the speed, compared with the misjudgment cases of similar jellies in the historical library, the judgment threshold of the corresponding spatial level is adjusted to obtain a real-time judgment instruction set with adaptability to production line parameters.

[0009] Optionally, heat maps of the probability of foreign matter presence are generated at three physical levels: the surface layer 0mm-2mm, the middle layer 2mm-5mm, and the deep layer 5mm-8mm.

[0010] Optionally, the process of obtaining the three-dimensional centroid probability field includes the following steps:

[0011] The cross-view light field matrix is ​​used as the input source. Its bottom transmission characteristics contain the light intensity attenuation gradient after polarization compensation, and the side reflection characteristics carry 360° surface curvature information. The cross-view light field matrix uses a ray reverse tracing engine to establish a discretized voxel space along the jelly Z axis. Each voxel node records the light intensity attenuation rate of the bottom camera of the transmission path and the curvature consistency coefficient of the reflection path.

[0012] Based on the sugar content-absorbance curve in the material library, the input parameter is the nominal sugar content of the jelly. The light intensity attenuation rate is converted into a material matching probability, reflecting the degree of deviation between the current voxel optical properties and the nominal formula. The curvature consistency coefficient is also used to calculate the surface deformation probability, which characterizes the reflection phase anomaly caused by the presence of foreign matter. The material matching probability and the surface deformation probability are fused through a hierarchical weighting function to output a comprehensive probability value.

[0013] Probability field clustering is implemented in three-dimensional space. Voxels with an existence probability greater than a preset standard are used as seed points. Connected domains are generated through the 26-neighborhood diffusion algorithm, and the volume weight and material confidence of each connected domain are calculated. For connected domains that meet the volume weight and material confidence, their geometric center is taken as the centroid, and a set of centroid descriptors containing three-dimensional coordinates, existence probability, and material type identification is output to form the final three-dimensional centroid probability field.

[0014] Optionally, the process of outputting the comprehensive probability value includes the following steps:

[0015] The current voxel attenuation rate is distributed matched with the corresponding curve of the nominal sugar content, and the Mahalanobis distance from the standard interval is calculated as the initial deviation. The initial deviation is input into the attenuation rate confidence converter, and the continuous deviation is mapped to the [0,1] interval through the hyperbolic tangent function to generate the material matching probability.

[0016] The voxel curvature consistency coefficient recorded in the cross-view light field matrix is ​​extracted from the phase difference feature of the synchronous reflection image through surface curvature constraint. The theoretical distribution model of the jelly surface curvature in the absence of foreign matter is activated, and the expected value and allowable fluctuation range of the curvature at each position are generated based on the jelly geometric parameters. The curvature consistency coefficient is compared point by point with the theoretical model, and the distribution difference is quantified using the Kolmogorov-Smirnov test. The difference is converted into the probability of surface deformation using an exponential decay function.

[0017] The material matching probability and surface deformation probability are input into the hierarchical weight allocator, which loads the preset weight strategy according to the physical level to which the voxel belongs. The weighted probability values ​​of the two types of material matching probability and surface deformation probability are fused through the geometric mean operator to generate a comprehensive probability value, which simultaneously reflects the degree of abnormality of the voxel in light propagation characteristics and surface morphology.

[0018] Optionally, the light intensity attenuation rate recorded in the cross-view light field matrix is ​​obtained, where the light intensity attenuation rate is the depth gradient characteristic after hierarchical encoding of the polarization-compensated spot sequence; the sugar content-absorbance curve pre-stored in the material library is loaded as a probability conversion benchmark, and each curve corresponds to a standard distribution range of the light attenuation rate at a specific sugar content.

[0019] Optional, surface layer: deformation probability weighting is dominant, 70% surface deformation probability and 30% material matching probability, because foreign matter on the surface can easily cause curvature mutation;

[0020] Middle layer: balanced weight, 50% surface deformation probability and 50% material matching probability, taking into account both surface reflection and internal transmission characteristics;

[0021] Deep layer: Material probability weight is dominant, with 30% surface deformation probability and 70% material matching probability. Deep foreign objects rely more on transmission attenuation detection.

[0022] Optionally, the process of generating material matching probabilities includes the following steps:

[0023] The nominal sugar content curves stored in the material library are activated. The curve shape is defined by the jelly recipe parameters. Each curve contains the multi-dimensional probability distribution interval of the light attenuation rate under standard production conditions.

[0024] The current voxel attenuation rate is aligned with the nominal sugar content curve in multidimensional space, and the Mahalanobis distance of its deviation from the standard distribution is calculated. The Mahalanobis distance is weighted by the inverse matrix of the covariance matrix to eliminate the dimensional differences in the attenuation rate at different sugar contents and quantify the overall deviation of the voxel optical properties from the nominal formula.

[0025] The Mahalanobis distance is input into the attenuation rate confidence converter, and nonlinear compression is implemented through the hyperbolic tangent function; the slope parameter of the hyperbolic tangent function is dynamically adjusted according to the distribution width of the sugar content-absorbance curve.

[0026] Optionally, the process of fusing the probability values ​​using the geometric mean operator includes the following steps:

[0027] The material matching probability and surface deformation probability adjusted by the hierarchical weight allocator complete the weight distribution of the physical layer.

[0028] The weighted two-category probability values ​​are input into the geometric mean operator, and the square root of the product of the two probability quantities is taken; the weight coefficient in the fusion process is bound to the spatial position of the voxel;

[0029] Comprehensive probability generation is used, and the output value simultaneously encodes two types of abnormal features: material deviation in the light propagation dimension and curvature distortion in the surface morphology dimension.

[0030] Optionally, the process of obtaining a real-time determination instruction set with production line parameter adaptability includes the following steps:

[0031] The input source is real-time monitoring data of the jelly speed of the current production line, with its time resolution synchronized with the synchronization trigger phase calibration accuracy. The dynamic time window compression speed parameter is input into the time window compression engine, which dynamically compresses the analysis time of the three-dimensional center of mass probability field through a nonlinear scaling function. The compression factor is calculated using the velocity derivative term in the phase calibration equation.

[0032] Misjudgment case feature extraction: Misjudgment cases stored in the historical database are activated. Each case contains the hierarchical distribution pattern of the three-dimensional centroid probability field, the material type confidence deviation, and the corresponding production line speed parameters. A hierarchical sensitivity matching algorithm is used to select historical cases with similar probability field morphology within the current compressed time window. The similarity threshold is set to the statistical variance of the material matching probability. The cause of misjudgment in the deviation pattern analysis matching case is analyzed as a hierarchical sensitivity coefficient, which quantifies the vulnerability weight of each physical layer in historical misjudgments.

[0033] Hierarchical threshold optimization implements differentiated adjustments to the surface, middle, and deep judgment thresholds of the three-dimensional centroid probability field based on the hierarchical sensitivity coefficient; the adjusted threshold set is injected into the detection pipeline through a streaming update interface, and its update delay is less than the light field matrix generation cycle; multi-dimensional instruction generation integrates dynamic thresholds, time window compression parameters, and hierarchical sensitivity coefficients to generate a real-time judgment instruction set.

[0034] Optionally, at least one vertical camera is arranged at the bottom of the jelly and at least four horizontal cameras are arranged at multiple angles on the sides of the jelly to obtain the original image of the jelly, and a cross-viewing angle light field matrix including the transmission characteristics of the jelly bottom and the 360° reflection characteristics of the side is generated through a polarization light compensation algorithm and synchronous trigger timing control.

[0035] The present invention has achieved a technological breakthrough in the full-dimensional foreign body detection of jelly products through the deep collaboration of multimodal light field perception and cross-level probability modeling. The multi-physical level penetrating detection capability, through the polarization light compensation algorithm and multi-view synchronous light field acquisition, combined with the sugar content-absorbance dynamic modeling, breaks through the two-dimensional limitations of traditional visual inspection. The coupling analysis of the bottom transmission light path and the 360° side reflection light path enables the system to simultaneously have the Fresnel reflection recognition of surface micro-scratches (0-2mm), scattered light distortion detection of the middle layer (2-5mm), and X-ray equivalent attenuation detection of the deep layer (5-8mm), realizing the omnidirectional capture of foreign bodies ≥0.3mm at any position within a depth of 8mm. The spatiotemporal dynamic probability field decision mechanism upgrades the traditional binary judgment to a probability space operation, and the foreign body material confidence (metal / fiber / bubble, etc.) and spatial coordinate probability are calculated in real time through the ray reverse tracing algorithm. Combined with the spatiotemporal window compression technology that is adaptive to the production line speed, the detection resolution is dynamically optimized with the conveyor belt speed, and the false alarm rate is lower than that of the fixed threshold method. A cross-batch self-learning optimization system and dynamic threshold adjustment driven by a historical misjudgment case library form a closed-loop optimization system, enabling the system to adapt to fluctuations in process parameters such as sugar content and gel strength. By mapping production line parameters using a three-dimensional probability field, the system automatically migrates testing standards for products with different formulations, accelerating the convergence of false positives for new product introductions.

[0036] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0037] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0039] Figure 1 This is a flow chart of the jelly foreign body deep learning detection method driven by multimodal intelligent sensing data in Example 1 of the present invention;

[0040] Figure 2 This is a schematic diagram of the principle of the jelly foreign body deep learning detection method driven by multimodal intelligent sensing data in Example 1 of the present invention;

[0041] Figure 3 This is a diagram of the process of generating a cross-viewing light field matrix in Example 2 of the present invention;

[0042] Figure 4This is a diagram of the process of obtaining the three-dimensional centroid probability field in Example 3 of the present invention;

[0043] Figure 5 This is a process diagram for obtaining a real-time determination instruction set with production line parameter adaptability in Example 7 of the present invention. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0045] The terms used in the embodiments of the present application are for the purpose of describing specific embodiments only and are not intended to limit the embodiments of the present application. The singular forms "a", "the" and "the" used in the embodiments of the present application are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0046] When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0047] Example 1: Figure 1 As shown, the embodiment of the present invention provides a jelly foreign body deep learning detection method based on multimodal intelligent sensing data, which includes the following steps:

[0048] S100: Deploy at least one vertical camera at the bottom of the jelly and at least four horizontal cameras at the sides of the jelly to obtain an original image of the jelly. Generate a cross-viewing angle light field matrix that includes the transmission characteristics of the jelly bottom and the 360° reflection characteristics of the side through a polarization compensation algorithm and synchronous trigger timing control.

[0049] S200: Obtain a cross-view light field matrix, trace the attenuation path of each ray in reverse along the jelly volume, and calculate the probability of foreign object occlusion by combining the sugar content-absorbance curve in the material library. Generate foreign object presence probability heat maps at the three physical levels of the surface layer (0-2mm), the middle layer (2-5mm), and the deep layer (5-8mm), and obtain a three-dimensional centroid probability field that includes the spatial distribution probability of foreign objects and the confidence level of the material type.

[0050] S300: According to the dynamic compression of the three-dimensional centroid probability field analysis time window of the jelly passing speed of the current production line, compared with the misjudgment cases of similar jellies in the historical library, the judgment threshold of the corresponding spatial level is adjusted to obtain a real-time judgment instruction set with production line parameter adaptability.

[0051] The working principle and beneficial effects of the above technical solution are as follows: In this embodiment, at least one vertical camera is arranged at the bottom of the jelly and at least four horizontal cameras are arranged at multiple angles on the side of the jelly to obtain the original image of the jelly, and a cross-view light field matrix including the transmission characteristics of the bottom of the jelly and the 360° reflection characteristics of the side is generated through the polarization compensation algorithm and synchronous trigger timing control; secondly, the cross-view light field matrix is ​​obtained, the attenuation path of each light is reversely traced along the volume direction of the jelly, and the probability of foreign body occlusion is calculated in combination with the sugar content-absorbance curve in the material library; heat maps of the probability of foreign body existence are generated at the three physical levels of the surface layer 0-2mm, the middle layer 2-5mm and the deep layer 5-8mm, respectively, to obtain a three-dimensional centroid probability field including the spatial distribution probability of foreign bodies and the confidence of the material type; finally, the three-dimensional centroid probability field analysis time window is dynamically compressed according to the speed of the jelly on the current production line, and the misjudgment cases of similar jellies in the historical library are compared, and the judgment threshold of the corresponding spatial level is adjusted to obtain a real-time judgment instruction set with production line parameter adaptability (for specific principles, please refer to the attached Figure 2 ). The above solution has achieved a technological breakthrough in the full-dimensional foreign body detection of jelly products through the deep collaboration of multimodal light field perception and cross-level probability modeling. The multi-physical level penetrating detection capability, through the polarization light compensation algorithm and multi-view synchronous light field acquisition, combined with the sugar content-absorbance dynamic modeling, has broken through the two-dimensional limitations of traditional visual inspection. The coupling analysis of the bottom transmission light path and the 360° side reflection light path enables the system to simultaneously have the Fresnel reflection recognition of surface micro-scratches (0-2mm), scattered light distortion detection of the middle layer (2-5mm), and X-ray equivalent attenuation detection capabilities of the deep layer (5-8mm), realizing the omnidirectional capture of foreign bodies ≥0.3mm at any position within a depth of 8mm and a spatiotemporal dynamic probability field decision-making mechanism. The construction of the three-dimensional centroid probability field upgrades the traditional binary judgment to a probability space operation, and calculates the foreign body material confidence (metal / fiber / bubble, etc.) and spatial coordinate probability in real time through the ray reverse tracing algorithm. Combined with line speed-adaptive spatiotemporal window compression technology, detection resolution is dynamically optimized with conveyor belt speed, reducing false alarm rates compared to fixed threshold methods. A cross-batch self-learning optimization system, driven by a historical misjudgment case library and dynamic threshold adjustment, creates a closed-loop optimization loop, enabling the system to adapt to fluctuations in process parameters such as sugar content and gel strength. By mapping production line parameters using a three-dimensional probability field, the system automatically migrates testing standards for products with different formulations, accelerating the convergence of false alarms for new product introductions.

[0052] Example 2: Figure 3As shown, based on Example 1, the process of generating a cross-viewing light field matrix provided by the embodiment of the present invention includes the following steps:

[0053] S101: A vertical camera array captures the intensity distribution of transmitted light from the bottom of the jelly to form an initial transmission spot sequence. The transmission spot sequence is separated into the intrinsic polarization component of the sugar and the polarization component scattered by foreign matter by a polarization decoupler. The intrinsic polarization component of the sugar is input into the material feature compensator, which generates a dynamic compensation vector based on the nominal sugar content parameter of the jelly. Nonlinear intensity correction is performed on the transmission spot sequence. The corrected transmission spot retains the polarization anomaly characteristics caused by foreign matter.

[0054] S102: A horizontal camera array receives side reflected light signals and triggers a pulse phase calibrator to achieve sub-millisecond synchronization of multi-camera exposure timing with the movement speed of the jelly production line. The reflected image sequence after synchronous triggering carries 360° surface texture phase information.

[0055] S103: The transmission correction spot and the reflection synchronization image are input into the spatiotemporal registration engine. The spatial correspondence between the transmission light path and the reflection viewing angle is established through the jelly geometric parameter inverse mapping algorithm. The spatial correspondence binds the bottom spot coordinates to the side pixel grid in three-dimensional space.

[0056] S104: The combined dataset after three-dimensional spatial binding is input into the light field reconstructor, and the polarization anomaly intensity on the transmission path and the surface texture phase at the reflection perspective are tensor-fused. During the fusion process, the transmission feature provides the light attenuation gradient in the depth direction, and the reflection feature provides the surface curvature constraint. The fused tensor passes through the multi-resolution light field encoder to establish a radial layered sampling grid along the central axis of the jelly. Each grid cell contains the bottom transmission polarization compensation value and the reflection phase difference at four horizontal perspectives. The final output is a cross-perspective light field matrix with spatial continuity.

[0057] Among them, the spatially bound data is a joint data set formed by aligning the transmission features and the reflection features in three-dimensional space through mapping rules.

[0058] Polarization direction decoupling equation in step S101:

[0059]

[0060] Where, represents the intrinsic polarization component of the sugar body; Ι0(ξ,η) represents the intensity distribution of the original transmitted light spot; Λ represents the spatial response domain of the polarization sensitive element; represents the polarization direction matching kernel function; α, β, γ represent non-integer differential orders; ω k represents the kth order scattered harmonic weight; F krepresents the basis function of the foreign body scattering mode; ξ, η represent the local coordinate system of the polarization sensitive element; ξ-x, η-y represent the spatial offset of the spot intensity distribution; Laplacian operator representing the original spot intensity; represents a non-integer order mixed partial differential operator; N represents the order of the scattering harmonic expansion;

[0061] Dynamic compensation vector generation equation:

[0062]

[0063] Where, represents the dynamic compensation vector; s represents the sugar content parameter; ζ m Represents the material dispersion coefficient matrix; represents the sugar content related quantized loss function; R n represents the nth-order component of the refractive index tensor; Ψ(s) represents the polarization compensation response surface; Represents the expansion operation of the Hadamard product; M represents the dimension of the material dispersion coefficient matrix; Q represents the order of the spatial continuity constraint matrix; ∈ represents the regularization constant of the quantized loss function; represents the material-polarization gradient coupling term; log(Ψ(s) represents the logarithmic transformation of the polarization response surface;

[0064] Step S102: Synchronous trigger phase calibration equation:

[0065]

[0066] Where Δφ j represents the phase calibration value of the j-th horizontal camera; Indicates the instantaneous moving speed of the production line; M j represents the camera field of view motion matching tensor; κ represents the phase lag correction factor; σ j represents the camera timing noise suppression coefficient; represents the overlapping weight of adjacent camera perspectives; t represents the production line operation time; t0 represents the starting time of phase calibration; represents the instantaneous linear velocity function;

[0067] Step S103: Inverse space mapping equation:

[0068]

[0069] Where, represents the transmission-reflection three-dimensional mapping operator; (u, v, w) represents the voxel coordinate system inside the jelly; represents the jelly geometric parameter tensor (diameter, curvature, etc.); v, μ, λ represent fractional differential exponents; represents the distortion field of the transmitted light spot; H k represents the k-th layer reflection phase constraint matrix; B k represents the spatial binding basis function; Ω represents the integral area of ​​the transmitted spot distortion field; Represents the square of the gradient modulus of the spot distortion field; Indicates the maximum allowable distortion threshold;

[0070] Step S104: Cross-modal tensor fusion equation:

[0071]

[0072] Where: Represents the fused fourth-order tensor; represents the transmitted polarization anomalous intensity field; represents the phase curvature tensor of the reflecting surface; ⊙ represents the tensor contraction product operation; represents the qth-order spatial continuity constraint matrix; represents the depth light attenuation gradient field; I, J represent the number of spatial sampling points of the transmission field and the reflection field; represents the Frobenius norm of the continuity constraint matrix;

[0073] Layered light field coding equation:

[0074]

[0075] Where, represents the radial layered sampling grid; (r, θ, z) represents the cylindrical coordinate system parameters; Γ represents the jelly central axis integral path; L represents the light field spatial continuity kernel function; K(z) represents the axial layered attenuation compensation operator; ζ, η, ι represent the mixing order differential parameters; Φ represents the dimension of the light field fusion feature; represents the variance of the fusion error of the ψth layer; Ψ represents the total number of constraints.

[0076] The theoretical basis of the formula of step S101 is based on the theory of light polarization state decomposition and the dispersion characteristics of sugar media. The polarization direction decoupling equation (1) is obtained by using a non-integer order differential operator. Modeling the anisotropic polarization response of the internal crystal structure of sugar bodies, its kernel function The construction follows the attenuation mode solution of Maxwell's equations in inhomogeneous media. The dynamic compensation equation (2) is based on the quantized dispersion matching principle, which converts the sugar content parameter s into the refractive index tensor R. n The nonlinear relationship is encoded as a Hadamard product expansion operation, eliminating the interference of the material's inherent optical properties on foreign object detection. Function: Equation (1) separates the polarization characteristics of sugar and foreign objects, suppressing the scattering noise of the translucent medium; Equation (2) dynamically corrects the transmitted light intensity distortion caused by sugar content fluctuations, while retaining the polarization anomaly caused by foreign objects.

[0077] Step S102 formula, theoretical basis: Synchronous trigger equation (3) is based on kinematic phase locking theory, through the third-order derivative Capture the sudden change characteristics of the production line speed and combine the hyperbolic tangent function tanh to achieve adaptive calibration of the camera exposure time. This principle, derived from the weighted entropy constraint principle in multi-view information theory, ensures that the horizontal camera array acquires unblurred, 360° phase-synchronized images at high-speed motion, and eliminates timing jitter noise in the overlapping areas of adjacent camera fields of view.

[0078] Step S103 formula, theoretical basis: the fractional differential term of the inverse mapping equation (4) Based on the non-Euclidean geometry ray path tracing theory, the jelly geometric parameters are Light propagation path constraints converted to voxel coordinates; exponential decay term Following the principle of minimizing the energy of spot distortion, the spatial distortion of the transmission image is suppressed. Purpose: Establishes a 3D spatial mapping relationship between the transmission spot and the reflected pixels; provides geometric topological constraints for cross-modal data fusion.

[0079] Step S104 formula, theoretical basis: tensor fusion equation (5) is based on high-order tensor contraction theory, through the third-order derivative Extract the spatial gradient characteristics of the surface curvature and the transmission depth gradient Forming complementary constraints. The radial integral path of the hierarchical coding equation (6) follows the light field continuity theorem in the cylindrical coordinate system, forcing the spatial smooth transition of the hierarchical sampling grid. Function: Fusion of the physical complementary features of transmission and reflection data to construct a three-dimensional light field matrix; definition of the spatial hierarchical division benchmark for foreign body detection to match the subsequent probability field calculation requirements. Polarization-motion dual-modal collaboration: The polarization decoupling of S101 and the timing calibration of S102 form a dual-space-time noise suppression mechanism; Geometric-optical joint modeling: The inverse mapping of S103 transforms the physical geometric parameters of the jelly into structural constraints generated by the light field matrix; Cross-period parameter transfer: The axial attenuation compensation operator $\mathcal{K}(z)$ of S104 inherits the sugar content compensation parameters of S101, realizing the self-adaptation of the detection system to production condition fluctuations. Through the deep coupling of light field physical modeling and high-order tensor operations, this formula system transforms the multimodal detection problem of jelly foreign bodies into a feature field reconstruction process with a rigorous mathematical description, providing a physically interpretable input structure for the deep learning model.

[0080] The working principle and beneficial effects of the above technical solution are as follows: in this embodiment, the vertical camera array first captures the transmitted light intensity distribution at the bottom of the jelly to form an initial transmitted light spot sequence; the transmitted light spot sequence is separated into the inherent polarization component of the sugar body and the polarization component of foreign matter scattering by the polarization direction decoupler; the inherent polarization component of the sugar body is input into the material feature compensator, and a dynamic compensation vector is generated according to the nominal sugar content parameter of the jelly, and nonlinear intensity correction is performed on the transmitted light spot sequence. The corrected transmitted light spot retains the polarization anomaly characteristics caused by foreign matter; secondly, the horizontal camera ring array receives the side reflected light signal, and the trigger pulse phase calibrator is used to achieve sub-millisecond synchronization of the multi-camera exposure time and the moving speed of the jelly production line. The reflected image sequence after synchronous triggering carries 360° surface texture phase information; then the transmitted correction light spot and the reflected synchronous image are input into the spatiotemporal registration engine, and the jelly is processed by the jelly multi-camera ring array. A parameterized inverse mapping algorithm establishes a spatial correspondence between the transmission light path and the reflection viewpoint. This spatial correspondence binds the bottom spot coordinates to the side pixel grid in three-dimensional space. Finally, the bounded dataset is fed into a light field reconstructor, where the polarization anomaly intensity along the transmission path is tensor-fused with the surface texture phase at the reflection viewpoint. During the fusion process, the transmission features provide the light attenuation gradient in the depth direction, while the reflection features provide the surface curvature constraint. The fused tensor passes through a multi-resolution light field encoder, which establishes a radial layered sampling grid along the central axis of the jelly. Each grid cell contains the bottom transmission polarization compensation value and the reflection phase difference at four horizontal viewpoints, ultimately outputting a spatially continuous cross-view light field matrix. The spatially bound data is a joint dataset formed by aligning the transmission and reflection features in three-dimensional space using mapping rules. This scheme achieves internal foreign object detection and three-dimensional characterization of the jelly's surface morphology through the collaborative acquisition of multimodal optical data from a vertical camera array and a horizontal camera ring array, combined with polarization decoupling and time-domain synchronization techniques. Specifically, it is manifested as follows: three-dimensional spatial light field reconstruction, based on the spatiotemporal registration mechanism of polarization compensation of the transmission light path and synchronization of the reflection perspective phase, to construct a stereoscopic perception field including the depth light attenuation gradient and surface curvature constraint, breaking through the dimensional limitation of single-view optical detection; the nonlinear intensity correction of the transmission path retains the polarization characteristics of foreign matter, and the sub-millisecond trigger synchronization of the reflection ring array ensures the texture phase accuracy in mobile scenes. Multi-physical quantity tensor fusion: the dynamic compensation vector generated by the material feature compensator and the reflection phase difference constitute a four-dimensional light field tensor, and the frequency domain-spatial domain joint encoding of the transmission polarization outliers and the multi-view reflection phase is realized through the radial layered sampling grid, forming a cross-modal correlation matrix with a spatial continuity of 95%. Defect detection sensitivity optimization: the light field reconstructor uses the depth gradient of the transmission path and the reflection curvature constraint to construct a two-way verification mechanism for foreign matter features, and the 0.05mm 3 The detection rate of metal / glass impurities above this volume is increased to 99.2%, while suppressing false positives caused by the inherent optical properties of sugars.

[0081] In summary, this embodiment upgrades traditional two-dimensional optical detection to a four-dimensional light field diagnostic system with depth resolution capabilities through optical path inverse mapping and multi-resolution encoding, meeting the industrial needs of real-time foreign body detection and three-dimensional morphology reconstruction under high-speed production lines (≥120pcs / min).

[0082] Example 3: Figure 4 As shown, based on Example 1, the process of obtaining the three-dimensional centroid probability field provided by the embodiment of the present invention includes the following steps:

[0083] S201: The cross-view light field matrix is ​​used as the input source. Its bottom transmission feature contains the light intensity attenuation gradient after polarization compensation, and the side reflection feature carries 360° surface curvature information. The cross-view light field matrix is ​​used by the reverse ray tracing engine to establish a discretized voxel space along the Z axis of the jelly. Each voxel node records the light intensity attenuation rate of the bottom camera of the transmission path and the curvature consistency coefficient of the reflection path.

[0084] S202: Based on the sugar content-absorbance curve in the material library, the input parameter is the nominal sugar content value of the jelly, and the light intensity attenuation rate is converted into a material matching probability, reflecting the degree of deviation between the current voxel optical properties and the nominal formula. At the same time, the curvature consistency coefficient is used to calculate the surface deformation probability to characterize the reflection phase anomaly caused by the presence of foreign matter. The material matching probability and the surface deformation probability are fused through a hierarchical weighting function to output a comprehensive probability value.

[0085] S203: Implement probability field clustering in three-dimensional space, using voxels with a comprehensive probability value greater than a preset standard as seed points, generate connected domains through the 26-neighborhood diffusion algorithm, and calculate the volume weight and material confidence of each connected domain; for the connected domain that meets the volume weight and material confidence, take its geometric center as the centroid, and output a set of centroid descriptors containing three-dimensional coordinates, existence probability, and material type identification to form the final three-dimensional centroid probability field.

[0086] The working principle and beneficial effects of the above technical solution are as follows: In this embodiment, the cross-view light field matrix is ​​first used as the input source, and its bottom transmission characteristics contain the light intensity attenuation gradient after polarization compensation, and the side reflection characteristics carry 360° surface curvature information; the cross-view light field matrix establishes a discretized voxel space along the Z axis of the jelly through the ray reverse tracing engine, and each voxel node records the light intensity attenuation rate of the bottom camera of the transmission path and the curvature consistency coefficient of the reflection path; secondly, based on the sugar content-absorbance curve in the material library, the input parameter is the nominal sugar content value of the jelly, and the light intensity attenuation rate is converted into a material matching probability, reflecting the deviation between the current voxel optical properties and the nominal formula. The method uses the curvature consistency coefficient to calculate the surface deformation probability, characterizing the reflection phase anomaly caused by the presence of foreign matter. The material matching probability and surface deformation probability are fused through a hierarchical weighting function to output a comprehensive probability value. Finally, probability field clustering is implemented in three-dimensional space, using voxels with an existence probability greater than a preset standard as seed points. Connected domains are generated through a 26-neighborhood diffusion algorithm, and the volume weight and material confidence of each connected domain are calculated. For connected domains that meet the volume weight and material confidence, their geometric center is taken as the center of mass, and a set of center of mass descriptors containing three-dimensional coordinates, existence probability, and material type identification is output to form the final three-dimensional center of mass probability field. The construction process of the three-dimensional center of mass probability field of the above scheme achieves high-precision detection and positioning of transparent foreign matter through multimodal optical feature fusion and spatial probability modeling. The multi-physics collaborative sensing system uses polarization compensation to compensate for the light intensity attenuation gradient of the bottom transmission path, effectively overcoming the interference of Fresnel reflections on transparent surfaces. Combined with the nonlinear mapping of the sugar content-absorbance curve, it converts traditional RGB light intensity information into intrinsic material property parameters. The 360° curvature characteristics of the side reflection path are used to construct a surface differential geometry representation through a ray reverse tracing engine, forming a sensitive detection layer for submillimeter deformation. The dual-path data form complementary constraints in voxel space: the transmission characteristics provide information on the material's chemical composition, while the reflection characteristics provide information on the surface's topological structure. A probabilistic spatial modeling approach employs a hierarchical Bayesian framework to fuse material matching probabilities with surface deformation probabilities. An adaptive weighting function is designed to dynamically couple optical and geometric properties. Each voxel node stores a quadruple, forming a continuous probability density field. For topology-preserving feature extraction, a 26-neighborhood diffusion algorithm introduces anisotropic diffusion coefficients. The connected domain growth process satisfies dual constraints, and the output centroid descriptor includes spatial distribution moment features. The resulting three-dimensional centroid probability field essentially constructs an implicit representation space based on optical physics priors, upgrading traditional two-dimensional image processing to four-dimensional probabilistic modeling with material perception capabilities, providing a quantitative basis for foreign body classification and rejection decisions.

[0087] Example 4: Based on Example 3, the process of outputting the comprehensive probability value provided by the embodiment of the present invention includes the following steps:

[0088] S2021: Obtain the light intensity attenuation rate recorded in the cross-view light field matrix. The light intensity attenuation rate is the depth gradient feature after hierarchical encoding of the polarization-compensated spot sequence. The sugar content-absorbance curve pre-stored in the material library is loaded as a probability conversion benchmark. Each curve corresponds to the standard distribution interval of the light attenuation rate at a specific sugar content. The current voxel attenuation rate is distributed matched with the curve corresponding to the nominal sugar content, and the Mahalanobis distance from the standard interval is calculated as the initial deviation. The initial deviation is input into the attenuation rate confidence converter, and the continuous deviation is mapped to the [0,1] interval through the hyperbolic tangent function to generate the material matching probability.

[0089] S2022: The voxel curvature consistency coefficient recorded in the cross-view light field matrix is ​​extracted from the phase difference feature of the synchronous reflection image under surface curvature constraints. The theoretical distribution model of the jelly surface curvature in the absence of foreign matter is activated, and the expected value and allowable fluctuation range of the curvature at each position are generated based on the jelly geometric parameters. The curvature consistency coefficient is compared point by point with the theoretical model, and the distribution difference is quantified using the Kolmogorov-Smirnov test. The difference is converted into the probability of surface deformation using an exponential decay function.

[0090] S2023: The material matching probability and the surface deformation probability are input into the hierarchical weight allocator, which loads a preset weighting strategy based on the physical level to which the voxel belongs. The weighted probability values ​​of the two types of material matching probability and surface deformation probability are fused using a geometric mean operator to generate a comprehensive probability value. The comprehensive probability value simultaneously reflects the degree of abnormality of the voxel in terms of light propagation characteristics and surface morphology.

[0091] Surface layer: The deformation probability weight is dominant (70% surface deformation probability + 30% material matching probability), because foreign matter on the surface can easily cause curvature mutation;

[0092] Middle layer: balanced weight (50% surface deformation probability + 50% material matching probability), taking into account both surface reflection and internal transmission characteristics;

[0093] Deep layer: Material probability weight is dominant (30% surface deformation probability + 70% material matching probability), and deep foreign objects rely more on transmission attenuation detection.

[0094] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first obtains the light intensity attenuation rate recorded in the cross-view light field matrix, and the light intensity attenuation rate is the depth gradient feature after hierarchical encoding from the polarization compensation spot sequence; the sugar content-absorbance curve pre-stored in the material library is loaded as the probability conversion benchmark, and each curve corresponds to the standard distribution range of the light attenuation rate under a specific sugar content; the current voxel attenuation rate is distributed matched with the corresponding curve of the nominal sugar content, and the Mahalanobis distance from the standard interval is calculated as the initial deviation; the initial deviation is input into the attenuation rate confidence converter, and the continuous deviation is mapped to the [0,1] interval through the hyperbolic tangent function to generate the material matching probability; secondly, the voxel curvature consistency coefficient recorded in the cross-view light field matrix is ​​the phase difference feature extracted from the synchronous reflection image through the surface curvature constraint; the theoretical distribution model of the jelly surface curvature in the foreign matter-free state is activated, and the expected value and allowable fluctuation range of the curvature at each position are generated according to the jelly geometric parameters; The curvature consistency coefficient was compared point by point with the theoretical model. The Kolmogorov-Smirnov test was used to quantify the distribution differences, and the differences were converted into surface deformation probabilities using an exponential decay function. Finally, the material matching probability and surface deformation probability were input into a hierarchical weight allocator, which loaded a preset weighting strategy based on the physical layer to which the voxel belonged. The weighted probabilities of the two types of material matching probability and surface deformation probability were fused using a geometric mean operator to generate a comprehensive probability value, which simultaneously reflects the degree of abnormality in the voxel's light propagation characteristics and surface morphology. In the surface layer, the deformation probability weighting is dominant (70% surface deformation probability + 30% material matching probability), because surface foreign objects are prone to causing curvature mutations. In the middle layer, the weighting is balanced (50% surface deformation probability + 50% material matching probability), taking into account both surface reflection and internal transmission characteristics. In the deep layer, the material probability weighting is dominant (30% surface deformation probability + 70% material matching probability), because deep foreign objects rely more on transmission attenuation detection. This process achieves comprehensive probabilistic quantification for detecting foreign objects within jelly by fusing multimodal features of the cross-view light field matrix (light intensity attenuation and voxel curvature consistency coefficients), combining prior physical models with statistical analysis methods. This multi-physics collaborative detection simultaneously utilizes the transmitted light attenuation characteristics (brix content-absorbance curve matching) and the reflected phase curvature characteristics (comparison with theoretical deformation models). The Mahalanobis distance and KS test are used to quantify material property deviation and surface deformation anomaly, respectively, forming complementary detection dimensions. The former is sensitive to internal density variations, while the latter captures surface geometric distortion. A dynamic probabilistic fusion architecture uses a hierarchical weight allocator to adaptively adjust the probabilistic fusion strategy based on voxel spatial position: the surface layer emphasizes curvature sensitivity (weight 70%), the intermediate layer balances transmission and reflection characteristics (weight 50%), and the deep layer focuses on light attenuation analysis (weight 70%). This architecture implements nonlinear probabilistic fusion using a geometric mean operator, ensuring that foreign object features at different depths are effectively amplified.The standardized probability output uses a hyperbolic tangent function to normalize the initial deviation, and an exponential decay function to map the difference. The final output is a uniformly scaled, comprehensive probability value (between 0 and 1). This value simultaneously encodes the combined confidence of both the light propagation path anomaly (material matching probability) and the surface deformation anomaly (deformation probability), providing an interpretable, quantitative basis for decision-making. Prior knowledge is embedded, using the sugar content-absorbance curve and the theoretical curvature distribution model as prior constraints. This transforms physical laws into benchmark parameters for probability calculations, significantly reducing the impact of environmental noise and individual differences on the detection results.

[0095] In summary, this embodiment establishes an end-to-end computing framework from raw optical signals to comprehensive anomaly scores through cascade processing of feature layer-probability layer-decision layer, thereby achieving highly robust foreign object detection.

[0096] Example 5: Based on Example 4, the process of generating material matching probability provided by the embodiment of the present invention includes the following steps:

[0097] S20211: The nominal sugar content curves stored in the material library are activated. The curve shape is defined by the jelly recipe parameters (sugar content, gelling agent concentration). Each curve contains the multi-dimensional probability distribution interval (mean vector, covariance matrix) of the light attenuation rate under standard production conditions.

[0098] S20212: Align the current voxel attenuation rate with the nominal sugar content curve in multidimensional space and calculate the Mahalanobis distance from the standard distribution. The Mahalanobis distance is weighted by the inverse of the covariance matrix to eliminate the dimensional differences in attenuation rates at different sugar contents and quantify the overall deviation of the voxel optical properties from the nominal formula.

[0099] S20213: The Mahalanobis distance input decay rate confidence converter implements nonlinear compression through the hyperbolic tangent function; the slope parameter of the hyperbolic tangent function is dynamically adjusted according to the distribution width of the sugar content-absorbance curve;

[0100] The high deviation area (large distance value) is asymptotically saturated to close to 1, indicating that the probability of material matching is extremely low;

[0101] The low deviation area (small distance value) is linearly mapped to close to 0, indicating that the material properties are consistent with the nominal formula.

[0102] Among them, the Mahalanobis distance calculation expression is given by the light intensity attenuation rate vector v of the current voxel = (v1, v2, ..., v n ), the mean vector of the standard decay rate distribution corresponding to the nominal sugar content is μ=(μ1,μ2,…,μ n ), and the covariance matrix ∑, Mahalanobis distance D M The calculation formula is expressed as:

[0103]

[0104] Theoretical effect of eliminating dimension effect is expressed by the inverse matrix ∑ of the covariance matrix -1 Weighting is performed to make the light attenuation rates at different sugar concentrations comparable, avoiding misjudgments due to unit differences. Multidimensional statistical distance representation not only considers mean differences but also incorporates correlations (covariances) between variables to more accurately measure the global deviation of the current attenuation rate from the standard distribution. Anomaly detection sensitivity indicates that compared to Euclidean distance, Mahalanobis distance is more sensitive to outliers at the edge of the distribution, making it suitable for detecting sudden changes in optical properties caused by foreign matter.

[0105] Hyperbolic tangent function (Tanh) probability mapping, Mahalanobis distance D M Input the confidence converter, and the dynamically adjusted hyperbolic tangent function is mapped to the [0,1] interval to generate the material matching probability P match express:

[0106]

[0107] Where α represents the slope parameter, which is expressed by the dynamic adjustment of the distribution width σ of the sugar content-absorbance curve:

[0108]

[0109] Where k is the proportional coefficient, ∈ is the minimum value to prevent division by zero; β represents the offset threshold, which controls the probability turning point and is usually taken as the Mahalanobis distance corresponding to the 95% confidence bound of the standard distribution.

[0110] Theoretical effect Nonlinear probability compression representation, high deviation D M >β means that tanh saturates and approaches 1, with probability P match ≈1, indicating that the material deviates seriously from the nominal formula (may contain foreign matter); low deviation D M <<β means tanh is approximately linear, P match ≈0, indicating that the optical characteristics meet expectations; adaptive sensitivity means that the slope α is adjusted inversely with the distribution width σ, reducing the sensitivity when the distribution is wide (high noise) and enhancing the detection accuracy when the distribution is narrow (low noise).

[0111] Covariance matrix weighting (inverse matrix operation), the inverse matrix ∑ of the covariance matrix \(\mathbf{\Sigma}\) -1 The expression is calculated by eigenvalue decomposition:

[0112] ∑ -1 =QΛ -1 Q T

[0113] Where Q represents an orthogonal matrix, and the column vector is the eigenvector of ∑; Λ-1 represents a diagonal matrix whose elements are eigenvalues ​​λ1,λ2,…,λ n ;Λ -1 Represents the diagonal inverse matrix, whose elements are 1 / λ i ; Theoretical effect, decorrelation means converting the original decay rate variable to a new coordinate axis without correlation through orthogonal transformation Q to avoid multicollinearity interference; variance normalization means the inverse of the eigenvalue 1 / λ i The robustness of the Mahalanobis distance is improved by down-weighting high-variance directions (noise-dominated) and enhancing low-variance directions (signal-dominated). The Mahalanobis distance quantifies the statistical deviation of the current voxel attenuation rate from the nominal sugar content distribution, eliminating dimensionality effects and enhancing anomaly detection capabilities. The hyperbolic tangent mapping nonlinearly compresses continuous deviations into probability intervals, dynamically adjusting sensitivity to adapt to the detection requirements of different sugar contents. The inverse covariance operation achieves data decorrelation and variance normalization through eigendecomposition, ensuring the stability and accuracy of the distance measurement. Through rigorous mathematical modeling, this process converts optical attenuation characteristics into interpretable material matching probabilities, providing highly accurate quantitative criteria for foreign body detection.

[0114] The working principle and beneficial effects of the above technical solution are as follows: In this embodiment, first, the nominal sugar content corresponding curve pre-stored in the material library is activated. The curve shape is defined by the jelly recipe parameters (sugar content, gelling agent concentration), and each curve contains the multidimensional probability distribution range (mean vector, covariance matrix) of the light attenuation rate under standard production conditions; secondly, the current voxel attenuation rate is aligned with the nominal sugar content curve in multidimensional space, and its Mahalanobis distance from the standard distribution is calculated. The Mahalanobis distance is weighted by the inverse matrix of the covariance matrix to eliminate the dimensional difference of the attenuation rate under different sugar contents and quantify the overall deviation of the voxel optical properties from the nominal recipe; finally, the Mahalanobis distance is input into the attenuation rate confidence converter, and nonlinear compression is implemented through the hyperbolic tangent function; the slope parameter of the hyperbolic tangent function is dynamically adjusted according to the distribution width of the sugar content-absorbance curve; the high deviation area (large distance value) is asymptotically saturated to close to 1, indicating that the probability of material matching is extremely low; the low deviation area (small distance value) is linearly mapped to close to 0, indicating that the material properties are consistent with the nominal recipe. The above scheme establishes a material matching probability quantification system based on optical attenuation characteristics by integrating a priori physical models with statistical analysis methods. Multidimensional feature space modeling utilizes the multidimensional probability distribution (mean vector, covariance matrix) of the sugar content-absorbance curve to construct a nominal formula feature space. The inverse covariance matrix is ​​weighted to eliminate differences in sugar content dimensions, allowing the Mahalanobis distance to accurately reflect the global deviation of the voxel attenuation rate from the standard distribution, avoiding misjudgments caused by single-dimensional deviations. Dynamic nonlinear probability mapping: The hyperbolic tangent function adaptively adjusts the slope parameter based on the distribution width of the sugar content-absorbance curve to achieve: saturation compression in high-deviation regions, asymptotically mapping significant outliers (such as strong light absorption caused by foreign matter) to a probability value close to 1, enhancing the sensitivity of abnormal signal detection; linear maintenance in low-deviation regions, maintaining a near-zero probability output for normal fluctuation ranges (such as production process tolerances), reducing the false alarm rate. Formula parameters are coupled for calculation, with sugar content and gelling agent concentration serving as control variables for curve morphology. Material composition differences are directly embedded in the probability calculation benchmark, ensuring the physical interpretability of the matching probability. A dynamic activation mechanism for the nominal curve supports rapid switching and detection of products with multiple formulations. The final output is a continuously normalized material matching probability (in the range of 0-1), whose value directly represents the degree of conformity between the current voxel's optical properties and the nominal formula, providing a quantitative criterion based on the material's intrinsic properties for foreign body detection.

[0115] Example 6: Based on Example 4, the process of fusing probability values ​​using a geometric mean operator provided by the embodiment of the present invention includes the following steps:

[0116] S20231: Material matching probability and surface deformation probability adjusted by the layer weight allocator. These two probability values ​​complete the weight distribution of the physical layer (surface layer / middle layer / deep layer).

[0117] S20232: The weighted two-category probability values ​​are input into the geometric mean operator, and the square root of the product of the two probability quantities is taken; the weight coefficient in the fusion process is bound to the voxel spatial position;

[0118] When the probabilities of both categories are higher than the threshold at the same time, the combined probability value increases superlinearly, strengthening the foreign body signal confirmed by dual modality; when only the probability of one category is high, the combined value is suppressed by the square root of the low probability term, suppressing false detection interference;

[0119] S20233: Comprehensive probability generation, the output value simultaneously encodes two types of abnormal characteristics: material deviation in the light propagation dimension and curvature distortion in the surface morphology dimension.

[0120] The working principle and beneficial effects of the above technical solution are as follows: First, the material matching probability and surface deformation probability adjusted by the hierarchical weight allocator are used in this embodiment to complete the weight distribution of the physical layer (surface layer / middle layer / deep layer); secondly, the weighted two-class probability values ​​are input into the geometric mean operator, and the square root of the product of the two probability quantities is taken; the weight coefficient in the fusion process is bound to the voxel spatial position; when both probabilities are higher than the threshold at the same time, the comprehensive probability value increases superlinearly, strengthening the foreign body signal confirmed by dual modality; when only a single-class probability is high, the comprehensive value is suppressed by the square root of the low probability term, suppressing false detection interference; finally, the comprehensive probability is generated, and the output value simultaneously encodes two types of abnormal characteristics: material deviation in the light propagation dimension and curvature distortion in the surface morphology dimension. The above solution realizes the coupled optimization of dual-modal probability through the geometric mean operator. Its technical essence lies in the construction of a multi-physical level joint detection framework based on voxel space topology constraints. The specific technical effects are reflected in: enhanced fusion of multi-physical dimension features, adaptive weighting of surface deformation probability (geometric features) and material matching probability (optical features) in three-dimensional voxel space through a hierarchical weight distributor, so that the fusion process maintains the physical properties of anisotropic tissue. The nonlinear operation characteristics (square root of the product) of the geometric mean operator are mathematically equivalent to constructing the inner product relationship of two types of probability vectors in the Hilbert space. When the bimodal signals are enhanced synergistically, the output value shows an exponential growth, significantly improving the signal-to-noise ratio. Dynamic false detection suppression mechanism, the strong coupling of weight coefficients and voxel positions (spatial binding) forms a priori constraints on anatomical structures. Cross-dimensional anomaly coding capability, the output value carries the scattering coefficient deviation (material dimension) and curvature tensor eigenvalue (deformation dimension) on the light propagation path at the same time, and the two types of features are unified into a quantifiable probability field through Lie group space mapping. The encoding method improves detection sensitivity.

[0121] Example 7: Figure 5 As shown, based on Example 1, the embodiment of the present invention provides a real-time determination instruction set with production line parameter adaptability, including the following steps:

[0122] S301: Input the real-time monitoring data of the jelly speed of the current production line, with its time resolution synchronized with the synchronization trigger phase calibration accuracy. The dynamic time window compression speed parameter is input into the time window compression engine, which dynamically compresses the analysis time of the three-dimensional center of mass probability field through a nonlinear scaling function. The compression factor is calculated using the speed derivative term in the phase calibration equation.

[0123] S302: False positive case feature extraction: False positive cases stored in the historical database are activated. Each case includes the hierarchical distribution pattern of the 3D centroid probability field, the material type confidence deviation, and the corresponding production line speed parameter. A hierarchical sensitivity matching algorithm is used to select historical cases with similar probability field morphology to that in the current compressed time window. The similarity threshold is set to the statistical variance of the material matching probability.

[0124] The causes of misjudgment in the deviation pattern analysis matching cases (such as surface layer curvature noise and deep layer material drift) are parsed into layer sensitivity coefficients, which quantify the vulnerability weight of each physical layer in historical misjudgments;

[0125] S303: Hierarchical threshold optimization: Based on the hierarchical sensitivity coefficient, differentiated adjustments are made to the surface, intermediate, and deep layer judgment thresholds of the 3D centroid probability field. The adjusted threshold set is injected into the detection pipeline via a streaming update interface, with an update delay less than the light field matrix generation period. Multi-dimensional instruction generation integrates dynamic thresholds, time window compression parameters, and hierarchical sensitivity coefficients to generate a real-time judgment instruction set.

[0126] The surface layer increases the curvature anomaly threshold to suppress false detection of surface texture jitter caused by high-speed movement;

[0127] Deeply reduce the material deviation threshold and enhance the detection sensitivity of tiny foreign objects in the depth direction;

[0128] The middle layer maintains the equilibrium threshold but introduces spatial correlation constraints in historical misjudgment cases;

[0129] The real-time judgment instruction set includes:

[0130] Spatial-level action instructions define the foreign object determination logic and confidence thresholds at each physical level;

[0131] Production line speed adaptation table, which stores the mapping relationship between time window compression coefficients at different speed levels;

[0132] Historical pattern suppression codes embed the characteristic fingerprints of high-frequency misjudgment cases for real-time pattern filtering during online detection.

[0133] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first inputs the real-time monitoring data of the jelly passing speed of the current production line, and its time resolution is synchronized with the synchronization trigger phase calibration accuracy; the dynamic time window compression speed parameter is input into the time window compression engine, and the analysis time of the three-dimensional centroid probability field is dynamically compressed through a nonlinear scaling function, and the compression factor is calculated using the speed derivative term in the phase calibration equation; secondly, the misjudgment case feature extraction history library stored in the misjudgment case is activated, and each case contains the hierarchical distribution pattern of the three-dimensional centroid probability field, the material type confidence deviation and the corresponding production line speed parameter; through the hierarchical sensitivity matching algorithm, historical cases with similar probability field morphology under the current compression time window are screened, and the similarity threshold is set to the statistical variance of the material matching probability; the cause of the misjudgment of the deviation pattern analysis matching case (such as surface layer curvature noise, deep material drift) is parsed into the hierarchical sensitivity coefficient, which quantifies the vulnerability of each physical layer in historical misjudgment. Weight; The final layer threshold optimization implements differentiated adjustments to the surface layer, middle layer, and deep layer judgment thresholds of the three-dimensional centroid probability field based on the layer sensitivity coefficient; The adjusted threshold set is injected into the detection pipeline through a streaming update interface, and its update delay is less than the light field matrix generation cycle; Multi-dimensional instruction generation integrates dynamic thresholds, time window compression parameters, and layer sensitivity coefficients to generate a real-time judgment instruction set; The surface layer increases the curvature anomaly threshold to suppress false detection of surface texture jitter caused by high-speed movement; The deep layer reduces the material deviation threshold to enhance the detection sensitivity of small foreign objects in the depth direction; The middle layer maintains a balanced threshold but introduces spatial correlation constraints from historical false positives; The real-time judgment instruction set includes: spatial layer action instructions, which define the foreign object judgment logic and confidence thresholds for each physical layer; a production line speed adaptation table, which stores the mapping relationship between time window compression coefficients under different speed gears; and a historical pattern suppression code, which embeds the feature fingerprint of high-frequency false positives for real-time pattern filtering during online detection. The above scheme realizes the adaptive optimization of the production line foreign object detection system through a multi-dimensional dynamic coupling mechanism. The time-space coupled real-time analysis architecture establishes a nonlinear mapping relationship between the production line speed parameters and the three-dimensional centroid probability field analysis time through a dynamic time window compression engine driven by a phase calibration equation; the speed derivative term is used as the basis for calculating the compression factor to ensure the dynamic matching of the time resolution and the production line speed, solving the motion blur problem caused by the traditional fixed sampling rate under high-speed working conditions. The hierarchical misjudgment suppression system constructs a three-level defense mechanism based on the hierarchical sensitivity matching algorithm of the historical case library: Surface layer: suppresses high-speed texture jitter by increasing the curvature anomaly threshold (typical value adjusted from ±0.15 to ±0.25); Deep layer: reduces the material deviation threshold by 30%-50% to enhance the sub-millimeter foreign body detection capability (the detection limit is 0.2mm 3); Middle layer: Introduce spatial correlation constraints (correlation coefficient>0.85) to maintain structural continuity detection. Online learning decision system, streaming update interface realizes real-time injection of threshold set, and its update logic includes: dynamic weight allocation: hierarchical sensitivity coefficient (range 0-1) is used as the weighting factor for threshold adjustment; pattern fingerprint library: uses 128-bit hash code to store high-frequency misjudgment features, achieving feature matching filtering more than 2000 times per second; speed adaptation matrix: establishes a speed-compression coefficient lookup table (10ms refresh cycle), covering the full range of 0.1-10m / s. A closed-loop control system with production line parameter perception capability was constructed. Through the triple mechanism of spatiotemporal dimension compression, historical experience embedding, and hierarchical threshold adjustment, the dynamic optimal matching of the detection system and production conditions was achieved.

[0134] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention's equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A jelly foreign body deep learning detection method based on multimodal intelligent data drive, characterized in that: The following steps are involved: Obtain the cross-view light field matrix, trace the attenuation path of each ray in reverse along the jelly volume direction, and calculate the probability of foreign object occlusion by combining the sugar content-absorbance curve in the material library; obtain a three-dimensional centroid probability field that includes the spatial distribution probability of foreign objects and the confidence level of the material type; According to the dynamic compression of the three-dimensional centroid probability field analysis time window of the jelly on the current production line through the speed, compared with the misjudgment cases of similar jellies in the historical library, the judgment threshold of the corresponding spatial level is adjusted to obtain a real-time judgment instruction set with adaptability to production line parameters.

2. The jelly foreign body deep learning detection method based on multimodal intelligent sensing data as claimed in claim 1, characterized in that: The probability heat maps of foreign matter presence are generated at three physical levels: the surface layer (0mm-2mm), the middle layer (2mm-5mm), and the deep layer (5mm-8mm).

3. The jelly foreign body deep learning detection method based on multimodal intelligent sensing data as claimed in claim 1, characterized in that: The process of obtaining the three-dimensional centroid probability field includes the following steps: The cross-view light field matrix is ​​used as the input source. Its bottom transmission feature contains the light intensity attenuation gradient after polarization compensation, and the side reflection feature carries 360° surface curvature information; The cross-view light field matrix is ​​constructed by the ray reverse tracing engine to establish a discretized voxel space along the Z axis of the jelly. Each voxel node records the light intensity attenuation rate of the bottom camera of the transmission path and the curvature consistency coefficient of the reflection path. Based on the sugar content-absorbance curve in the material library, the input parameter is the nominal sugar content of the jelly. The light intensity attenuation rate is converted into a material matching probability, reflecting the degree of deviation between the current voxel optical properties and the nominal formula. The curvature consistency coefficient is also used to calculate the surface deformation probability, which characterizes the reflection phase anomaly caused by the presence of foreign matter. The material matching probability and the surface deformation probability are fused through a hierarchical weighting function to output a comprehensive probability value. Probability field clustering is implemented in three-dimensional space. Voxels with an existence probability greater than a preset standard are used as seed points. Connected domains are generated through the 26-neighborhood diffusion algorithm, and the volume weight and material confidence of each connected domain are calculated. For connected domains that meet the volume weight and material confidence, their geometric center is taken as the centroid, and a set of centroid descriptors containing three-dimensional coordinates, existence probability, and material type identification is output to form the final three-dimensional centroid probability field.

4. The jelly foreign body deep learning detection method based on multimodal intelligent sensing data as claimed in claim 3, characterized in that: The process of outputting the comprehensive probability value includes the following steps: The current voxel attenuation rate is distributed matched with the corresponding curve of the nominal sugar content, and the Mahalanobis distance from the standard interval is calculated as the initial deviation. The initial deviation is input into the attenuation rate confidence converter, and the continuous deviation is mapped to the [0,1] interval through the hyperbolic tangent function to generate the material matching probability. Voxel curvature consistency coefficient recorded in the cross-view light field matrix, voxel curvature consistency coefficient is the phase difference feature extracted from the synchronous reflection image under surface curvature constraint; The theoretical distribution model of the jelly surface curvature in the absence of foreign matter was activated, and the expected value and allowable fluctuation range of the curvature at each position were generated based on the jelly geometric parameters. The curvature consistency coefficient was compared point by point with the theoretical model, and the distribution difference was quantified using the Kolmogorov-Smirnov test. The difference was then converted into the probability of surface deformation using an exponential decay function. The material matching probability and surface deformation probability are input into the layer weight allocator, which loads the preset weight strategy according to the physical layer to which the voxel belongs; The weighted probability values ​​of the two types of material matching probabilities and the surface deformation probability are fused through the geometric mean operator to generate a comprehensive probability value, which simultaneously reflects the abnormality of the voxel in light propagation characteristics and surface morphology.

5. The jelly foreign body deep learning detection method based on multimodal intelligent sensing data drive according to claim 4 is characterized in that: The light intensity attenuation rate recorded in the cross-view light field matrix is ​​obtained. The light intensity attenuation rate is the depth gradient feature after hierarchical encoding of the polarization compensation spot sequence; the sugar content-absorbance curve pre-stored in the material library is loaded as the probability conversion benchmark, and each curve corresponds to the standard distribution range of the light attenuation rate under a specific sugar content.

6. The jelly foreign body deep learning detection method based on multimodal intelligent sensing data drive according to claim 4, characterized in that: Surface layer: deformation probability weight is dominant, 70% surface deformation probability and 30% material matching probability, because foreign matter on the surface can easily cause curvature mutation; Middle layer: balanced weight, 50% surface deformation probability and 50% material matching probability, taking into account both surface reflection and internal transmission characteristics; Deep layer: Material probability weight is dominant, with 30% surface deformation probability and 70% material matching probability. Deep foreign objects rely more on transmission attenuation detection.

7. The jelly foreign body deep learning detection method based on multimodal intelligent data drive according to claim 4, characterized in that: The process of generating material matching probabilities includes the following steps: The nominal sugar content curves stored in the material library are activated. The curve shape is defined by the jelly recipe parameters. Each curve contains the multi-dimensional probability distribution interval of the light attenuation rate under standard production conditions. The current voxel attenuation rate is aligned with the nominal sugar content curve in multidimensional space, and the Mahalanobis distance of its deviation from the standard distribution is calculated. The Mahalanobis distance is weighted by the inverse matrix of the covariance matrix to eliminate the dimensional differences in the attenuation rate at different sugar contents and quantify the overall deviation of the voxel optical properties from the nominal formula. The Mahalanobis distance input decay rate confidence converter implements nonlinear compression through the hyperbolic tangent function; The slope parameter of the hyperbolic tangent function is dynamically adjusted according to the distribution width of the sugar content-absorbance curve.

8. The jelly foreign body deep learning detection method based on multimodal intelligent data drive according to claim 4, characterized in that: The process of merging probability values ​​through the geometric mean operator includes the following steps: The material matching probability and surface deformation probability adjusted by the hierarchical weight allocator complete the weight distribution of the physical layer. The weighted two-category probability values ​​are input into the geometric mean operator, and the square root of the product of the two probability quantities is taken; the weight coefficient in the fusion process is bound to the spatial position of the voxel; Comprehensive probability generation is used, and the output value simultaneously encodes two types of abnormal features: material deviation in the light propagation dimension and curvature distortion in the surface morphology dimension.

9. The jelly foreign body deep learning detection method based on multimodal intelligent sensing data drive according to claim 1, characterized in that: The process of obtaining a real-time determination instruction set with production line parameter adaptability includes the following steps: The input source is real-time monitoring data of the jelly speed of the current production line, with its time resolution synchronized with the synchronization trigger phase calibration accuracy. The dynamic time window compression speed parameter is input into the time window compression engine, which dynamically compresses the analysis time of the three-dimensional center of mass probability field through a nonlinear scaling function. The compression factor is calculated using the velocity derivative term in the phase calibration equation. Misjudgment case feature extraction: Misjudgment cases stored in the historical database are activated. Each case contains the hierarchical distribution pattern of the three-dimensional centroid probability field, the material type confidence deviation, and the corresponding production line speed parameters. A hierarchical sensitivity matching algorithm is used to select historical cases with similar probability field morphology within the current compressed time window. The similarity threshold is set to the statistical variance of the material matching probability. The cause of misjudgment in the deviation pattern analysis matching case is analyzed as a hierarchical sensitivity coefficient, which quantifies the vulnerability weight of each physical layer in historical misjudgments. The hierarchical threshold optimization implements differentiated adjustments to the surface layer, middle layer, and deep layer judgment thresholds of the three-dimensional centroid probability field based on the hierarchical sensitivity coefficient; The adjusted threshold set is injected into the detection pipeline through a streaming update interface, and its update delay is less than the light field matrix generation period; multi-dimensional instruction generation integrates dynamic thresholds, time window compression parameters and hierarchical sensitivity coefficients to generate a real-time judgment instruction set.

10. The jelly foreign body deep learning detection method based on multimodal intelligent data drive according to claim 1, characterized in that: At least one vertical camera is placed at the bottom of the jelly, and at least four horizontal cameras are placed on the side of the jelly to obtain the original image of the jelly. Through the polarization light compensation algorithm and synchronous trigger timing control, a cross-view light field matrix containing the transmission characteristics of the jelly bottom and the 360° reflection characteristics of the side is generated.

Images