A jewelry identification and grading method and system based on image recognition

Abstract

The application relates to the technical field of image recognition, and discloses a jewelry identification and grading method and system based on image recognition, which comprises the following steps: collecting an original image sequence of a high-refractive transparent medium target under a dynamic light source environment; calculating a global pixel displacement field between adjacent image frames and analyzing a global rigid body motion vector in the global pixel displacement field; inversely compensating the global pixel displacement field by using the global rigid body motion vector, and stripping to obtain a luminance space evolution track representing the refractive characteristic of the target; calculating the fluctuation variance of the track and determining the shape similarity between the track and a template; and determining the grade of the target according to the shape similarity. By inversely hedging the background displacement noise, the application eliminates the interference of mechanical shaking of a conveying mechanism on microscopic optical phase change characteristics, realizes deep decoupling of physical structure characteristics and environmental noise, and improves the grading stability of the transparent medium under complex dynamic working conditions.

Description

Technical Field

[0001] This invention relates to a jewelry identification and grading method and system based on image recognition, belonging to the field of image recognition technology. Background Technology

[0002] Currently, extracting the spatial distribution features of jewelry using image acquisition equipment and outputting identification results is a common practice in the industry. Existing visual identification systems focus on acquiring physical features such as geometric contours, color saturation, and internal inclusions in static images, and combine them with preset models to complete jewelry grading. Essentially, they still rely on the spatial dimension mapping of static features, lack the ability to identify and classify high-frequency dynamic phase transition features in image sequences, and cannot effectively suppress nonlinear noise interference caused by background motion in the image domain. This has certain application value in meeting conventional identification needs. In addition to optimizing hardware vibration isolation design, improving algorithm logic is also a technical direction to improve sorting accuracy. For example, Chinese invention patent with authorization announcement number CN120446140B discloses a pearl flaw sorting system and method based on machine vision. By stitching multiple frames of images into a panoramic image to be detected, flaw estimation and luster analysis are performed based on the brightness value and color opposite dimension after color space conversion.

[0003] However, due to the optical properties of high-dispersion transparent media, key evaluation indicators such as fire and scintillation originate from the dynamic evolution process caused by the refraction of light inside the jewelry. Existing identification processes implicitly rely on the idealized assumption that the jewelry is physically stationary relative to the acquisition equipment. In industrial automated production line scenarios, the mechanical vibration generated by the conveyor belt causes global rigid body displacement noise to be mixed into the acquired image sequence. This macroscopic global displacement is highly coupled with the microscopic, high-frequency local refraction pixel displacement generated by the internal cross-section of the jewelry in the pixel motion field. Existing optical flow calculation methods lack an effective mechanism to remove the physical source of the motion field, resulting in nonlinear distortion of the output optical response trajectory due to environmental jitter interference.

[0004] Therefore, the technical problem to be solved by this invention is how to establish a dynamic reference system in the pixel domain through image processing algorithms to counteract the global motion components introduced by the environment, thereby achieving accurate extraction of the microscopic refractive features inside the transparent medium. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A jewelry identification and grading method based on image recognition, comprising the following steps: Step S1: Obtain the original image sequence of a high refractive index transparent medium target that changes continuously over time under a dynamic light source environment; Step S2: Calculate the global pixel displacement field between adjacent image frames in the original image sequence; Step S3: Analyze the global rigid body motion vector in the global pixel displacement field. The global rigid body motion vector represents the background displacement noise introduced by the conveying mechanism. Step S4: Use the global rigid body motion vector to perform reverse compensation and offset on the global pixel displacement field, so as to separate the brightness spatial evolution trajectory that characterizes the internal refractive properties of the high refractive index transparent medium target from the aliased pixel motion. Step S5: Calculate the pixel energy integral of the brightness spatial evolution trajectory within a preset time window, and determine the fluctuation variance of the pixel energy integral within the preset time window. Step S6: Map the brightness space evolution trajectory to the multidimensional vector space, and calculate the morphological similarity between the multidimensional vector space and the templates of each level in the standard level library. Step S7: Index the preset grading matrix according to morphological similarity and fluctuation variance, and output category labels that characterize the physical cut grade and internal clarity grade of the high refractive index transparent medium target.

[0006] Preferably, step S2 includes: step S21, smoothing and denoising each frame of the original image sequence using a spatial filter; step S22, extracting pixel feature point sets from adjacent image frames; and step S23, generating a global pixel displacement field using a dense optical flow algorithm based on the spatial coordinate offset of the pixel feature point sets between adjacent image frames.

[0007] Preferably, step S3 includes: step S31, statistically analyzing the distribution probability of motion vectors of feature points within the global pixel displacement field; step S32, identifying the peak interval with the highest proportion in the distribution probability and conforming to the spatial displacement consistency constraint as the background motion principal component; and step S33, fitting the background motion principal component with a homography transformation matrix to obtain the global rigid body motion vector.

[0008] Preferably, step S4 includes: step S41, constructing a motion compensation coordinate image model using global rigid body motion vectors; step S42, projecting the global pixel displacement field onto the motion compensation coordinate image model to cancel low-frequency macroscopic pixel displacement components; and step S43, extracting the local pixel brightness phase transition data after cancellation processing and reconstructing the brightness spatial evolution trajectory.

[0009] Preferably, in step S5, the variance of fluctuation The calculation rules are as follows: ,in, The brightness space evolution trajectory in the 1st The pixel energy integral value of each sampling node. The total number of sampling frames within the preset time window. It is the arithmetic mean of the pixel energy integral values ​​within the total number of sampled frames.

[0010] Preferably, the output category label characterizing the internal clarity grade of a high-refractive-index transparent medium target includes: step S51, monitoring pixel energy integral value. The evolution slope over time; step S52, when the evolution slope exceeds the preset negative jump threshold, and the fluctuation variance... When the deviation exceeds the preset threshold, it is determined that there is an opaque inclusion inside the high refractive index transparent medium target, and the corresponding image space coordinates are marked.

[0011] Preferably, step S6 includes: step S61, encoding the luminance spatial evolution trajectory into a feature vector with temporal characteristics; step S62, calculating the cumulative distortion distance between the feature vector and the grade template; and step S63, aligning the cumulative distortion distance on the time axis using an elastic matching algorithm to eliminate the nonlinear distortion of the time axis caused by the non-uniform motion of the light source.

[0012] Preferably, the output category label characterizing the physical cutting grade of the high refractive index transparent medium target includes: step S64, extracting the color phase angle difference of the high refractive index transparent medium target under different refractive phases according to the elastic matching algorithm; step S65, inputting the color phase angle difference and morphological similarity as independent variables into the preset classification network model to obtain the proportional symmetry grade determination result.

[0013] Preferably, before the calculation in step S2, the method further includes the following preprocessing steps: Step S91, detecting the global average brightness of the original image sequence; Step S92, when the deviation rate between the average brightness and the preset brightness benchmark exceeds 5%, performing reverse compensation on the grayscale values ​​of the original image sequence through a linear gain function based on the deviation rate, and using the compensated original image sequence as the calculation input for step S2; in step S1, the original image sequence is synchronously acquired by at least two sensing units with a distribution angle between 45° and 90°; Step S4, by fusing the brightness spatial evolution trajectories under different viewpoints, constructing a stereo optical feature tensor characterizing the internal geometric structure of the high refractive index transparent medium target.

[0014] A jewelry identification and grading system based on image recognition, comprising: The image acquisition module is used to acquire a sequence of original images of a high-refractive-index transparent medium target that changes continuously over time under a dynamic light source environment. The displacement field calculation module is used to calculate the global pixel displacement field between adjacent image frames based on the original image sequence. The noise extraction module is used to parse the global rigid body motion vector in the global pixel displacement field in order to extract the background displacement noise introduced by the external environment; The trajectory reconstruction module, connected to the noise extraction module, is used to perform inverse compensation and offset on the global pixel displacement field using the global rigid body motion vector, and strip away the brightness spatial evolution trajectory that characterizes the internal refractive properties of the high refractive index transparent medium target. The energy analysis module, connected to the trajectory reconstruction module, is used to calculate the pixel energy integral of the brightness spatial evolution trajectory within a preset time window and determine the fluctuation variance of the pixel energy integral within the preset time window. The grading output module is used to calculate the morphological similarity between the luminance spatial evolution trajectory and the grading templates in the standard grading library, and index the preset grading matrix according to the morphological similarity and fluctuation variance to output category labels that characterize the physical cut grade and internal clarity grade of the high refractive index transparent medium target.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In the jewelry identification and grading method, a dynamic reference system offset algorithm is established in the image sequence processing process by introducing a motion decoupling mechanism between the background contour edge pixel set and the local highlight pixel cluster. This significantly improves the robustness of the image recognition model in complex dynamic backgrounds and eliminates the interference of mechanical vibration on the extraction of microscopic optical features in industrial identification environments. This invention utilizes the rigid body affine transformation relationship of the physical boundary of the jewelry between adjacent frames to calculate the rotation matrix and translation matrix representing the global displacement. Based on this, reverse spatial compensation is performed on the absolute displacement vector of the highlight pixel cluster. This dynamic reference system offset mechanism constructed in the image pixel domain enables the final generated relative spatial displacement vector to be stripped of external physical motion noise and purely present the intrinsic refraction law caused by the internal cutting structure of the jewelry, ensuring the consistency of feature extraction under non-steady-state acquisition conditions.

[0016] 2. The constructed three-dimensional spatiotemporal feature tensor realizes the dimensional leap from static spatial features to dynamic temporal response features, solving the problem of loss of dynamic evolution information of fire and color in traditional visual identification. By stitching the pixel brightness spatial distribution features of multiple frames of images along the time axis, this invention transforms the pixel grayscale statistics of a single viewpoint into a trajectory sequence that represents the continuous phase change of light and shadow. This mechanism utilizes the data correlation and temporal verification between adjacent frames to completely capture the high-frequency flicker frequency and color gradation span of jewelry under relatively moving light sources, thereby providing high-dimensional data support with physical correlation for cut grade determination.

[0017] 3. The dynamic time warping matching mechanism between the optical response trajectory sequence and the standard-level trajectory library improves the system's stability in recognizing non-uniform motion and complex refraction paths. Due to the complexity of the internal refraction path of jewelry, the fire and shimmer exhibit non-linear scaling characteristics on the time axis. This invention calculates the dynamic time warping distance to achieve accurate measurement of trajectory morphological similarity under time distortion conditions. This feature coupling method not only reduces the system's rigid dependence on the uniformity of light source motion, but also effectively distinguishes light and shadow jumps caused by minute cutting differences through morphological matching of trajectory envelopes, enhancing the recognition accuracy of the classification model when dealing with high refractive index transparent media. Attached Figure Description

[0018] Figure 1 This is a flowchart of the dynamic image recognition-based jewelry identification and grading process involved in the present invention. Figure 2 This invention relates to the logical processing architecture and multidimensional decision graph of the jewelry grading system. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A jewelry identification and grading method based on image recognition includes the following steps: Step S1: Obtain the original image sequence of a high refractive index transparent medium target that changes continuously over time under a dynamic light source environment; Step S2: Calculate the global pixel displacement field between adjacent image frames in the original image sequence; Step S3: Analyze the global rigid body motion vector in the global pixel displacement field. The global rigid body motion vector represents the background displacement noise introduced by the conveying mechanism. Step S4: Use the global rigid body motion vector to perform reverse compensation and offset on the global pixel displacement field, so as to separate the brightness spatial evolution trajectory that characterizes the internal refractive properties of the high refractive index transparent medium target from the aliased pixel motion. Step S5: Calculate the pixel energy integral of the brightness spatial evolution trajectory within a preset time window, and determine the fluctuation variance of the pixel energy integral within the preset time window. Step S6: Map the brightness space evolution trajectory to the multidimensional vector space, and calculate the morphological similarity between the multidimensional vector space and the templates of each level in the standard level library. Step S7: Index the preset grading matrix according to morphological similarity and fluctuation variance, and output category labels that characterize the physical cut grade and internal clarity grade of the high refractive index transparent medium target.

[0021] Preferably, step S2 includes: step S21, smoothing and denoising each frame of the original image sequence using a spatial filter; step S22, extracting pixel feature point sets from adjacent image frames; and step S23, generating a global pixel displacement field using a dense optical flow algorithm based on the spatial coordinate offset of the pixel feature point sets between adjacent image frames.

[0022] Preferably, step S3 includes: step S31, statistically analyzing the distribution probability of motion vectors of feature points within the global pixel displacement field; step S32, identifying the peak interval with the highest proportion in the distribution probability and conforming to the spatial displacement consistency constraint as the background motion principal component; and step S33, fitting the background motion principal component with a homography transformation matrix to obtain the global rigid body motion vector.

[0023] Preferably, step S4 includes: step S41, constructing a motion compensation coordinate image model using global rigid body motion vectors; step S42, projecting the global pixel displacement field onto the motion compensation coordinate image model to cancel low-frequency macroscopic pixel displacement components; and step S43, extracting the local pixel brightness phase transition data after cancellation processing and reconstructing the brightness spatial evolution trajectory.

[0024] Preferably, in step S5, the variance of fluctuation The calculation rules are as follows: ,in, The brightness space evolution trajectory in the 1st The pixel energy integral value of each sampling node. The total number of sampling frames within the preset time window. It is the arithmetic mean of the pixel energy integral values ​​within the total number of sampled frames.

[0025] Preferably, the output category label characterizing the internal clarity grade of a high-refractive-index transparent medium target includes: step S51, monitoring pixel energy integral value. The evolution slope over time; step S52, when the evolution slope exceeds the preset negative jump threshold, and the fluctuation variance... When the deviation exceeds the preset threshold, it is determined that there is an opaque inclusion inside the high refractive index transparent medium target, and the corresponding image space coordinates are marked.

[0026] Preferably, step S6 includes: step S61, encoding the luminance spatial evolution trajectory into a feature vector with temporal characteristics; step S62, calculating the cumulative distortion distance between the feature vector and the grade template; and step S63, aligning the cumulative distortion distance on the time axis using an elastic matching algorithm to eliminate the nonlinear distortion of the time axis caused by the non-uniform motion of the light source.

[0027] Preferably, the output category label characterizing the physical cutting grade of the high refractive index transparent medium target includes: step S64, extracting the color phase angle difference of the high refractive index transparent medium target under different refractive phases according to the elastic matching algorithm; step S65, inputting the color phase angle difference and morphological similarity as independent variables into the preset classification network model to obtain the proportional symmetry grade determination result.

[0028] Preferably, before the calculation in step S2, the method further includes the following preprocessing steps: Step S91, detecting the global average brightness of the original image sequence; Step S92, when the deviation rate between the average brightness and the preset brightness benchmark exceeds 5%, performing reverse compensation on the grayscale values ​​of the original image sequence through a linear gain function based on the deviation rate, and using the compensated original image sequence as the calculation input for step S2; in step S1, the original image sequence is synchronously acquired by at least two sensing units with a distribution angle between 45° and 90°; Step S4, by fusing the brightness spatial evolution trajectories under different viewpoints, constructing a stereo optical feature tensor characterizing the internal geometric structure of the high refractive index transparent medium target.

[0029] A jewelry identification and grading system based on image recognition, comprising: The image acquisition module is used to acquire a sequence of original images of a high-refractive-index transparent medium target that changes continuously over time under a dynamic light source environment. The displacement field calculation module is used to calculate the global pixel displacement field between adjacent image frames based on the original image sequence. The noise extraction module is used to parse the global rigid body motion vector in the global pixel displacement field in order to extract the background displacement noise introduced by the external environment; The trajectory reconstruction module, connected to the noise extraction module, is used to perform inverse compensation and offset on the global pixel displacement field using the global rigid body motion vector, and strip away the brightness spatial evolution trajectory that characterizes the internal refractive properties of the high refractive index transparent medium target. The energy analysis module, connected to the trajectory reconstruction module, is used to calculate the pixel energy integral of the brightness spatial evolution trajectory within a preset time window and determine the fluctuation variance of the pixel energy integral within the preset time window. The grading output module is used to calculate the morphological similarity between the luminance spatial evolution trajectory and the grading templates in the standard grading library, and index the preset grading matrix according to the morphological similarity and fluctuation variance to output category labels that characterize the physical cut grade and internal clarity grade of the high refractive index transparent medium target.

[0030] Example 1: In the continuous operation of a large-scale automated jewelry appraisal production line, the conveyor mechanism introduces mechanical vibration when carrying a high-refractive-index transparent medium target through a dynamic light source environment. This mechanical vibration manifests as a low-frequency global rigid body displacement in the original image sequence. The background displacement noise introduced by the external physical environment and the high-frequency local fire and scintillation pixel shift caused by the complex cross-section refraction inside the transparent medium overlap in the global pixel displacement field. The overlap phenomenon causes nonlinear distortion of the optical response trajectory extracted by the conventional optical flow tracing algorithm, resulting in the loss of spatiotemporal continuity of high-frequency dynamic optical features in the two-dimensional image sequence, and producing technical problems such as cutting fire quantization distortion and misclassification.

[0031] The image acquisition module acquires a sequence of original images of a high-refractive-index transparent medium target over time. The displacement field calculation module calculates the global pixel displacement field between adjacent image frames using a dense optical flow algorithm. The noise extraction module statistically analyzes the distribution probability of feature point motion vectors within this global pixel displacement field, identifies the peak interval with the highest probability distribution that meets the spatial displacement consistency constraint as the background motion principal component, and the trajectory reconstruction module fits this background motion principal component using a homography transformation matrix to obtain the global rigid body motion vector characterizing the background displacement noise of the conveying mechanism. It then uses this global rigid body motion vector to construct a motion compensation coordinate mapping model. The trajectory reconstruction module projects the global pixel displacement field onto this motion compensation coordinate mapping model to cancel out low-frequency global pixel displacement components. The separation logic of the relative spatial displacement satisfies the mathematical relationship. ,in, To characterize the relative spatial displacement vector of local specular pixel clusters, It is an absolute spatial displacement vector. The rotation matrix is ​​obtained by decomposing the rigid body affine transformation matrix. The translation matrix is ​​obtained by decomposing the rigid body affine transformation matrix. Using the initial centroid coordinates of the local highlight pixel clusters, the trajectory reconstruction module reconstructs the intrinsic brightness spatial evolution trajectory stripped of environmental noise based on the relative spatial displacement vector and the brightness gradient values ​​of the local highlight pixel clusters between adjacent image frames. The inverse compensation hedging mechanism is based on the small displacement quasi-static assumption, that is, within the extremely short sampling period of adjacent image frames, the mechanical vibration amplitude generated by the conveying mechanism is much smaller than the physical size of the jewelry, and the nonlinear disturbance caused by the incident angle change on the internal refraction path is on the order of high-order infinitesimal. By calculating the rigid body motion vector of the background edge pixel set, the system establishes a dynamic follower reference frame attached to the outer contour of the jewelry. The algebraic subtraction operation performed in this reference frame filters out the macroscopic translation and rotation motion of the jewelry relative to the imaging sensor. The relative spatial displacement vector obtained at this time is... It can approximately linearly characterize the intrinsic displacement of the internal sectional reflection path caused by dynamic light source scanning. After completing the macroscopic displacement stripping in the single-view reference plane, the image acquisition module calls at least two sensing units with a distribution angle between 45° and 90°. The trajectory reconstruction module reconstructs the spatial optical features based on the epipolar geometry principle. The processor extracts the common local highlight pixel feature point pairs between the original images of each sensing unit, calculates the essential matrix between the corresponding camera physical coordinate systems, and calculates the absolute depth scalar of each local highlight pixel cluster in three-dimensional space. Taking the imaging view plane of the main sensing unit as the two-dimensional reference plane, the trajectory reconstruction module takes the two-dimensional brightness space evolution trajectory independently generated by each sensing unit along the calculated absolute depth scalar. The depth scalar is orthogonally stitched along its direction to output a complete 3D optical feature tensor containing lateral pixel offset, vertical pixel offset, and depth phase transition. This tensor configuration maps the spatial continuity of the optical phase transition into a data matrix for downstream computation. During the reconstruction process, based on epipolar geometric constraints, the system uses the essential matrix to perform spatial triangulation on the local specular pixel clusters captured by the main and auxiliary sensing units. By calculating the projection coordinate deviation of the specular center on the image plane at different viewpoints, and combining this with the intrinsic parameter matrix of the sensing unit and pre-calibrated relative pose parameters, the intersection point of the light beams in 3D space is solved. The resulting depth scalar and 2D pixel coordinates together constitute a 3D coordinate vector. The stereo optical feature tensor is a four-dimensional spatiotemporal manifold formed by the time sampling sequence of the coordinate vector, which is used to characterize the three-dimensional optical path evolution law of the internal reflective interface of jewelry under the excitation of dynamic light source.

[0032] The intrinsic brightness spatial evolution trajectory forms the input tensor of the downstream classification network. Based on the intrinsic brightness spatial evolution trajectory, the energy analysis module calculates the pixel energy integral within a preset time window, and determines the fluctuation variance of the pixel energy integral within the preset time window. The quantification rules for volatility variance satisfy the formula ,in, The brightness space evolution trajectory in the 1st The pixel energy integral value of each sampling node. The total number of sampling frames within the preset time window. The hierarchical output module maps the intrinsic brightness space evolution trajectory to a multi-dimensional vector space, encoding it as a feature vector, and calculates the cumulative distortion distance between this feature vector and each level template in the standard level library. An elastic matching algorithm aligns the cumulative distortion distance on the time axis, eliminating time axis distortion caused by non-uniform light source motion and extracting morphological similarity. During this process, the level templates in the standard level library are pre-associated with the physical phase labels of the light source motion. The elastic matching algorithm achieves non-linear alignment between real-time acquired time-series frames and template frames with a defined incident angle phase by searching for the minimum cost path in the distance matrix. This alignment is achieved at each time node after alignment. Each corresponds to a specific incident phase node in the template. By reading the index mapping relationship on the alignment path, the system can inversely deduce the refraction phase of the local highlight pixel cluster in the real-time image, thereby transforming the brightness fluctuation in the time domain into the optical response characteristics in the angular phase domain. The graded output module indexes the preset graded matrix based on morphological similarity and fluctuation variance, and outputs category labels representing the physical cut grade and internal clarity grade of the high refractive index transparent medium target. The processor constructs an optical flow spatial decoupling algorithm based on affine transformation constraints in the two-dimensional image pixel coordinate system. It performs spatial separation of the low-frequency global displacement caused by physical vibration and the high-frequency intrinsic optical response caused by the refraction of the internal facet of the transparent medium in the data domain according to the algebraic subtraction operator. The system eliminates the interference of external pipeline jitter on the optical phase transition characteristics from the bottom layer of image processing. The output optical response trajectory sequence maps the physical structural characteristics inside the transparent medium. The visual identification device outputs the corresponding identification grade label while maintaining the basic physical structure of the sensing unit unchanged.

[0033] Example 2: Current industrial production lines experience external mechanical vibration and light source flicker interference. The conveyor belt drive motor generates 50Hz mechanical harmonics, which are superimposed with 20dB Gaussian white noise from the dynamic light source. An image acquisition environment including an industrial camera and a variable frequency conveyor belt is used as the verification platform. The industrial camera's sampling rate is selected as 120 frames per second. The variable frequency conveyor belt carries a high-refractive-index transparent medium target through a multi-dimensional dynamic light field. The total number of sampling frames within a preset time window is limited by the constraints of real-time data acquisition and the integrity of the optical flicker period. When the light source phase switching frequency increases, the total number of sampling frames is reduced to suppress temporal aliasing of the feature tensor. The processor determines the width of the preset time window by multiplying the light source phase switching period by the camera sampling rate. The light source switching frequency is set to 2Hz, and the camera sampling rate is set to 120 frames per second. Based on the above product relationship, the total number of sampling frames is determined. It is 60.

[0034] The control group used optical flow feature extraction logic without removing background displacement noise, while the experimental group used an optical flow spatial decoupling algorithm based on affine transformation constraints. The image acquisition module acquired the original image sequence affected by 50Hz mechanical harmonic interference. Under this input condition, the peak value of the absolute spatial displacement vector of the local highlight pixel cluster reached 12.5 pixels, and the direction of the absolute spatial displacement vector showed irregular oscillation. The trajectory reconstruction module used the formula... Calculate the relative spatial displacement vector, where, To characterize the relative spatial displacement vector of local specular pixel clusters, It is an absolute spatial displacement vector. The rotation matrix is ​​obtained by decomposing the rigid body affine transformation matrix. The translation matrix is ​​obtained by decomposing the rigid body affine transformation matrix. As the initial centroid coordinates of the local highlight pixel cluster, after the experimental group stripped the global rigid body motion vector, the peak value of the output relative spatial displacement vector converged within the range of 2.1 to 2.3 pixels. This reverse compensation hedging step filters out the environmental displacement introduced by external mechanical jitter.

[0035] The energy analysis module is based on the formula Calculate the variance of the fluctuation of the brightness spatial evolution trajectory within a preset time window, where, For the variance of the fluctuation, The brightness space evolution trajectory in the 1st The pixel energy integral value of each sampling node. The total number of sampling frames within the preset time window. The total number of sampled frames is the arithmetic mean of the pixel energy integral values ​​within the total number of sampled frames. The evolution trend of fluctuation variance under conditions increasing from 15 to 120, and the total number of sampling frames. When the variance is less than 30, the fluctuation variance exhibits a discrete oscillation state. The feature extraction module outputs the spatial evolution trajectory of the broken brightness, and the total number of sampling frames. When the variance is within the range of 45 to 75, the fluctuation variance converges to a baseline range of 0.05 to 0.08. This range constitutes a high signal-to-noise ratio working window for feature extraction, with a total number of sampling frames. After 90, background noise from adjacent illumination cycles aliased, and the fluctuation variance showed a nonlinear surge. This surge caused degradation of the input tensor of the subsequent classification network. The hierarchical output module encoded the brightness spatial evolution trajectory within the high signal-to-noise ratio working window into a feature vector. The elastic matching algorithm aligned the cumulative distortion distance and extracted morphological similarity. Under the condition of continuously applying 20dB background Gaussian noise and 50Hz mechanical harmonics, the trajectory of local highlight pixel clusters in the control group broke, and the classification accuracy of the control group dropped to 72.4%. The experimental group input the brightness spatial evolution trajectory reconstructed based on the relative spatial displacement vector and the corresponding fluctuation variance. The experimental group maintained a classification accuracy of 98.7%. The system filtered out the noise interference of the conveying mechanism through the inverse compensation hedging of the global pixel displacement field and the boundary calibration of the preset time window, and output an objective category label characterizing the physical level of the high refractive index transparent medium target.

[0036] Example 3: This example combines Figures 1 to 2 This document describes a jewelry identification and grading method and system based on image recognition, such as... Figure 1 As shown, step S1 acquires the original image sequence of the high-refractive-index transparent medium target under dynamic light source environment that changes continuously over time; step S2 calculates the global pixel displacement field between adjacent image frames in the original image sequence; step S3 parses the global rigid body motion vector in the global pixel displacement field to characterize the background displacement noise introduced by the conveying mechanism; step S4 uses the global rigid body motion vector to perform reverse compensation and offset on the global pixel displacement field to separate the brightness spatial evolution trajectory characterizing the internal refractive characteristics of the high-refractive-index transparent medium target from the aliased pixel motion; step S5 calculates the pixel energy integral of the brightness spatial evolution trajectory within a preset time window and determines the fluctuation variance of the pixel energy integral within the preset time window; step S6 maps the brightness spatial evolution trajectory to a multi-dimensional vector space and calculates the morphological similarity between the multi-dimensional vector space and each grade template in the standard grade library; and step S7 indexes the preset grading matrix according to the morphological similarity and fluctuation variance, and outputs the category label characterizing the physical cut grade and internal clarity grade of the high-refractive-index transparent medium target.

[0037] like Figure 2As shown, the system executes graded tasks by constructing a hierarchical logical architecture. The system starts data processing from the root node, namely the original image sequence of the high-refractive-index transparent medium. After determining whether there is background displacement noise through the global pixel displacement field, the system identifies the principal component of background motion and analyzes the global rigid body motion vector to fit the principal component of background motion. After performing the reverse compensation hedging operation and confirming that the brightness space evolution trajectory has been obtained by stripping, the logic is divided into two parallel paths. Branch A is the internal clarity decision path, which performs pixel energy integral fluctuation analysis by calculating the fluctuation variance S. Branch B is the physical cut decision path, which performs multi-dimensional spatial morphological similarity calculation by extracting morphological similarity. Finally, the system gathers the two decision information to output the jewelry physical cut grade and internal clarity grade category label.

[0038] Example 4: In an automated identification pipeline, a conveyor mechanism carries a high-refractive-index transparent medium target through a dynamic light source environment. The internal inclusion features exhibit localized transient energy loss in the brightness space evolution trajectory. This missing signal, along with grayscale fluctuations caused by background displacement noise, aliases in the time-domain data stream. This presents a technical challenge for the system to accurately extract nonlinear brightness jump features from the synthesized pixel motion field. The displacement field calculation module constructs pixel correspondences between adjacent image frames using feature point matching technology. The noise extraction module performs multiple rounds of discrete iterations on the global pixel displacement field to fit the homography transformation matrix. In each iteration, four non-collinear pixels are randomly selected. Image point pairs are used to calculate candidate homography transformation matrices. The processor calculates the reprojection error of the remaining feature points based on the candidate homography transformation matrix and counts the number of feature points whose reprojection error is less than a preset pixel deviation value. When the proportion of feature points whose reprojection error meets the accuracy requirements reaches more than 75% of the total number of global feature points, the system locks the current candidate homography transformation matrix and stops iterating. The trajectory reconstruction module parses the global rigid body motion vector based on the locked homography transformation matrix and projects the pixel displacement of the current image frame onto the reference coordinate system to remove low-frequency macroscopic displacement components. The calculation logic of relative spatial displacement in this process is constrained by mathematical relationships. ,in, To characterize the relative spatial displacement vector of local specular pixel clusters, It is an absolute spatial displacement vector. and These are the rotation matrix and translation matrix obtained from the homography transformation matrix decomposition, respectively. The initial centroid coordinates of the local highlight pixel cluster are given.

[0039] The energy analysis module extracts the intrinsic brightness spatial evolution trajectory and calculates the pixel energy integral value. As the slope of the sampling nodes evolves, the system establishes an environmental baseline distribution by collecting flawless standard samples to determine the negative jump threshold for identifying the presence of internal inclusions. The energy analysis module statistically analyzes the standard deviation of the pixel energy integral values ​​of the standard samples throughout the entire cycle of dynamic light source phase switching. The negative jump threshold is set to the standard deviation of the fluctuation. Three times that of the differential sequence of brightness spatial evolution trajectory monitored in real time by the graded output module. When sampling node The difference at that point exceeds the negative transition threshold, and the corresponding variance of the fluctuation exceeds the threshold. When the deviation exceeds a preset threshold, the system determines that an opaque inclusion exists within the high-refractive-index transparent medium target and locks the corresponding image spatial coordinates. To eliminate misjudgments caused by light source phase switching or highlight spot movement out of the field of view, the system introduces a spatiotemporal consistency check. Since light source phase switching is a globally regular motion, the resulting brightness abrupt changes exhibit rhythmic characteristics of multiple highlight clusters evolving synchronously in spatial distribution. Energy loss caused by internal inclusions is usually limited to a specific local trajectory and has definite three-dimensional spatial coordinates. The system determines whether the negative jump of the difference sequence occurs only at a specific local coordinate and whether that coordinate remains spatially stationary in the following reference frame, thereby logically decoupling the local physical defect signal from the global light and shadow changes. The fluctuation variance... The calculation rules are as follows: ,in, The brightness space evolution trajectory in the 1st The pixel energy integral value of each sampling node. The total number of sampling frames within the preset time window. It is the arithmetic mean of the pixel energy integral values ​​within the total number of sampled frames.

[0040] The hierarchical output module aligns feature vectors with hierarchical templates using an elastic matching algorithm. This algorithm constructs a dimensional path based on a dynamic time warping path. The distance matrix, where The number of frames for the input feature vector. As the reference frame number for the grade template, the graded output module recursively calculates matrix elements frame by frame to determine the matching cost. The matching cost at each sampling moment is determined by the Euclidean distance between the current input vector component and the template vector component. The system searches for the global minimum cost path from the starting node to the ending node in the distance matrix to offset the nonlinear distortion of the time axis caused by the non-uniform motion of the light source. The cost value corresponding to the minimum path is extracted as the morphological similarity. The graded output module maps the morphological similarity and color phase angle difference as input variables to the preset classification network. Finally, it outputs category labels that characterize the physical cut grade and internal clarity grade of the high refractive index transparent medium target. The processor realizes the physical alignment of background motion noise and the intrinsic refractive signal of the target based on the iterative estimation process of the homography transformation matrix. The system uses the negative jump threshold based on environmental benchmark statistical calibration to realize the quantitative identification of internal micro-defects. The feature extraction process eliminates the dependence on ideal static working conditions, so that the grade category labels output by the visual identification device can objectively characterize the physical structural properties of the high refractive index transparent medium target.

[0041] Example 5: In the offline calibration scenario, the system fills the standard grade library by collecting a set of standard reference samples with determined physical properties. The processor adjusts the dynamic light source to perform a complete phase evolution action within a preset light position switching cycle. The image acquisition module acquires the original image sequence of the standard reference sample at a fixed sampling frequency of 120Hz. The trajectory reconstruction module calls the optical flow spatial decoupling algorithm to remove the residual displacement component and performs tensor normalization mapping on the extracted intrinsic brightness spatial evolution trajectory. The normalized data is stored in the memory as a grade template and associated with the corresponding grade category label. This process ensures that the feature vectors calculated in the subsequent online identification process have objective physical consistency with the grade template in terms of energy dimensions and spatial manifold through preset data alignment logic.

[0042] When the system faces a new physical identification environment or a replacement of the image acquisition unit, the processor initiates a pre-baseline calibration program to determine the initial operating parameters. Under the condition of no-load operation of the conveyor mechanism, the image acquisition module acquires a background image sequence and sends it to the displacement field calculation module to determine the global pixel displacement field. The noise extraction module statistically analyzes the statistical distribution characteristics of this global pixel displacement field to calculate the intrinsic noise standard deviation under the current environment. The system is based on the intrinsic noise standard deviation. The exposure time of the image acquisition module is adjusted according to a linear function relationship with a preset signal-to-noise ratio threshold, and the environmental zero-point reference is determined by integrating the pixel energy of the empty background after the dynamic light source is turned on. The calibration process eliminates the influence of individual sensor differences and external light intensity fluctuations, and puts the system into a preset ready state.

[0043] Example 6: In the standard grade library construction scenario, the image acquisition module acquires a reference image sequence of a standard sample with determined physical properties. The processor drives a dynamic light source to switch the array light position of 4×8 light-emitting units in a preset time sequence within a 3D coordinate system. The displacement field calculation module performs pixel displacement decoupling operations on the acquired original image sequence, converting the extracted intrinsic brightness spatial evolution trajectory into a sequence containing... The system stores the feature vectors of each component in an index table in memory. It then determines the center point vector of that level by analyzing the feature vector distribution characteristics of over 100 sets of standard samples at the same level. and its distribution density gradient This density gradient provides a statistical distribution benchmark for the elastic matching logic in the subsequent detection stage; simultaneously, the processor performs color separation processing on the color components of local highlight pixel clusters to extract hue component feature values, according to the formula... Calculate the color phase angle difference for each sampling node. ,in, For color phase angle difference, The blue difference component, The red difference component represents the phase angle difference of this color. The distribution range that evolves over time is stored in the grading template as a criterion for evaluating cutting quality. The grading output module calls a classification network model that uses a support vector machine classifier based on radial basis function kernel function. The processor extracts the morphological similarity scalar of the feature vector after elastic matching and alignment, and modifies it with the color phase angle difference. The root mean square (RMS) values ​​of the time series are merged into a two-dimensional input evaluation vector, which is mapped to the high-dimensional feature space of the support vector machine (SVM) classifier. The classifier establishes the hyperplane decision boundary based on pre-trained standard grade sample datasets and outputs discrete grade values ​​representing the symmetry of the target's physical cutting ratio. In the specific mapping logic, the classifier constructs a two-dimensional feature vector using the normalized morphological similarity scalar and the RMS value of the color phase angle difference as the input space. The pre-set hyperplane decision boundary maps this input space to a high-dimensional kernel space through a radial basis function (RBF) kernel function. Within the kernel space, a nonlinear decision surface is constructed using the weights of the trained support vectors. When a real-time input feature point falls into a specific geometric region divided by the hyperplane, the system maps it to a pre-set discrete grade according to the region index, for example, 0 represents Excellent (EX) and 1 represents Very Good (VG). This achieves a structured quantitative conclusion from multi-dimensional optical features to the cutting ratio grade. The processor addresses the pre-set grade matrix using a two-dimensional array lookup table logic, defining the row address index variable of the grade matrix as the fluctuation variance obtained statistically based on a pre-set time window. Within the specified numerical range, the column address index variable is a proportionally symmetric discrete grade value. The processor verifies the row and column address index variables calculated in real time and reads the pre-written comprehensive evaluation conclusion from the corresponding cross data storage unit of the grade matrix, outputting the final category label characterizing the physical cut grade and internal clarity grade of the high refractive index transparent medium target.

[0044] When the system encounters fluctuations in ambient light intensity or gain drift in the image acquisition unit, the processor drives the system to sample pixel energy integrals on a neutral gray calibration plate with constant reflectivity during the self-test phase before detection begins. The energy analysis module calculates the environmental attenuation factor based on the ratio of the sampled reference energy value to the measured energy integral value under the current environment. The processor utilizes environmental degradation factors A linear gain correction is applied to the fluctuation variance threshold in the preset classification matrix. If the measured pixel energy integral after correction is lower than the preset minimum sensitivity level, the processor outputs an alarm signal and blocks the output action of the classification tag. This process enables the identification system to maintain the consistency of physical dimensions in the extraction of brightness spatial evolution trajectory under different physical environments through real-time monitoring of environmental disturbance parameters and closed-loop correction of judgment threshold.

[0045] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A jewelry identification and grading method based on image recognition, characterized in that, Includes the following steps: Step S1: Obtain the original image sequence of a high refractive index transparent medium target that changes continuously over time under a dynamic light source environment; Step S2: Calculate the global pixel displacement field between adjacent image frames in the original image sequence; Step S3: Analyze the global rigid body motion vector in the global pixel displacement field. The global rigid body motion vector represents the background displacement noise introduced by the conveying mechanism. Step S4: Use the global rigid body motion vector to perform reverse compensation and offset on the global pixel displacement field, so as to separate the brightness spatial evolution trajectory that characterizes the internal refractive properties of the high refractive index transparent medium target from the aliased pixel motion. Step S5: Calculate the pixel energy integral of the brightness spatial evolution trajectory within a preset time window, and determine the fluctuation variance of the pixel energy integral within the preset time window. Step S6: Map the brightness space evolution trajectory to the multidimensional vector space, and calculate the morphological similarity between the multidimensional vector space and the templates of each level in the standard level library. Step S7: Index the preset grading matrix according to morphological similarity and fluctuation variance, and output category labels that characterize the physical cut grade and internal clarity grade of the high refractive index transparent medium target.

2. The jewelry identification and grading method based on image recognition according to claim 1, characterized in that, Step S2 includes: Step S21, smoothing and denoising each frame of the original image sequence using a spatial filter; Step S22, extracting pixel feature point sets from adjacent image frames; Step S23, generating a global pixel displacement field using a dense optical flow algorithm based on the spatial coordinate offset of the pixel feature point sets between adjacent image frames.

3. The jewelry identification and grading method based on image recognition according to claim 1, characterized in that, Step S3 includes: Step S31, statistically analyzing the distribution probability of feature point motion vectors within the global pixel displacement field; Step S32, identifying the peak interval with the highest proportion in the distribution probability that meets the spatial displacement consistency constraint as the background motion principal component; Step S33, fitting the background motion principal component through the homography transformation matrix to obtain the global rigid body motion vector.

4. The jewelry identification and grading method based on image recognition according to claim 1, characterized in that, Step S4 includes: Step S41, constructing a motion compensation coordinate image model using global rigid body motion vectors; Step S42, projecting the global pixel displacement field onto the motion compensation coordinate image model to cancel low-frequency macroscopic pixel displacement components; Step S43, extracting the local pixel brightness phase transition data after cancellation processing and reconstructing the brightness spatial evolution trajectory.

5. The jewelry identification and grading method based on image recognition according to claim 1, characterized in that, In step S5, the variance of fluctuation The calculation rules are as follows: ,in, The brightness space evolution trajectory in the 1st The pixel energy integral value of each sampling node. The total number of sampling frames within the preset time window. It is the arithmetic mean of the pixel energy integral values ​​within the total number of sampled frames.

6. The jewelry identification and grading method based on image recognition according to claim 5, characterized in that, Outputting category labels characterizing the internal clarity grade of a high-refractive-index transparent medium target includes: Step S51, monitoring pixel energy integral values. The evolution slope over time; step S52, when the evolution slope exceeds the preset negative jump threshold, and the fluctuation variance... When the deviation exceeds the preset threshold, it is determined that there is an opaque inclusion inside the high refractive index transparent medium target, and the corresponding image space coordinates are marked.

7. The jewelry identification and grading method based on image recognition according to claim 1, characterized in that, Step S6 includes: Step S61, encoding the luminance spatial evolution trajectory into a feature vector with temporal characteristics; Step S62, calculating the cumulative distortion distance between the feature vector and the level template; Step S63, aligning the cumulative distortion distance on the time axis using an elastic matching algorithm to eliminate the nonlinear distortion of the time axis caused by the non-uniform motion of the light source.

8. The jewelry identification and grading method based on image recognition according to claim 7, characterized in that, The output category label characterizing the physical cutting grade of the high refractive index transparent medium target includes: step S64, extracting the color phase angle difference of the high refractive index transparent medium target under different refractive phases according to the elastic matching algorithm; step S65, inputting the color phase angle difference and morphological similarity as independent variables into the preset classification network model to obtain the proportional symmetry grade determination result.

9. The jewelry identification and grading method based on image recognition according to claim 1, characterized in that, Before the calculation in step S2, the method also includes the following preprocessing steps: Step S91, detecting the global average brightness of the original image sequence; Step S92, when the deviation rate between the average brightness and the preset brightness benchmark exceeds 5%, the gray values ​​of the original image sequence are inversely compensated according to the deviation rate through a linear gain function, and the compensated original image sequence is used as the calculation input for step S2. In step S1, the original image sequence is synchronously acquired by at least two sensing units with a distribution angle between 45° and 90°; Step S4, by fusing the brightness spatial evolution trajectory under different viewpoints, a stereo optical feature tensor characterizing the internal geometric structure of the high refractive index transparent medium target is constructed.

10. A jewelry appraisal and grading system based on image recognition, used to implement the jewelry appraisal and grading method based on image recognition as described in claim 1, characterized in that, include: The image acquisition module is used to acquire a sequence of original images of a high-refractive-index transparent medium target that changes continuously over time under a dynamic light source environment. The displacement field calculation module is used to calculate the global pixel displacement field between adjacent image frames based on the original image sequence. The noise extraction module is used to parse the global rigid body motion vector in the global pixel displacement field in order to extract the background displacement noise introduced by the external environment; The trajectory reconstruction module, connected to the noise extraction module, is used to perform inverse compensation and offset on the global pixel displacement field using the global rigid body motion vector, and strip away the brightness spatial evolution trajectory that characterizes the internal refractive properties of the high refractive index transparent medium target. The energy analysis module, connected to the trajectory reconstruction module, is used to calculate the pixel energy integral of the brightness spatial evolution trajectory within a preset time window and determine the fluctuation variance of the pixel energy integral within the preset time window. The grading output module is used to calculate the morphological similarity between the luminance spatial evolution trajectory and the grading templates in the standard grading library, and index the preset grading matrix according to the morphological similarity and fluctuation variance to output category labels that characterize the physical cut grade and internal clarity grade of the high refractive index transparent medium target.

Images