Egg surface microcrack detection system and method based on image analysis

Abstract

The invention discloses an egg surface microcrack detection system and method based on image analysis, and relates to the field of surface defect nondestructive detection.The method comprises the steps that a process identification interface module obtains a processing process identifier of an egg in real time; the multi-angle annular light source module is provided with a multi-band annular light source array comprising visible light and near-infrared LED lamp beads which are independently controlled, a polaroid group and a beam splitter prism are integrated, and visible light polarization images and near-infrared polarization images on the surfaces of the eggs are synchronously collected; the thermal excitation enhancement unit obtains a thermal infrared image; the process filtering module extracts the axial elongation of the mechanical stress microcrack after graded transportation, and calculates the mesh fractal dimension of the thermal stress microcrack after UV disinfection; the multi-modal feature fusion module generates a dual-channel crack probability graph; and the dynamic feedback control module outputs a micro-crack risk grade and adjusts the rotating speed of the objective table and the light source intensity. The method has the advantages that accurate judgment of mechanical and thermal stress cracks is realized, and curved surface reflection interference is broken through.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-destructive testing of surface defects, in particular to an egg surface micro-crack detection system and method based on image analysis. BACKGROUND

[0002] Eggs are prone to surface micro-cracks due to mechanical impact or thermal stress during processing procedures such as washing, grading transportation, UV disinfection, and oil coating. These micro-cracks are difficult to identify by the naked eye, but they can significantly reduce the shelf life of eggs and increase the risk of microbial invasion. Existing detection techniques mainly rely on visible light imaging combined with morphological processing. However, during actual operation, the curved structure of the egg leads to uneven illumination of the fixed light source, the high light area covers the crack features, and the curved surface reflection interferes. At the same time, the characteristic differences between mechanical stress cracks (axial extension) and thermal stress cracks (fractal network) are not targeted, making it difficult to distinguish the types of cracks. A single sensor cannot simultaneously capture surface texture (visible light), subsurface structure (near-infrared), and thermal conductivity characteristics (thermal infrared), making the image basis for judgment more one-sided. In addition, the detection parameters are not dynamically adjusted according to the production line procedures, and the efficiency and accuracy are difficult to balance. SUMMARY

[0003] To solve the above technical problems, an egg surface micro-crack detection system and method based on image analysis are provided. This technical solution solves the problem of high micro-crack miss rate caused by the fixed detection parameters being difficult to adapt to the differentiated characteristics of mechanical stress and thermal stress cracks, and being easily disturbed by reflection in the curved surface imaging of eggs.

[0004] To achieve the above purposes, the technical solution adopted by the present application is as follows:

[0005] The egg surface micro-crack detection system and method based on image analysis include:

[0006] A process identification interface module is connected to the production line control system to obtain the processing procedure identification of the eggs in real time. The procedure identification includes the washing, grading transportation, UV disinfection, and oil coating procedure states.

[0007] A multi-angle ring light source module is provided with a multi-wavelength ring light source array containing independently controlled visible light and near-infrared LED lamp beads. The incident angle of each light source is adjusted in real time based on the egg curvature model, and a polarizer group and a light splitting prism are integrated to simultaneously collect visible light polarization images and near-infrared polarization images of the egg surface.

[0008] A thermal excitation enhancement unit is integrated with a pulse laser at the center of the light source array. The temperature difference in the crack area is excited by a short-time thermal shock to enhance the contrast of infrared thermal imaging and obtain thermal infrared images.

[0009] Process filtering module: select filtering path based on process identification, perform directional constraint morphological filtering on visible light image after hierarchical transportation, strengthen mechanical stress cracks, and extract axial elongation of mechanical stress microcracks; perform anisotropic diffusion filtering on thermal infrared image after UV disinfection, enhance thermal stress cracks, and calculate network fractal dimension of thermal stress microcracks;

[0010] Multimodal feature fusion module: based on visible light image, near-infrared image and thermal infrared image, extract crack edge features and crack projection features based on polarization angle offset, fuse visible light texture features and near-infrared depth features, generate double-channel crack probability map, crack projection features include vertical maximum peak value, vertical projection range, circularity, light texture gradient, near-infrared penetration depth and local temperature rise;

[0011] Dynamic feedback control module: build crack depth distribution model based on process identification, output microcrack risk level according to axial elongation, network fractal dimension and depth value, call historical crack judgment threshold of corresponding process, output crack positioning result with process label and confidence score, and adjust stage speed and light source intensity.

[0012] Preferably, the incident angle of each light source is adjusted in real time based on the egg curvature model, which specifically includes:

[0013] The three-dimensional vision sensor is used to acquire the point cloud data of the egg surface in real time, and an ellipsoid equation is fitted to establish an egg curvature model, wherein the ellipsoid equation takes the major axis of the egg and the equatorial radius as parameters;

[0014] The critical values of optical scattering attenuation characteristics, curved spot and polarization modulation cross-validation are obtained, the principal curvature vector of the current detection point is calculated according to the egg curvature model, and the universal adjusting mechanism of each LED lamp bead in the ring-shaped light source array is driven to make the angle between the incident light direction and the surface normal vector less than the critical value;

[0015] The reflecting light is separated into two imaging channels by the synchronous control of the light splitting prism, the first channel generates a visible light polarization image through a 490nm-600nm band-pass filter and a 0° polarizer, and the second channel generates a near-infrared polarization image through an 850nm band-pass filter and a 90° polarizer;

[0016] The thermal infrared image is acquired by a medium wave infrared thermal imager within a few seconds after the 500μs thermal shock of the pulsed laser, and the thermal shock duration is determined by matching experiments of the sampling period of the infrared thermal imager and the maximum temperature rise response time of the crack area.

[0017] Preferably, the filtering path based on process identification specifically includes:

[0018] A linear structural element parallel to the long axis is generated based on the egg curvature model after the hierarchical transportation process, and a directional erosion operation is performed along the projection direction of the normal of the equatorial plane of the egg to suppress the interference noise of non-mechanical stress cracks;

[0019] A directional dilation operation is used to strengthen the continuity of the axial cracks and perform constraint morphological filtering;

[0020] The filtered binary image is skeletonized, the angle deviation θ between the main axis direction of the crack and the long axis of the egg is calculated, and the effective extension length L e is calculated based on the angle deviation through L e , and the axial extension rate η is calculated through and output to the dynamic feedback control module, where L0 is the original projection length and D is the equatorial diameter of the egg;

[0021] After the UV disinfection process, a diffusion tensor is constructed based on the thermal gradient field, and the thermal conductivity coefficient is enhanced along the normal direction of the crack;

[0022] The heat noise in the non-crack area is smoothed through partial differential equation iteration, the crack connected domain of the filtered image is extracted, and the network fractal dimension of the thermal stress crack is calculated using box counting method.

[0023] Preferably, the process filtering module further comprises:

[0024] Process logic judgment unit: real-time analysis of process identification, when the identification is hierarchical transportation, trigger directional constraint morphological filtering, when the identification is UV disinfection, trigger thermal excitation enhancement unit synchronously, and start anisotropic diffusion filtering;

[0025] Filtering path switching unit: disable crack strengthening filtering during cleaning and oiling process, and only perform basic noise reduction;

[0026] Based on the numerical range of the axial extension rate η and the network fractal dimension, the weight distribution factor of the subsequent multi-modal feature fusion module is dynamically adjusted, wherein the numerical range of the network fractal dimension is obtained through historical experimental data, η>0.35 determines the mechanical stress crack, and the fractal dimension>1.2 determines the thermal stress crack.

[0027] Preferably, the extraction of crack edge features and crack projection features based on polarization angle offset, fusion of visible light texture features and near-infrared depth features, and generation of a double-channel crack probability map specifically includes:

[0028] The gradient amplitude difference of 0° and 90° polarization direction is calculated for the visible light polarization image and the near-infrared polarization image respectively, a polarization angle offset feature map is generated, and the maximum value point in the feature map is extracted as the crack edge feature;

[0029] extract vertical projection features, geometric and optical features along the long axis direction of the egg, including vertical maximum wave peak value, vertical projection range, circularity, light texture gradient, near-infrared penetration depth and local temperature rise;

[0030] The polarization angle offset feature and the light texture gradient in the crack projection feature are combined into a fusion feature vector, which is input into the polarization suppression channel to generate an edge probability map that suppresses the curved surface reflection interference;

[0031] The near-infrared penetration depth and the local temperature rise in the crack projection feature are linearly weighted to generate a temperature difference enhancement probability map;

[0032] Based on the weight distribution factor output by the process filtering module, the mechanical stress related features and the thermal stress related features are dynamically weighted and output to the edge probability map and the temperature difference enhancement probability map respectively to form a double-channel crack probability map, wherein the mechanical stress related features include the vertical maximum wave peak value and the vertical projection range, and the thermal stress related features include the network fractal dimension and the local temperature rise.

[0033] Preferably, the vertical projection features, geometric and optical features extracted along the long axis direction of the egg specifically include:

[0034] Based on the egg curvature model, all projection features are extracted along the equatorial plane normal projection direction;

[0035] The vertical projection range is calculated by the gray projection curve along the long axis direction of the egg, and is located in the interval where the curve value is greater than the maximum wave peak value;

[0036] The light texture gradient is quantified by the mean value of the Sobel operator gradient amplitude of the crack area in the visible light image;

[0037] The near-infrared penetration depth is calculated by the pixel intensity decay rate of the crack area in the near-infrared polarization image;

[0038] The circularity is determined based on the ratio of the short axis to the long axis of the minimum circumscribed ellipse of the crack connected domain;

[0039] The local temperature rise is extracted by the maximum temperature difference between the crack area and the adjacent normal area in the thermal infrared image.

[0040] Preferably, the crack depth distribution model is constructed based on the process identification, and the micro-crack risk level is output according to the axial extension rate, the network fractal dimension and the depth value, specifically including:

[0041] Based on the process identification selection input parameter, when the process identification is classified transportation, the axial extension rate and the local temperature rise value are input; when the process identification is UV disinfection, the network fractal dimension and the local temperature rise value are input;

[0042] The depth estimation value of the process identified as the grading transportation is calculated based on the linear weighting of the axial elongation rate and the local temperature rise value, and the depth estimation value of the process identified as the UV sterilization is calculated based on the reticular fractal dimension and the local temperature rise value, wherein the weight coefficient is calibrated by the historical crack sample;

[0043] The average thickness of the eggshell is 0.3-0.4mm, the risk level is associated with the eggshell thickness, when the crack depth does not exceed half of the eggshell thickness, it is recorded as low risk, when the crack depth reaches between half and equal to the eggshell thickness, it is recorded as medium risk, and when the crack depth significantly exceeds the eggshell thickness, it is recorded as high risk;

[0044] Based on the depth estimation values calculated by different processes, the risk level of the current process is compared, and the micro-crack risk level is output.

[0045] Preferably, the output crack positioning result with process label and confidence score adjusts the rotation speed of the stage and the intensity of the light source, which specifically includes:

[0046] When the stage rotation makes the current detection point of the egg reach the laser incidence area of the thermal excitation enhancement unit, based on the surface normal vector output by the egg curvature model, the 500μs short-time thermal shock of the pulsed laser is triggered to make the transient temperature rise of the crack excitation area, and the goniometer adjustment mechanism of the LED lamp bead is adjusted in real time to make the angle between the light source incidence direction and the surface normal vector less than the angle threshold, and the angle threshold is obtained through the cross-validation experiment of the optical scattering attenuation characteristics and the polarization modulation;

[0047] Within a few seconds after the thermal shock ends, the visible light polarization image of the first channel and the near-infrared polarization image of the second channel are synchronously collected through the spectrometer prism, and the thermal infrared image is collected through the mid-wave infrared thermal imager;

[0048] If the risk level is low, the rotation speed of the stage is maintained at the first rotation speed, and the intensity of the annular light source is increased, the first rotation speed is calculated by matching the equatorial circumference and the resolution of the imaging system, and the light source intensity increase value is determined through the compensation experiment of the visible light texture gradient;

[0049] If the risk level is medium, the rotation speed of the stage is reduced to the second rotation speed, and the secondary thermal excitation is triggered to enhance the crack contrast, and the second rotation speed is determined by the mapping relationship between the minimum time required for complete sampling of thermal diffusion and the angular velocity of the stage;

[0050] If the risk level is high, the stage rotation is immediately stopped, and an alarm signal with a UV sterilization process label is output.

[0051] Preferably, the egg surface micro-crack detection method based on image analysis comprises:

[0052] Connect the production line control system, real-time acquisition of the processing procedure identification of the eggs, the procedure identification including the cleaning, grading transportation, UV disinfection, and oil coating procedure states;

[0053] A multi-band ring light source array containing independently controlled visible light and near-infrared LED lamp beads is arranged, the incidence angle of each light source is adjusted in real time based on the egg curvature model, and a polarizer group and a light splitting prism are integrated to synchronously collect visible light polarization images and near-infrared polarization images on the surface of the egg, a pulsed laser is integrated in the center of the light source array, the temperature difference of the crack area is excited through short-time thermal shock to enhance the contrast of infrared thermal imaging, and a thermal infrared image is acquired;

[0054] Based on the procedure identification, a direction-constrained morphological filter is performed on the visible light image after grading transportation to strengthen mechanical stress cracks and extract the axial elongation rate of the mechanical stress micro-cracks, an anisotropic diffusion filter is performed on the thermal infrared image after UV disinfection to enhance thermal stress cracks and calculate the network fractal dimension of the thermal stress micro-cracks;

[0055] Based on the visible light image, the near-infrared image and the thermal infrared image, crack edge features and crack projection features based on polarization angle offset are extracted, visible light texture features and near-infrared depth features are fused, a double-channel crack probability map is generated, and the crack projection features include the vertical maximum peak value, the vertical projection range, the circularity, the light texture gradient, the near-infrared penetration depth and the local temperature rise;

[0056] Based on the procedure identification, a crack depth distribution model is constructed, the micro-crack risk level is output according to the axial elongation rate, the network fractal dimension and the depth value, the historical crack determination threshold of the corresponding procedure is called, the crack positioning result with the procedure label and the confidence score are output, and the rotation speed of the object table and the light source intensity are adjusted.

[0057] Compared with the prior art, the beneficial effects of the present application are as follows:

[0058] The present application proposes an egg surface micro-crack detection system and method based on image analysis, an egg ellipsoid curvature model is fitted through three-dimensional point cloud, surface normal vectors are calculated, LED incidence angles are adjusted in real time to suppress curved surface reflection, and dynamic optical regulation is formed. Meanwhile, the light splitting prism separates the double channels of the visible light polarization image (0° polarization) and the near-infrared polarization image (90° polarization), and the thermal shock and infrared sampling time sequence are matched, so that the crack area illumination uniformity is improved through dynamic light adjustment based on the curvature model, and the polarization imaging signal-to-noise ratio is improved. Meanwhile, the time difference of synchronous acquisition of three modalities (visible light texture, near-infrared penetration depth, and thermal infrared temperature rise) is less than 100 ms, and motion blur is avoided;

[0059] The application provides an egg surface micro-crack detection system and method based on image analysis, which forms a differential filter engine through directional constraint morphological filtering in a mechanical stress channel and anisotropic diffusion filtering in a thermal stress channel. In combination with a dynamic weighted fusion device, a polarization angle offset and a projection feature are input to obtain a double-channel crack probability map, and a stage-light linkage strategy is obtained based on a crack depth model, so that the mechanical and thermal stress crack identification accuracy is improved through process identification driven feature weight distribution. At the same time, the detection efficiency is also improved through risk classification adjustment of the stage rotation speed. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The figure is a flowchart of the application;

[0061] Figure 2 The figure is a flowchart of real-time adjustment of the incidence angles of the light sources in the application based on an egg curvature model;

[0062] Figure 3 The figure is a flowchart of selection of a filtering path based on a process identification in the application;

[0063] Figure 4 The figure is a flowchart of a process filtering module further included in the application;

[0064] Figure 5 The figure is a flowchart of extraction of crack edge features and crack projection features based on a polarization angle offset, fusion of visible light texture features and near-infrared depth features, and generation of a double-channel crack probability map in the application;

[0065] Figure 6 The figure is a flowchart of extraction of vertical projection features, geometric and optical features along the egg long axis direction in the application;

[0066] Figure 7 The figure is a flowchart of construction of a crack depth distribution model based on a process identification in the application, output of a micro-crack risk level according to an axial elongation rate, a network fractal dimension, and a depth value;

[0067] Figure 8 The figure is a flowchart of output of a crack positioning result with a process label and a confidence score, and adjustment of a stage rotation speed and a light source intensity in the application. DETAILED DESCRIPTION

[0068] The following description is used to disclose the application so that those skilled in the art can implement the application. The preferred embodiments in the following description are only used as examples, and other obvious modifications can be thought of by those skilled in the art.

[0069] REFERENCE Figure 1 As shown in the figure, the egg surface micro-crack detection system and method based on image analysis comprises:

[0070] Process identification interface module: connected to the production line control system, real-time acquisition of the processing process identification of the eggs, including cleaning, grading transportation, UV disinfection, and oil coating process status;

[0071] Multi-angle ring light source module: set up a multi-band ring light source array containing independently controlled visible light and near-infrared LED lamp beads, the incidence angle of each light source is adjusted in real time based on the egg curvature model, and a polarizer group and a light splitting prism are integrated to simultaneously collect visible light polarization images and near-infrared polarization images on the surface of the egg;

[0072] Thermal excitation enhancement unit: integrate a pulsed laser in the center of the light source array, excite the temperature difference of the crack area through short-time thermal shock, enhance the contrast of infrared thermal imaging, and obtain thermal infrared images;

[0073] Process filtering module: based on the process identification, select the filtering path, perform directional constraint morphological filtering on the visible light image after grading transportation, strengthen the mechanical stress cracks, and extract the axial elongation rate of the mechanical stress micro-cracks, perform anisotropic diffusion filtering on the thermal infrared image after UV disinfection, enhance the thermal stress cracks, and calculate the network fractal dimension of the thermal stress micro-cracks;

[0074] Multi-modal feature fusion module: based on the visible light image, near-infrared image and thermal infrared image, extract crack edge features and crack projection features based on polarization angle offset, fuse visible light texture features and near-infrared depth features, generate a double-channel crack probability map, and crack projection features include vertical maximum peak value, vertical projection range, circularity, light texture gradient, near-infrared penetration depth and local temperature rise;

[0075] Dynamic feedback control module: based on the process identification, construct a crack depth distribution model, output the micro-crack risk level according to the axial elongation rate, network fractal dimension, and depth value, call the historical crack judgment threshold of the corresponding process, output the crack positioning result with process label and confidence score, and adjust the rotation speed of the stage and the intensity of the light source.

[0076] The process recognition interface module communicates with the production line PLC system in real time to obtain the current process identification of the eggs (cleaning, grading transportation, UV disinfection, and oiling), and based on the process dynamic switching detection strategy, a process perception layer is constructed. Through the optical imaging layer, a multi-band ring light source array (visible light + near-infrared LED) is used to realize double-channel synchronous imaging by integrating a polarizer and a beam splitter. At the same time, a thermal excitation unit (pulsed laser) is used to excite the transient temperature rise of the crack area, thereby enhancing the contrast of the thermal infrared image. The optical imaging layer is constructed. The feature fusion layer selects a differentiated filtering path according to the process type, extracts the axial elongation rate of mechanical stress cracks or the network fractal dimension of thermal stress cracks, and fuses the three modal image features to generate a double-channel crack probability map. The decision layer outputs the risk level based on the crack depth model and the historical threshold library, dynamically adjusts the rotation speed of the stage and the intensity of the light source, and realizes real-time optimization of the detection parameters. The process perception layer, the optical imaging layer, the feature fusion layer, and the decision control layer can form a closed-loop control egg surface micro-crack detection system.

[0077] In use, the process perception layer identifies that the egg enters the grading transportation state, triggering directional constraint morphological filtering. The ring light source automatically adjusts the LED incident angle based on the egg ellipsoid curvature model (ensuring that the angle between the light and the normal is less than a critical value), and synchronously collects visible light polarization images (490-600 nm, 0° polarization) and near-infrared polarization images (850 nm, 90° polarization). Then, directional erosion operation is performed on the visible light image (projection along the equatorial plane normal), and non-mechanical stress noise is suppressed; directional dilation operation is used to strengthen the axial crack continuity, and after skeletonization, the angle deviation θ between the crack principal axis and the major axis is calculated, and the axial elongation rate η is output through the formula. If η>0.35, it is determined as a mechanical stress crack, and the dynamic feedback module calls the historical threshold of the grading transportation process, increases the light source intensity and maintains the basic rotation speed of the stage, and outputs the high-confidence crack positioning result.

[0078] The process state of the production line can be used as a decision variable of the detection strategy to specifically enhance the crack features caused by different stress sources (mechanical / thermal stress), and to solve the problem of interference in egg curved surface imaging by combining three modal data of polarized light (suppressing curved surface reflection), near-infrared (penetrating depth perception), and thermal infrared (temperature difference enhancement). At the same time, through real-time adjustment of the rotation speed of the stage and the intensity of the light source, the detection efficiency and accuracy can be effectively balanced (such as high-risk stop alarm and low-risk speed-up scanning). In addition, the crack positioning result with process label and confidence score is output, directly guiding the production line sorting decision, and reducing the cost of manual re-inspection.

[0079] Referring to Figure 2 The incident angle of each light source is adjusted in real time based on the egg curvature model, which specifically includes:

[0080] The surface point cloud data of the egg is acquired in real time by a three-dimensional vision sensor, and an ellipsoid equation is fitted to establish an egg curvature model, wherein the ellipsoid equation takes the long axis and the equatorial radius of the egg as parameters;

[0081] The critical value of the optical scattering attenuation characteristic, the curved surface light spot and the polarization modulation cross verification is obtained, the principal curvature vector of the current detection point is calculated according to the egg curvature model, the gimbal adjusting mechanism of each LED lamp bead in the annular light source array is driven, so that the angle between the incident light and the surface normal vector is less than the critical value;

[0082] The reflected light is separated into two imaging channels by a dichroic prism, the first channel generates a visible light polarization image through a 490nm-600nm band-pass filter and a 0° polarizer, and the second channel generates a near-infrared polarization image through an 850nm band-pass filter and a 90° polarizer.

[0083] Within a few seconds after the pulsed laser emits a 500μs thermal shock, a thermal infrared image is acquired by a medium-wave infrared thermal imager, and the thermal shock duration is determined by matching experiments of the infrared thermal imager sampling period and the maximum temperature rise response time of the crack area.

[0084] High-precision point cloud data of the egg surface is captured in real time by a three-dimensional vision sensor (such as a structured light or a laser scanner), and an ellipsoid equation specific to the egg (parameters: long axis a and equatorial radius b) is fitted by the least squares method to establish a dynamically updated egg curvature model. Meanwhile, the annular light source array is composed of independently controlled visible light (490-600nm) and near-infrared (850nm) LED lamp beads, each lamp bead is integrated with a gimbal adjusting mechanism (such as a multi-axis gimbal driven by a miniature stepper motor), and is equipped with a polarizer group and a dichroic prism optical assembly.

[0085] During its dynamic light control, the principal curvature vector of the current detection point is calculated based on the curvature model, and the critical value is determined by the scattering attenuation characteristic, the curved surface light spot distortion threshold and the polarization modulation effect cross verification through optical experiment calibration. The gimbal mechanism adjusts the incident angle of each LED in real time to ensure that the angle between the incident light and the surface normal vector is always less than the critical value (for example, ≤15°), forming a dual-channel synchronous imaging. At the same time, the dichroic prism separates the reflected light into two imaging channels: the first channel generates a visible light polarization image that suppresses specular reflection through a 490-600nm band-pass filter and a 0° polarizer; the second channel generates a near-infrared polarization image that enhances the contrast of deep cracks through an 850nm band-pass filter and a 90° polarizer. In addition, after the pulsed laser applies a 500μs short-time thermal shock, the medium-wave infrared thermal imager is triggered to acquire a transient thermal infrared image according to the matching relationship between the infrared thermal imager sampling period and the maximum temperature rise response time of the crack area (the delay time from the end of the thermal shock to the peak temperature rise is determined by pre-experiment), forming a thermal excitation cooperative triggering mode.

[0086] That is, the uneven illumination problem caused by the fixed light source on the egg curvature can be overcome by the dynamic light adjustment driven by the curvature model, the crack area illumination uniformity is improved by avoiding the high light area from covering the crack, the polarization image signal-to-noise ratio is improved, and the visible light texture details and near-infrared penetration depth information are synchronously obtained through the collaborative design of the light splitting prism and the double-channel polarization filter, thereby solving the problem that a single light source cannot detect the surface and subsurface simultaneously. Based on the 500 μs thermal shock duration matched with the temperature rise response characteristics, the temperature difference contrast between the crack area and the non-crack area is enhanced (the local temperature rise feature extraction error of the thermal infrared image is less than or equal to 0.5 °C), thereby avoiding the temperature difference signal ambiguity caused by the timing mismatch. In addition, the gimbal mechanism tracks the normal vector in real time, strictly limits the incident angle within a critical value, effectively suppresses the interference of the curved surface reflection on the polarization imaging, and improves the crack detection rate.

[0087] Referring to Figure 3 The selecting filter path based on the process identifier specifically includes:

[0088] After the hierarchical transportation process, a linear structural element parallel to the major axis is generated based on the egg curvature model, a directional erosion operation is performed along the projection direction of the egg equatorial plane normal, and the interference noise points of non-mechanical stress cracks are suppressed;

[0089] Directional dilation operation is used to strengthen the axial crack continuity for constraint morphological filtering;

[0090] Skeletonization processing is performed on the filtered binary image, the angle deviation θ between the crack principal axis direction and the egg major axis is calculated, and the effective extension length L e is calculated based on the angle deviation through L e , and the axial extension rate η is calculated through and output to the dynamic feedback control module, where L0 is the original projection length, and D is the equatorial diameter of the egg;

[0091] After the UV disinfection process, a diffusion tensor is constructed based on the thermal gradient field, and the thermal conductivity coefficient is enhanced along the crack normal;

[0092] The heat noise in the non-crack area is smoothed through the partial differential equation iteration, the crack connected domain of the filtered image is extracted, and the net-like fractal dimension of the thermal stress crack is calculated by using the box counting method.

[0093] When the process identification is "graded transportation", the system automatically triggers the direction-constrained morphological filtering. For example, directional erosion is first performed on the visible light image to eliminate horizontal interference, then homodirectional expansion is performed to connect the broken crack segments, and finally the eta value is output through skeletonization and geometric calculation. If eta>0.35 (the threshold value comes from historical experimental data), it is determined as a mechanical stress crack. When the identification is "UV disinfection", the thermal excitation unit is activated synchronously to collect thermal infrared images. The anisotropic diffusion filter is used to enhance the thermal conduction contrast of the crack area (the temperature difference between the crack area and the normal area is increased by 2-3°C), and then the connected domain is extracted to calculate the fractal dimension. If the fractal dimension is greater than 1.2, it is determined as a thermal stress crack. At the same time, during the cleaning and oiling process, the enhanced filtering is disabled, and only the basic Gaussian noise reduction is performed.

[0094] That is, after graded transportation, directional erosion can be used to eliminate horizontal noise points (such as transportation belt friction marks) in a targeted manner, while preserving the integrity of the axial cracks (the signal-to-noise ratio is improved). After UV disinfection, the thermal diffusion filter enhances the characteristics of the network cracks (the error of fractal dimension calculation is less than 5%), avoiding the masking of micro-cracks by thermal noise. At the same time, the axial elongation rate eta is corrected by the curvature model to eliminate the angle deviation caused by the projection distortion of the egg curved surface, improving the accuracy; the box counting method quantifies the complexity of the crack network, breaking through the limitations of manual experience interpretation, and improving the recognition rate of thermal stress cracks. In addition, only the key processes (graded transportation / UV disinfection) enable enhanced filtering, reducing the computational load and better meeting the real-time detection needs of the production line.

[0095] Referring to Figure 4 As shown in the figure, the process filtering module further includes:

[0096] Process logic judgment unit: real-time analysis of process identification, when the identification is graded transportation, trigger direction-constrained morphological filtering, when the identification is UV disinfection, trigger the thermal excitation enhancement unit synchronously, and start anisotropic diffusion filtering;

[0097] Filter path switching unit: disable crack-enhanced filtering during cleaning and oiling processes, and only perform basic noise reduction;

[0098] Based on the numerical range of the axial elongation rate eta and the network fractal dimension, the weight distribution factor of the subsequent multi-modal feature fusion module is dynamically adjusted, wherein the numerical range of the network fractal dimension is obtained through historical experimental data, eta>0.35 determines a mechanical stress crack, and fractal dimension>1.2 determines a thermal stress crack.

[0099] The process logic judgment unit builds a condition triggering mechanism by real-time analysis of the process identification (cleaning, staging transportation, UV disinfection, oiling) of the production line control system: when the identification is staging transportation, the directional constraint morphological filtering algorithm is called; when the identification is UV disinfection, the thermal excitation enhancement unit (pulsed laser) is activated synchronously and the anisotropic diffusion filtering algorithm is started. The filtering path switching unit dynamically configures the processing channel based on the process identification: during the cleaning / oiling process, only basic noise reduction algorithms such as Gaussian filtering are enabled, and crack enhancement filtering is disabled.

[0100] By dynamically switching filtering algorithms based on process identification, the crack characteristics caused by different stress sources (axial cracks of mechanical stress / reticular cracks of thermal stress) are targetedly enhanced, the signal-to-noise ratio of crack characteristics is improved, and during the cleaning / oiling process, the enhanced filtering is disabled and only basic noise reduction is performed to reduce invalid calculation amount; at the same time, based on the quantitative threshold of axial elongation rate and fractal dimension, the feature weight is dynamically distributed to avoid interference of non-target crack features, so that the misjudgment rate of the subsequent fusion module is reduced. At the same time, the axial elongation rate and the fractal dimension are input into the dynamic feedback module as quantitative parameters to construct a crack depth distribution model and output a risk level, realizing closed-loop control from filtering enhancement to risk decision-making, and improving the overall confidence score of the system.

[0101] Referring to Figure 5 As shown in the figure, the extraction of crack edge features and crack projection features based on polarization angle offset, the fusion of visible light texture features and near-infrared depth features, and the generation of a dual-channel crack probability map specifically include:

[0102] For visible light polarization images and near-infrared polarization images, the gradient amplitude difference of 0° and 90° polarization directions is calculated respectively to generate a polarization angle offset feature map, and the maximum value points in the feature map are extracted as crack edge features;

[0103] The vertical projection feature, geometric and optical feature along the long axis direction of the egg are projected, including the vertical maximum wave peak value, vertical projection range, circularity, light texture gradient, near-infrared penetration depth and local temperature rise;

[0104] The polarization angle offset feature and the light texture gradient in the crack projection feature are combined into a fusion feature vector, which is input into the polarization suppression channel to generate an edge probability map that suppresses the interference of curved surface reflection;

[0105] The near-infrared penetration depth and local temperature rise in the crack projection feature are linearly weighted to generate a temperature difference enhancement probability map;

[0106] Based on the weight distribution factor output by the process filtering module, the mechanical stress related features and the thermal stress related features are dynamically weighted and output to the edge probability map and the temperature difference enhancement probability map respectively, forming a double-channel crack probability map, wherein the mechanical stress related features include the vertical maximum peak value and the vertical projection range, and the thermal stress related features are the network fractal dimension and the local temperature rise.

[0107] For the visible light polarization image (490-600 nm band pass filter + 0° polarizer generated) and the near-infrared polarization image (850 nm band pass filter + 90° polarizer generated), the gradient amplitude (such as Sobel operator gradient amplitude) of 0° and 90° polarization directions is calculated respectively, so that the gradient difference map of the two polarization directions can be generated by pixel-level difference operation, and the local maximum value points in the map are extracted as crack edge feature points to form a polarization angle offset feature map. This process effectively suppresses the interference of curved surface reflection and highlights the crack edge.

[0108] Based on the egg curvature model (ellipsoidal surface equation fitting), the vertical maximum peak value of the gray projection curve peak value along the long axis direction, the vertical projection range of the curve value greater than the maximum peak value interval, the light texture gradient of the Sobel gradient mean value of the crack region of the visible light image, the near-infrared penetration depth of the pixel intensity attenuation rate of the crack region of the near-infrared image, and the local temperature rise of the temperature difference maximum value of the crack and normal region of the thermal infrared image are extracted. Through the equatorial plane normal projection direction unified coordinate, combined with the egg geometric model (short axis / long axis ratio calculation circularity), the feature quantization is realized.

[0109] Through the edge probability map channel, the polarization angle offset feature and the light texture gradient are combined into a fusion vector, which is input into the polarization suppression channel to generate an edge probability map that suppresses curved surface reflection. Through the temperature difference enhancement probability map channel, the near-infrared penetration depth and the local temperature rise are linearly weighted (the weight is calibrated by experiment), to generate a temperature difference enhancement probability map that highlights the thermal characteristics of the crack. Based on the weight distribution factor output by the process filtering module (such as when η>0.35, the mechanical stress weight is high), the mechanical stress features (vertical maximum peak value, projection range) and the thermal stress features (network fractal dimension, local temperature rise) are dynamically weighted and output to the double-channel probability map.

[0110] That is, by combining polarization angle offset features with a spectroscopic prism dual-channel imaging, mirror reflection noise can be significantly suppressed, edge detection signal-to-noise ratio is improved, and based on process identification (such as hierarchical transportation / UV disinfection) dynamic weighting mechanical stress and thermal stress features, the crack probability map is adapted to different stress sources, and the detection rate is reduced. At the same time, visible light texture gradient describes the surface topography, near-infrared penetration depth reflects the subsurface structure, and thermal infrared temperature rise reveals the difference in crack heat conduction. The fusion of the three breaks through the limitations of a single mode and solves the problem of insufficient penetration depth perception of micro-cracks. In addition, the dual-channel probability map synchronously generates edge and thermal features, replacing the traditional step-by-step detection process, so that the detection speed is improved, and the quantization error of axial extension rate η and network fractal dimension is ≤5%.

[0111] Referring to Figure 6 As shown in the figure, the vertical projection feature extraction along the long axis direction of the egg projection extraction perpendicular projection feature, geometric and optical feature specifically includes:

[0112] Based on the curvature model of the egg, all projection features are extracted along the normal projection direction of the equatorial plane;

[0113] The vertical projection range is calculated by the gray projection curve along the long axis direction of the egg, and is located in the interval inside the curve value greater than the maximum peak value;

[0114] The light texture gradient is quantified by the mean value of the Sobel operator gradient amplitude of the crack area in the visible light image;

[0115] The near-infrared penetration depth is calculated by the pixel intensity attenuation rate of the crack area in the near-infrared polarization image;

[0116] The circularity is determined based on the ratio of the short axis to the long axis of the minimum circumscribed ellipse of the crack connected domain;

[0117] The local temperature rise is extracted by the temperature difference maximum value of the crack area and the adjacent normal area in the thermal infrared image.

[0118] Based on the point cloud data of the egg surface obtained by the three-dimensional vision sensor, the ellipsoid equation (parameters are long axis a and equatorial radius b) is fitted, and the dynamic egg curvature model is established. The coordinate system is set along the normal projection direction of the equatorial plane, ensuring that all feature extraction is aligned with the geometric principal axis of the egg, and the gray value of the visible light image is projected along the long axis direction to generate a one-dimensional gray curve. The vertical projection range is defined as the continuous interval with a curve value higher than the maximum peak value P max , that is, x∈[x start , x end ], where the gray value I(x)>P max , reflecting the diffusion span of the crack in the long axis direction.

[0119] In the crack area of the visible light image, the Sobel operator is applied to calculate the pixel gradient amplitude The average value of the gradient amplitude of all pixels in the region is taken as the light texture gradient value. This value represents the sharpness of the crack edge and the surface roughness.

[0120] The pixel intensity decay rate of the crack region is measured using near-infrared polarized images (generated by 850 nm band-pass filtering and 90° polarizer) (I0 is the normal region intensity, I c is the crack region intensity), through the exponential decay model where k is the material attenuation coefficient calibration value.

[0121] The minimum circumscribed ellipse fitting is performed on the crack binary connected domain, and the ratio of the short axis r min to the long axis r max is calculated. When β approaches 0, it indicates that the crack is a narrow line, and when it approaches 1, it indicates that it is nearly circular (such as a hole defect).

[0122] In the thermal infrared image, the pixel of the crack region is taken as the region of interest (ROI), and the maximum temperature difference ΔT max between it and the adjacent normal region (5-pixel annular region outside the ROI) is calculated to exclude the interference of background temperature fluctuations.

[0123] Referring to Figure 7 , the crack depth distribution model is constructed based on the process identification, and according to the axial extension rate, the network fractal dimension, and the depth value, the micro-crack risk level is output, which specifically includes:

[0124] Based on the process identification selection input parameter, when the process identification is classified transportation, the axial extension rate and the local temperature rise value are input; when the process identification is UV disinfection, the network fractal dimension and the local temperature rise value are input;

[0125] Based on the linear weighting of the axial extension rate and the local temperature rise value, the depth estimation value during the process identification of classified transportation is calculated, and based on the network fractal dimension and the local temperature rise value, the depth estimation value of the process identification of UV disinfection is calculated, wherein the weight coefficient is calibrated through historical crack samples.

[0126] The average thickness of the eggshell is 0.3-0.4mm, the risk level is associated with the eggshell thickness, when the crack depth does not exceed half of the eggshell thickness, it is recorded as low risk, when the crack depth reaches half to equal of the eggshell thickness, it is recorded as medium risk, when the crack depth significantly exceeds the eggshell thickness, it is recorded as high risk;

[0127] Based on the depth estimation value calculated by different processes, the risk level of the current process is compared, and the micro-crack risk level is output.

[0128] The process identification interface is used to obtain the process identification of the production line (graded transportation / UV disinfection) in real time, and the corresponding parameter channel is activated dynamically: Graded transportation process: collect axial elongation η (from the directional constraint morphological filter output) ) and local temperature rise ΔT (maximum temperature difference between crack area and normal area) extracted by thermal infrared image

[0129] UV disinfection process: call the network fractal dimension (fractal dimension of connected domain calculated by box counting method) and local temperature rise ΔT. A historical crack sample database is established, and the weight coefficient is calibrated (such as k1η+k2ΔT for graded transportation, and m1D f +m2ΔT for UV disinfection, ki and mi are determined by sample regression analysis). The depth estimation value can be calculated by linear weighting: depth 运输 =ω1×η+ω2×ΔT in graded transportation, and depth 消毒 =υ1×D f +υ2×ΔT in UV disinfection. The eggshell thickness threshold (0.3-0.4mm) is associated, and when the risk is low, depth≤0.5×thickness average (i.e. <0.175mm); when the risk is medium, 0.5×thickness average<depth≤thickness average (0.175-0.35mm); when the risk is high, depth>thickness average (>0.35mm).

[0130] The evaluation model can be dynamically switched by process identification (mechanical stress focuses on axial elongation, and thermal stress focuses on network fractal dimension), so that the depth estimation error is reduced, and the absolute threshold (0.175mm / 0.35mm) is set based on the physical properties of the eggshell thickness, and the risk misjudgment rate is reduced. At the same time, the morphological features (η / D_f) and thermal features (ΔT) are fused, which solves the unstable problem of single sensor detection on the curved surface of chicken eggs, and improves the confidence of crack depth detection.

[0131] Referring to Figure 8 , the output belt process label crack positioning result and confidence score adjusts the rotation speed of the stage and the intensity of the light source, which specifically includes:

[0132] When the stage rotation makes the current detection point of the egg reach the laser incidence area of the thermal excitation enhancement unit, based on the surface normal vector output by the egg curvature model, the pulse laser emits a 500μs short-time thermal shock, which makes the transient temperature rise of the crack area, and according to the incidence angle threshold calculated by the curvature model, the universal adjusting mechanism of the LED lamp bead is adjusted in real time, so that the angle between the light source incidence direction and the surface normal vector is less than the angle threshold, and the angle threshold is obtained by optical scattering attenuation characteristic and polarization modulation cross-validation experiment;

[0133] The visible light polarization image of the first channel and the near-infrared polarization image of the second channel are synchronously collected by a light splitting prism within several seconds after the thermal shock ends, and a thermal infrared image is collected by a mid-wave infrared thermal imager;

[0134] If the risk level is low, the rotation speed of the object table is maintained at the first rotation speed, and the intensity of the annular light source is increased, the first rotation speed is calculated by matching the equatorial circumference and the resolution of the imaging system, and the light source intensity increase value is determined through compensation experiment verification of the visible light texture gradient;

[0135] If the risk level is medium, the rotation speed of the object table is reduced to the second rotation speed, and the secondary thermal excitation is triggered to enhance the crack contrast, the second rotation speed is determined by the mapping relationship between the minimum time required for complete sampling of thermal diffusion and the angular velocity of the object table.

[0136] If the risk level is high, the rotation of the object table is immediately stopped, and an alarm signal with a UV disinfection process label is output.

[0137] When the rotation of the object table causes the egg detection point to enter the laser action area, the system calculates the surface normal vector according to the egg curvature model (ellipsoidal surface equation fitting) generated by the three-dimensional vision sensor in real time. And trigger the pulsed laser to emit 500us short-time thermal shock (the time is pre-calibrated through the matching experiment of the sampling period of the infrared thermal imager and the crack temperature rise response), to excite the transient temperature rise of the crack area. Synchronously drive the gimbal adjustment mechanism of the annular light source array, according to the critical value of the incident angle calculated by the curvature model (determined by the cross verification experiment of optical scattering attenuation and polarization modulation), dynamically adjust the angle of the LED lamp bead, ensure that the angle between the incident direction of the light source and the surface normal vector is always less than the critical threshold (such as ≤15°), avoid the interference of curved surface reflection to imaging. Form a thermal excitation trigger and optical synchronous control.

[0138] Within seconds after the end of thermal shock (matching the time of thermal diffusion peak), the reflected light is separated by a light splitting prism, the first channel passes through a 490-600 nm band-pass filter and a 0° polarizer to generate a visible light polarization image; the second channel passes through an 850 nm band-pass filter and a 90° polarizer to generate a near-infrared polarization image; a mid-wave infrared thermal imager synchronously collects a thermal infrared image, forming a three-modality data synchronous acquisition mechanism. That is, at low risk (crack depth ≤ 50% of eggshell thickness), maintain the first rotation speed (calculated by matching the equatorial circumference length and imaging resolution, for example: circumference length 200 mm / 0.1 mm resolution → rotation speed 2 rpm), and increase the intensity of the annular light source (the increase value is verified by visible light texture gradient compensation experiments, such as increasing by 20% to enhance the contrast of surface texture); at medium risk (crack depth reaches 50%-100% of eggshell thickness), the stage is lowered to the second rotation speed (determined by mapping the minimum time required for complete sampling of thermal diffusion and angular velocity, for example: thermal diffusion requires 200 ms → rotation speed is reduced to 0.5 rpm), and a secondary thermal excitation is triggered to enhance the crack area temperature difference contrast; at high risk (crack depth > eggshell thickness), the stage is immediately stopped, an alarm signal with a "UV disinfection process" label is output (thermal stress cracks are prone to occur during the disinfection process), and the crack positioning result and confidence score (generated based on the dual-channel crack probability map output by the multi-modality feature fusion module) are marked.

[0139] That is, the light source incidence angle can be dynamically controlled by the curvature model (rather than fixed angle illumination), combined with the polarizer set to suppress specular reflection, to solve the uneven illumination problem of egg curved surface imaging, and to improve the crack edge signal-to-noise ratio compared with traditional annular light sources. At the same time, the rotation speed-light source intensity linkage mechanism driven by risk level can increase the speed and light source at low risk to ensure detection efficiency; reduce the speed combined with secondary thermal excitation at medium risk to ensure complete sampling of thermal diffusion signals and avoid missed detection (axial extension rate η and net fractal dimension calculation error ≤ 5%); high risk alarm directly associated with UV disinfection process to locate the defect source of production link and reduce 70% of manual recheck cost.

[0140] The egg surface micro-crack detection method based on image analysis, characterized in that it comprises:

[0141] Connect the production line control system to obtain the processing procedure identification of the eggs in real time, and the procedure identification includes the cleaning, grading transportation, UV disinfection, and oil coating procedure states;

[0142] Set a multi-band annular light source array containing independently controlled visible light and near-infrared LED lamp beads, the incidence angle of each light source is adjusted in real time based on the egg curvature model, and a polarizer set and a light splitting prism are integrated to synchronously collect visible light polarization images and near-infrared polarization images on the surface of the eggs, a pulsed laser is integrated in the center of the light source array to excite the temperature difference of the crack area through short-time thermal shock and enhance the infrared thermal imaging contrast to obtain a thermal infrared image;

[0143] Based on the process identification, the filtering path is selected, the direction constraint morphological filtering is performed on the visible light image after hierarchical transportation, the mechanical stress cracks are strengthened, and the axial extension rate of the mechanical stress microcracks is extracted, the anisotropic diffusion filtering is performed on the thermal infrared image after UV disinfection, the thermal stress cracks are enhanced, and the reticular fractal dimension of the thermal stress microcracks is calculated;

[0144] Based on the visible light image, the near-infrared image and the thermal infrared image, the crack edge feature and the crack projection feature based on the polarization angle offset are extracted, the visible light texture feature and the near-infrared depth feature are fused, the double-channel crack probability map is generated, and the crack projection feature includes the vertical maximum peak value, the vertical projection range, the circular degree, the light texture gradient, the near-infrared penetration depth and the local temperature rise.

[0145] Based on the process identification, the crack depth distribution model is constructed, the microcrack risk level is output according to the axial extension rate, the reticular fractal dimension and the depth value, the historical crack judgment threshold of the corresponding process is called, the crack positioning result with the process label and the confidence score are output, and the rotation speed of the object table and the light source intensity are adjusted.

[0146] In summary, the advantages of the present application are that the multi-modal feature fusion and dynamic parameter adjustment are driven by the process identification, the accurate discrimination of mechanical and thermal stress cracks is realized, the curved surface reflection interference is broken through, and the detection efficiency is improved.

[0147] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. An image analysis-based system for detecting microcracks on the surface of eggs, characterized in that, include: Process identification interface module: Connects to the production line control system to obtain the processing process identifier of the eggs in real time. The process identifier includes the status of cleaning, grading and transportation, UV disinfection, and oiling processes. Multi-angle ring light source module: Set up a multi-band ring light source array containing independently controlled visible light and near-infrared LED beads. The incident angle of each light source is adjusted in real time based on the curvature model of an egg. It also integrates a polarizer group and a beam splitter to simultaneously acquire visible light polarization images and near-infrared polarization images of the egg surface. Thermal excitation enhancement unit: A pulsed laser is integrated at the center of the light source array to excite the temperature difference in the crack area through short-time thermal shock, thereby enhancing the contrast of infrared thermal imaging and acquiring thermal infrared images; Process filtering module: Selects filtering path based on process identifier, performs directional constraint morphological filtering on visible light image after graded transportation to enhance mechanical stress cracks and extract axial extension rate of mechanical stress microcracks, performs anisotropic diffusion filtering on thermal infrared image after UV disinfection to enhance thermal stress cracks and calculates the network fractal dimension of thermal stress microcracks. Multimodal feature fusion module: Based on visible light images, near-infrared images and thermal infrared images, extract crack edge features and crack projection features based on polarization angle offset, fuse visible light texture features and near-infrared depth features to generate a dual-channel crack probability map. Crack projection features include vertical maximum peak value, vertical projection range, circularity, optical texture gradient, near-infrared penetration depth and local temperature rise. Dynamic feedback control module: Based on process identifiers, a crack depth distribution model is constructed. According to the axial elongation, network fractal dimension, and depth value, the microcrack risk level is output. The historical crack judgment threshold of the corresponding process is called, and the crack location result and confidence score with process label are output. The stage rotation speed and light source intensity are adjusted.

2. The image analysis-based microcrack detection system for eggs according to claim 1, characterized in that, The incident angle of each light source is adjusted in real time based on the egg curvature model, specifically including: Real-time point cloud data of egg surface is acquired by a 3D vision sensor, and an egg curvature model is established by fitting the ellipsoid equation, where the ellipsoid equation uses the egg's major axis and equatorial radius as parameters. Obtain the critical value for cross-validation of optical scattering attenuation characteristics, curved surface spot and polarization modulation. Calculate the principal curvature vector of the current detection point based on the egg curvature model. Drive the universal adjustment mechanism of each LED bead in the ring light source array so that the angle between the incident light direction and the surface normal vector is less than the critical value. The synchronously controlled beam splitter separates the reflected light into two imaging channels. The first channel generates a visible light polarized image through a 490nm-600nm bandpass filter and a 0° polarizer, while the second channel generates a near-infrared polarized image through an 850nm bandpass filter and a 90° polarizer. Within seconds after the pulsed laser emits a 500μs thermal shock, a mid-wave infrared thermal imager is used to acquire thermal infrared images. The duration of the thermal shock is determined by matching the sampling period of the infrared thermal imager with the response time of the maximum temperature rise in the crack zone.

3. The image analysis-based microcrack detection system for eggs according to claim 2, characterized in that, The selection of the filtering path based on the process identifier specifically includes: After the graded transportation process, based on the egg curvature model, linear structural elements parallel to the long axis are generated, and directional corrosion calculation is performed along the normal projection direction of the egg equatorial plane to suppress interference noise from non-mechanical stress cracks. Directional expansion calculations are employed to enhance the continuity of axial cracks through constrained morphological filtering. The filtered binary image is skeletonized, and the angle deviation θ between the crack principal axis and the long axis of the egg is calculated. Based on the angle deviation, L... e =L0×cosθ, calculate the effective extension length L e Then through Calculate the axial elongation η and output it to the dynamic feedback control module, where L0 is the original projected length and D is the equatorial diameter of the egg. After the UV disinfection process, a diffusion tensor is constructed based on the thermal gradient field to enhance the thermal conductivity along the crack normal. The thermal noise in the non-crack region is smoothed by iteratively smoothing the partial differential equation, the crack connected domain of the filtered image is extracted, and the mesh fractal dimension of the thermal stress crack is calculated by box counting.

4. The image analysis-based microcrack detection system for eggs according to claim 3, characterized in that, The process filtering module also includes: Process logic judgment unit: Real-time parsing of process identifiers. When the identifier is graded transportation, it triggers directional constraint morphological filtering. When the identifier is UV disinfection, it synchronously triggers the thermal excitation enhancement unit and starts anisotropic diffusion filtering. Filter path switching unit: Disables crack enhancement filtering during cleaning and oiling processes, and only performs basic noise reduction; Based on the numerical range of axial elongation η and network fractal dimension, the weight allocation factor of the subsequent multimodal feature fusion module is dynamically adjusted. The numerical range of network fractal dimension is obtained through historical experimental data. η>0.35 indicates mechanical stress crack, and fractal dimension>1.2 indicates thermal stress crack.

5. The image analysis-based microcrack detection system for eggs according to claim 4, characterized in that, The extraction of crack edge features and crack projection features based on polarization angle offset, and the fusion of visible light texture features and near-infrared depth features to generate a dual-channel crack probability map specifically includes: For visible light polarization images and near-infrared polarization images, the gradient magnitude difference in the polarization direction at 0° and 90° is calculated respectively, and a polarization angle offset feature map is generated. The maximum point in the feature map is extracted as the crack edge feature. Vertical projection features, geometric and optical features are extracted by projecting along the long axis of the egg, including vertical maximum peak value, vertical projection range, circularity, optical texture gradient, near-infrared penetration depth and local temperature rise; The polarization angle offset feature and the optical texture gradient in the crack projection feature are combined to form a fused feature vector, which is then input into the polarization suppression channel to generate an edge probability map that suppresses surface reflection interference. The near-infrared penetration depth and local temperature rise in the crack projection features are linearly weighted to generate a temperature difference enhancement probability map. Based on the weight allocation factor output by the process filtering module, the mechanical stress-related features and thermal stress-related features are dynamically weighted and output to the edge probability map and the temperature difference enhancement probability map, respectively, to form a dual-channel crack probability map. The mechanical stress-related features include the vertical maximum peak value and the vertical projection range, while the thermal stress-related features are the network fractal dimension and the local temperature rise.

6. The image analysis-based microcrack detection system for eggs according to claim 5, characterized in that, The extraction of vertical projection features, geometric and optical features by projecting along the long axis of the egg specifically includes: Based on the egg curvature model, all projection features are extracted along the normal projection direction of the equatorial plane. The vertical projection range is calculated using the grayscale projection curve along the long axis of the egg, and lies within the range where the curve value is greater than the maximum peak value. The optical texture gradient is quantized by the mean magnitude of the Sobel operator gradient in the crack region of the visible light image; The near-infrared penetration depth is calculated by the pixel intensity attenuation rate in the crack region of the near-infrared polarized image. Circularity is determined based on the ratio of the minor axis to the major axis of the minimum circumscribed ellipse of the cracked connected domain; Local temperature rise is extracted by the maximum temperature difference between the cracked area and the adjacent normal area in the thermal infrared image.

7. The image analysis-based microcrack detection system for eggs according to claim 6, characterized in that, The crack depth distribution model constructed based on process identifiers, which outputs the microcrack risk level according to axial elongation, network fractal dimension, and depth value, specifically includes: Input parameters are selected based on the process identifier. When the process identifier is graded transportation, the axial elongation and local temperature rise value are input; when the process identifier is UV disinfection, the network fractal dimension and local temperature rise value are input. Based on the linear weighting of axial elongation and local temperature rise, the depth estimate of the process identification during graded transportation is calculated. Based on the network fractal dimension and local temperature rise, the depth estimate of the process identification during UV disinfection is calculated. The weighting coefficients are calibrated using historical crack samples. The average thickness of eggshells was 0.3–0.4 mm. The risk level was correlated with the eggshell thickness. When the crack depth did not exceed half the eggshell thickness, it was recorded as low risk. When the crack depth reached half the eggshell thickness or equal to it, it was recorded as medium risk. When the crack depth significantly exceeded the eggshell thickness, it was recorded as high risk. Based on the depth estimates calculated for different processes, the risk level of the current process is compared, and the microcrack risk level is output.

8. The image analysis-based microcrack detection system for eggs according to claim 7, characterized in that, The output includes crack location results and confidence scores with process labels, and the adjustment of stage rotation speed and light source intensity specifically includes: When the stage rotates and the current detection point of the egg reaches the laser incident area of ​​the thermal excitation enhancement unit, the pulsed laser is triggered to emit a short thermal shock of 500μs based on the surface normal vector output by the egg curvature model, causing the temperature of the excitation crack area to rise transiently. According to the critical value of the incident angle calculated by the curvature model, the universal adjustment mechanism of the LED beads is adjusted in real time to make the angle between the incident direction of the light source and the surface normal vector less than the angle threshold. The angle threshold is obtained through cross-verification experiments of optical scattering attenuation characteristics and polarization modulation. Within seconds after the thermal shock, the visible light polarization image of the first channel and the near-infrared polarization image of the second channel are simultaneously acquired by a beam splitter, and the thermal infrared image is acquired by a mid-wave infrared thermal imager. If the risk level is low, the stage rotation speed is maintained at the first rotation speed, and the intensity of the ring light source is increased. The first rotation speed is calculated by matching the equatorial plane circumference with the imaging system resolution. The increase in light source intensity is verified and determined by the visible light texture gradient compensation experiment. If the risk level is medium, the stage rotation speed is reduced to the second rotation speed, and a secondary thermal excitation is triggered to enhance crack contrast. The second rotation speed is determined by the mapping relationship between the minimum time required for complete thermal diffusion sampling and the stage angular velocity. If the risk level is high, immediately stop the rotation of the stage and output an alarm signal with a UV disinfection process label.

9. A method for detecting microcracks on the surface of eggs based on image analysis, characterized in that, include: Connect to the production line control system to obtain the processing status of eggs in real time. The processing status includes cleaning, grading and transportation, UV disinfection, and oiling. A multi-band ring light source array containing independently controlled visible light and near-infrared LED beads is set up. The incident angle of each light source is adjusted in real time based on the curvature model of an egg. A polarizer group and a beam splitter are integrated to simultaneously acquire visible light polarization images and near-infrared polarization images of the egg surface. A pulsed laser is integrated at the center of the light source array to excite the temperature difference in the crack area through short-time thermal shock, thereby enhancing the contrast of infrared thermal imaging and acquiring thermal infrared images. Based on the process identification, the filtering path is selected. After graded transportation, the visible light image is subjected to directional constraint morphological filtering to enhance mechanical stress cracks and extract the axial extension rate of mechanical stress microcracks. After UV disinfection, the thermal infrared image is subjected to anisotropic diffusion filtering to enhance thermal stress cracks and calculate the network fractal dimension of thermal stress microcracks. Based on visible light images, near-infrared images, and thermal infrared images, crack edge features and crack projection features based on polarization angle offset are extracted. Visible light texture features and near-infrared depth features are fused to generate a dual-channel crack probability map. Crack projection features include vertical maximum peak value, vertical projection range, circularity, optical texture gradient, near-infrared penetration depth, and local temperature rise. A crack depth distribution model is constructed based on process identifiers. The microcrack risk level is output according to the axial elongation, network fractal dimension, and depth value. The historical crack judgment threshold of the corresponding process is called up, and the crack location result and confidence score with process label are output. The stage rotation speed and light source intensity are adjusted.

Images