Method and device for detecting wet fish fillets based on polarization reflection and transmission cooperative measurement

Abstract

The present application relates to the technical field of optical nondestructive testing, and discloses a wet fish piece detection method and device based on polarization reflection and transmission cooperative measurement. The method comprises the following steps: fish pieces are sequentially passed through the interval arranged reflection detection station and transmission detection station by using a conveying assembly, multi-band and multi-angle parallel / orthogonal polarization reflection images and transmission images are collected and cross-station alignment is performed; high light suppression reflection quantity is adaptively obtained according to polarization response difference, thickness proxy quantity is constructed from the transmission images, normal reflection baseline is established in combination with local fiber direction, and thickness compensation residual error is obtained; abnormal enhancement index is calculated by fusing multi-modal optical features, candidate abnormal regions are extracted, and abnormal category labels are output. The device comprises a conveying assembly, a flattening and thickness limiting assembly, an extinction detection bin, a reflection acquisition assembly, a transmission acquisition assembly and a processor. The present application is used for reducing the interference of water film highlight, thickness change and texture direction on wet fish piece abnormal detection.

Description

Technical Field

[0001] This invention relates to the technical field of optical nondestructive testing, specifically to a method and apparatus for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission. Background Technology

[0002] After being cut, skinned, deboned, washed, and drained, fish fillets typically enter the conveyor line in a wet, flexible, sheet-like form for trimming, quality inspection, grading, or packaging. In actual production, common problems with wet fish fillets include blood spots or bruises, fascia residue, skin residue, bone spurs or hard foreign objects, tissue damage, contamination, and uneven thickness. These abnormalities affect product appearance, grade, and food safety control. The surface of wet fish fillets contains a water film, mucus, and free water, easily creating specular highlights. Variations in fillet thickness, folds, curls, and local overlaps alter reflection and transmission intensity. Muscle fiber texture directionality and differences in fish species also affect image feature stability. Therefore, ordinary visible light vision or single-spectrum detection methods are prone to false positives, false negatives, and parameter drift.

[0003] For example, in the prior art, CN118592482A discloses an automated fish fillet defect trimming device and its control method, which removes fish fillet defects through image acquisition, deep learning recognition, and robotic arm trimming; CN118758892A discloses a non-destructive rapid detection method for the texture of fish meat at different freshness levels, which uses near-infrared spectroscopy and Raman spectroscopy fusion data to predict texture parameters such as elasticity, hardness, and chewiness of fish meat; CN122016713A discloses a near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon fillets, which achieves total colony count detection through near-infrared hyperspectral imaging and spectral modeling; and CN106959311A discloses a detection device and method for bone-containing fish fillets, which identifies bone-containing fish fillets through an X-ray detection device.

[0004] The aforementioned existing technologies focus on defect repair, texture or freshness evaluation, microbial index detection, and bone spur detection, respectively. However, most only address a single index, a single abnormality category, or subsequent repair stages, and have not yet formed a polarization-reflection-transmission fusion measurement scheme suitable for online multi-category abnormality detection of wet fish fillets. In practical applications, the following drawbacks exist: First, wet water films and mucus easily generate strong specular reflection, causing low-contrast abnormalities such as blood spots, fascia, and skin residue to be masked. Second, uneven thickness of the fish fillets, folds, and local overlaps can cause changes in reflectance and transmission intensity, easily confusing normal thick areas with true abnormal areas. Third, ordinary visible light or single reflectance spectrum detection is difficult to reliably identify small, slender abnormalities such as bone spurs or hard foreign objects. Fourth, if reflective illumination and transmission backlight are set in the same vertical position and operate simultaneously, crosstalk between reflective and transmission optical paths is easily generated. Fifth, in wide-span transport detection, small-sized polarization beam splitters or ordinary polarizers cannot guarantee the synchronization and consistency of the two polarization images within the complete field of view. Sixth, there is a lack of processing mechanisms to establish thickness surrogate quantities using transmission images and to perform thickness compensation for polarization reflection characteristics. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and apparatus for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission.

[0006] In a first aspect, this invention proposes a method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission, the method comprising: S1, using a conveying assembly, the wet fish fillet to be tested is sent through a reflection detection station and a transmission detection station set at intervals along the conveying direction, and reflection images and transmission images of the wet fish fillet to be tested are acquired respectively; the reflection images include parallel polarization reflection images and orthogonal polarization reflection images; S2, determine the correspondence between the wet fish fillet to be tested at the reflection detection station and the transmission detection station based on the motion state of the conveying component, and align the corresponding areas in the reflection image and the transmission image with spatial coordinates; S3. Determine the specular suppression coefficient based on the response difference between the parallel polarization reflection image and the orthogonal polarization reflection image. Then, weightedly fuse the parallel polarization reflection image and the orthogonal polarization reflection image based on the specular suppression coefficient to obtain the specular suppression reflection amount. Construct a thickness surrogate quantity to characterize transmission attenuation based on the aligned transmission image. Extract local fiber direction or texture direction features based on the specular suppression reflection amount, and combine it with the normal reflection baseline model constructed using the thickness surrogate quantity. Compare the specular suppression reflection amount with the normal reflection baseline model to obtain the thickness compensation residual. S4. The thickness compensation residual and the multimodal features extracted from the reflection and transmission images are scaled and fused to obtain the abnormal enhancement index of the wet fish fillet to be tested. The multimodal features include one or more of the following: polarization difference, multi-angle reflection difference, transmission anomaly, edge sharpness feature and orientation consistency feature. S5: Extract candidate abnormal regions and calculate region features based on the comparison results of the abnormality enhancement index and the discrimination threshold. Then, perform region type discrimination based on the region features and output the corresponding abnormality category label.

[0007] In the above technical solution, polarization reflection and transmission dual-station collaborative imaging is used, and dynamic thickness surrogate is used to compensate for the attenuation of the reflection spectrum caused by the inherent thickness of the fish slice in real time. This reduces the background interference caused by the uneven thickness of the fish slice on the reflection response, and greatly improves the contrast of the abnormal area enhancement and the signal-to-noise ratio of the initial defect screening.

[0008] Furthermore, the spatial coordinate alignment in step S2 includes: In response to the conveying components operating under constant speed conveying conditions, based on time delay Determine the time delay of transmission acquisition relative to reflection acquisition, and... The corresponding time is taken as the corresponding time of the transmission image of the corresponding region, where... The distance between the center of the reflection detection station and the transmission detection station. For conveying speed, The moment the reflected image was acquired; Using pulse equivalent q The encoder determines the number of trigger delay pulses based on the center distance between the reflection detection station and the transmission detection station: ,in, For rounding operations; in response to the non-constant conveyor speed, when the cumulative pulse value of the encoder reaches... At that time, the corresponding transmission image rows are acquired or retrieved to determine the longitudinal correspondence of the corresponding regions in the direction of travel. The encoder accumulates pulse values ​​when acquiring reflected images; Based on time delay Or trigger delay pulse number Determine the initial correspondence between the corresponding regions using the reflection-transmission coordinate transformation matrix. The initial spatial coordinate mapping is completed, and the result is corrected based on the outer contour, centroid, leading edge point, or boundary features of the wet fish fillet to be tested, resulting in spatial coordinate alignment between the corresponding regions of the reflected and transmitted images; whereby the coordinate mapping is expressed as: ; In the formula, These are the pixel coordinates in the reflected image. These correspond to the pixel coordinates in the transmitted image.

[0009] In the above technical solution, by delaying the time and finely aligning the encoder pulses for both constant and non-constant speed operating conditions, and in conjunction with the coordinate transformation matrix, the spatial registration accuracy of the corresponding areas of the reflected and transmitted images is improved, and the registration errors and false edge effects caused by conveying speed fluctuations, slippage, and attitude changes are reduced.

[0010] Furthermore, the specular suppression reflection amount is constructed by fusing the following methods: Calculate specular indication based on normalized parallel polarization reflectance images and normalized orthogonal polarization reflectance images: ; In the formula, This is the normalized parallel polarization reflection image at the k-th band and the m-th illumination angle; This is the corresponding normalized orthogonal polarization reflectance image; This is the balance coefficient for the two polarization channels; The value is set to a positive number to avoid the denominator being zero. This means that only the positive values ​​within the parentheses should be retained. Determine the highlight suppression coefficient based on the highlight indication: ; In the formula, The positive threshold for highlight suppression is determined based on the statistical distribution of highlight indicators in areas without obvious specular highlights in normal fish fillets. A numeric truncation function that restricts input values ​​to the range of 0 to 1; By using the specular suppression coefficient, the parallel polarization reflection image and the orthogonal polarization reflection image are fused at the pixel level to construct the first... The first band, the first Specular reflection suppression at each illumination angle: .

[0011] In the above technical solution, by introducing a dynamic specular indication based on the enhancement degree of two polarization signals and a pixel-level suppression coefficient, adaptive differentiation and precise suppression of specular highlights on wet surface water films and mucus surfaces are achieved. In areas without obvious specular highlights, the background information of the flesh with parallel polarization is fully preserved, and in areas with strong specular highlights, effective diffuse reflection signals with orthogonal polarization are automatically supplemented, which greatly improves the physical extraction contrast and image signal-to-noise ratio of surface and subsurface defects under strong reflective interference.

[0012] Furthermore, in step S3, the thickness proxy is constructed as follows: ; ; In the formula, This is a normalized transmission image; This is the transmittance after numerical truncation; The default value is a positive number, used to avoid zero input in logarithmic operations; The normal reflection baseline model is characterized by the following formula: ; In the formula, For the predicted reflection baseline model, Mapping function for local fiber direction or texture direction features Obtained through training based on corresponding region samples from normal fish fillets; In the above technical solution, by constructing a thickness surrogate quantity and a reflection baseline model trained on normal samples, it is possible to estimate the reflection baseline of normal fish fillets under corresponding thickness and texture direction conditions, so that the residuals mainly characterize the abnormal response that deviates from the normal reflection mode, which greatly enhances the ability to highlight small and hidden defects.

[0013] Furthermore, methods for extracting local fiber orientation or texture orientation features include: The texture direction input image is constructed based on the amount of specular suppression reflection under preset bands and preset illumination angles, or by fusing the specular suppression reflection amounts under multiple bands and illumination angles. Calculate the local gradients of the input image in different directions according to the texture direction, and use a Gaussian smoothing kernel to convolve and smooth the gradient energy to construct a structure tensor. The local principal gradient direction angle is determined based on the gradient energy component and cross-correlation component in the structural tensor. The local principal gradient direction angle is then rotated by π / 2 and normalized according to the π period to obtain the local fiber direction or texture direction features.

[0014] In the above technical solution, by constructing a structural tensor from the reflective image after high light suppression and extracting the directional features of local muscle fibers, the interference of water film reflection on the extraction of local texture direction is reduced, thereby improving the adaptability of the normal reflective baseline model to the reflection differences of different local texture directions and reducing the risk of abnormal misjudgment caused by the natural changes in the direction of muscle fibers.

[0015] Furthermore, the thickness compensation residual is expressed as: ; in, To compensate for the thickness residual; This corresponds to the amount of specular reflection suppression; This corresponds to the normal reflection baseline model; The polarization difference is expressed as: ; In the formula, This is a measure of polarization difference; Preset to positive numbers; This is a normalized parallel polarization reflection image; This is a normalized orthogonal polarization reflectance image; The difference in reflection from multiple angles is expressed as follows: ; In the formula, This represents the difference in reflection from multiple angles. and The first The amount of highlight suppression reflection in each band under different illumination angles; Preset to positive numbers; The thickness compensation residual and polarization difference were averaged and aggregated for each detection band and each illumination angle to obtain the comprehensive thickness compensation residual and comprehensive polarization difference; the multi-angle reflection difference was averaged and aggregated for each detection band to obtain the comprehensive multi-angle reflection difference. Transmission anomalies are constructed as follows: ; ; ; In the formula, The local normal thickness background is calculated by taking the median value of the thickness proxy set of all pixels within a local sliding window. The current position of the pixel to be measured Thickness at the agent level; To measure the position of the pixel to be measured (x,y) A preset local sliding window centered on the target area; H(u,v) The position of each neighboring pixel within the local window (u, v) The thickness of the agent at the point satisfies , This is an abnormal amount of transmission. For standard defect-free fish fillets, the thickness within the normal area is the agent quantity. With corresponding local normal thickness background The standard deviation of the residuals between the differences; This indicates the standard deviation calculated within the normal range of standard defect-free fish fillets; Preset to positive numbers; Edge sharpness features G(x,y) Calculated as follows: The gradient magnitudes are calculated for both the comprehensive thickness compensation residual and the transmission anomaly, thus obtaining the sharpness of the reflection edge. G R (x,y) and the sharpness of the transmission edge G T (x,y) ; ; ; In the formula, To compensate for residual thickness; G R (x,y) For the sharpness of the reflection edge; G T (x,y) For the sharpness of the transmission edge; x , y These are the pixel coordinates of the image; Based on the comparison between the transmission anomaly and the preset edge gating threshold, an adaptive gating selection is performed between the reflection edge sharpness and the transmission edge sharpness: when the transmission anomaly at the pixel location is greater than or equal to the preset edge gating threshold, the transmission edge sharpness is used as the edge sharpness feature of the pixel location; otherwise, the reflection edge sharpness is used as the edge sharpness feature. Directional consistency feature Extract using the following method: Calculate the local gradient for the specular suppression reflection amount or thickness compensation residual, and construct the corresponding structure tensor. : ; In the formula, and These represent the input specular suppression reflection amount or thickness compensation residual. direction and Gradient of direction, For Gaussian smoothing kernel, This represents the convolution operation; Let the structure tensor The two eigenvalues ​​are and ,and Then the directional consistency feature The calculation is as follows: ; In the formula, It is a preset positive number.

[0016] The scale normalization is performed on the comprehensive thickness compensation residual, comprehensive polarization difference, comprehensive multi-angle reflection difference, transmission anomaly, edge sharpness feature, and direction inconsistency value obtained from the direction consistency feature transformation to obtain the corresponding normalized feature values. The anomaly enhancement index is then calculated according to the following formula: ; ; In the formula, This is an abnormal enhancement index; This is a normalized value for the absolute value of the comprehensive thickness compensation residual; This is a normalized quantity for the comprehensive polarization difference. This is a normalized quantity that integrates the differences in reflection from multiple angles; This is the normalized value of the transmission anomaly. This is a normalization factor for the edge sharpness feature; For quantities with inconsistent directions, the normalized quantity is used. to The weights are non-negative and the sum of the weights is 1. They can be determined through training with calibration samples or pre-set according to the detection task. Norm [·] indicates that the input features are dimensionless according to the statistical range of normal fish fillet samples, so that different physical quantities are mapped to a uniform scale.

[0017] In the above technical solution, multimodal features are deeply characterized from multiple physical dimensions such as polarization decoupling, multi-angle scattering, transmission attenuation, boundary abrupt changes and tissue fiber arrangement orientation, providing a multi-dimensional decoupling feature space for distinguishing defects with highly intertwined optical responses such as bone spurs, fascia, skin residues, and blood spots.

[0018] Furthermore, the process of extracting candidate anomaly regions and calculating region features includes: Based on the comparison results between the anomaly enhancement index and the preset discrimination threshold, threshold segmentation and connected component analysis are performed to extract candidate anomaly region masks; Based on camera spatial calibration parameters or planar calibration matrix, the total number of pixels contained in the candidate anomalous region mask is converted into the physical region area, and the aspect ratio is calculated based on the minimum bounding rotation rectangle of the candidate anomalous region mask. Within the range defined by the candidate anomalous region mask, the pixel mean values ​​within the region are calculated for the corresponding thickness proxy, transmission anomalous value, comprehensive polarization difference, comprehensive multi-angle reflection difference, orientation consistency feature, and edge sharpness feature, respectively, to obtain the corresponding average thickness proxy, average transmission anomalous value, average polarization difference, average multi-angle reflection difference, average orientation consistency feature, and average edge sharpness feature. Within the range defined by the candidate anomalous region mask, the pixel mean values ​​within the region are calculated for the thickness compensation residuals corresponding to each band and each illumination angle, to obtain the average reflection residual values ​​corresponding to each band and each illumination angle.

[0019] In the above technical solution, by constructing a candidate abnormal region mask based on the anomaly enhancement index and extracting the physical region area and the region features within the corresponding spatial range, the misjudgment caused by single pixel noise is reduced, and the correspondence between the candidate region and the physical size and the features of each region is established.

[0020] Furthermore, based on regional features, the system identifies the region type and outputs the corresponding anomaly category label, specifically including: If the area of ​​the candidate abnormal region is within the preset area range, the mean value of the reflection residual corresponding to the candidate abnormal region in the 660±10nm band exceeds the normal reflection residual range corresponding to that band, and the aspect ratio of the candidate abnormal region is less than or equal to the preset aspect ratio threshold, then the abnormal category label is determined as a blood spot or ecchymosis type region. If the area of ​​the candidate abnormal region is within the preset area range, the mean values ​​of the reflection residuals of the candidate abnormal region in the 450±10nm and 530±10nm bands both exceed the normal reflection residual range corresponding to the corresponding bands, the average polarization difference of the candidate abnormal region is greater than or equal to the polarization difference threshold, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, the average directional consistency feature is less than or equal to the directional consistency threshold, and the aspect ratio of the candidate abnormal region is greater than the preset aspect ratio threshold, then the abnormal category label is determined as a fascial residual region. If the mean reflection residual values ​​of the candidate abnormal region in the 450±10nm and 530±10nm bands both exceed the normal reflection residual range corresponding to the corresponding bands, the area of ​​the candidate abnormal region is greater than the upper limit of the preset area range, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, and the average directional consistency feature is less than or equal to the directional consistency threshold, then the abnormal category label is determined as a skin residue region. If the area of ​​the candidate abnormal region is within the preset area range, the average transmission abnormality of the candidate abnormal region is greater than or equal to the first judgment threshold, the average edge sharpness feature is greater than or equal to the second judgment threshold, and the aspect ratio of the candidate abnormal region is greater than the preset aspect ratio threshold, then the abnormal category label is determined as a bone spur or hard foreign body region. If the area of ​​the candidate abnormal region is within the preset area range, the average edge sharpness feature of the candidate abnormal region is greater than or equal to the second judgment threshold, the average directional consistency feature is less than or equal to the directional consistency threshold, and the average reflection residual value corresponding to at least two of the four reflection bands of 450±10nm, 530±10nm, 660±10nm and 810±15nm exceeds the normal reflection residual range corresponding to the corresponding band, but the candidate abnormal region does not meet the judgment conditions for blood spot or ecchymosis type regions, and its average transmission abnormality is less than the first judgment threshold, then the abnormal category label is determined as a tissue damage type region. If the average reflection residual value of the candidate abnormal region in at least one of the 450±10nm, 530±10nm and 810±15nm bands exceeds the normal reflection residual range corresponding to the corresponding band, the average polarization difference of the candidate abnormal region is greater than or equal to the polarization difference threshold, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, the average transmission anomaly is less than the first judgment threshold, the area of ​​the candidate abnormal region is within the preset area range, and the aspect ratio of the candidate abnormal region is less than or equal to the preset aspect ratio threshold, then the abnormal category label is determined as a contamination attachment type area. If the average thickness proxy of the candidate abnormal region exceeds the normal thickness proxy range, the average transmission abnormality is greater than or equal to the first judgment threshold, and the mean reflection residual values ​​corresponding to each band of 450±10nm, 530±10nm, 660±10nm and 810±15nm do not exceed the normal reflection residual range corresponding to the corresponding band, and the candidate abnormal region does not meet the judgment conditions of blood spot or ecchymosis area, fascia residue area, skin residue area, bone spur or hard foreign body area, tissue damage area and contamination attachment area, then the abnormal category label is determined as uneven thickness, folded edge or local overlap area. Among them, the normal reflection residual range, polarization difference threshold, multi-angle reflection difference threshold, direction consistency threshold, first judgment threshold, second judgment threshold, normal thickness proxy range, preset area range, and preset aspect ratio threshold corresponding to each band are determined based on the statistical distribution of normal fish fillet samples and corresponding abnormal calibration samples.

[0021] In the above technical solution, by using the cross-logic judgment of the regional features and morphological features such as area and aspect ratio, high-precision grading and semantic classification of blood spots, fascia, skin residue, bone spurs, damage, contamination and overlapping areas are achieved, avoiding the misjudgment of harmless folds and overlaps as malignant foreign body defects that need to be removed.

[0022] Furthermore, step S2 also includes normalizing and correcting the acquired reflectance and transmission images separately. The normalization correction includes: Reflection image normalization: for the first The first band, the first Original reflection image at each illumination angle The image is processed to obtain a normalized reflectance image. : ; In the formula, This is a dark field image; Use a whiteboard or a reference area image within the same frame as a reference. Preset positive numbers; normalized reflectance image Including normalized parallel polarization reflection images and normalized orthogonal polarization reflection image ; Transmission image normalization: normalization of the original transmission image The image is processed to obtain a normalized transmission image. : ; In the formula, This is a dark-field transmission image; This is a transmission reference image; It is a preset positive number.

[0023] In the above technical solution, by using dark field subtraction combined with reference domain division correction, the systematic errors caused by factors such as camera dark response, light source intensity drift and uneven spatial distribution of illumination are reduced. The absolute light intensity of the image is standardized and mapped to a uniform and drift-free relative reflectivity and relative transmittance, which ensures the high robustness of subsequent multimodal feature extraction and defect judgment.

[0024] Secondly, this invention proposes a wet fish fillet detection device based on the combined measurement of polarization reflection and transmission, the device comprising: The conveying assembly is used to carry and transport the wet fish fillets to be tested through the reflection detection station and the transmission detection station; The flattening and thickness limiting component is set in front of or at the reflection detection station and is used to guide, flatten, hold and limit the thickness of the wet fish fillet to be tested. The testing chamber surrounds the reflection testing station and the transmission testing station; The reflection acquisition component is set at the reflection detection station and is used to acquire the reflection image information of the wet fish fillet to be tested; The transmission acquisition component is set at the transmission detection station and is spaced apart from the reflection acquisition component. It is used to acquire transmission images of the wet fish fillets to be tested. The processor, electrically connected to the reflection acquisition component and the transmission acquisition component, is used to receive the reflection image and the transmission image, and to execute the corresponding steps of the wet fish fillet detection method based on the combined measurement of polarization reflection and transmission described in the first aspect.

[0025] In the above technical solution, by integrating "mechanical flattening, optical extinction, dual-station spatial isolation and on-demand staggered acquisition, and multimodal processing", a detection pipeline with high physical constraints and resistance to external interference is formed, which suppresses the physical interference of edge curling, reflection and external stray light from the source.

[0026] Furthermore, the reflection acquisition component includes: An illumination source is positioned above the conveying assembly to emit multi-band detection light from different incident directions into the detection field of view of the reflection detection station to illuminate the wet fish fillets to be tested. The linear polarization component includes a first linear polarization component and a second linear polarization component respectively disposed on the output light paths of illumination source one and illumination source two, for forming linearly polarized detection light with a consistent preset incident polarization direction; the polarization beam splitter is used to separate the reflected light from the detection field of view according to two polarization directions parallel and orthogonal to the preset incident polarization direction, to obtain parallel polarized reflection signal and orthogonal polarized reflection signal respectively; Parallel acquisition camera and orthogonal acquisition camera are respectively set on the first and second outgoing optical paths of polarization beam splitter, and are used to acquire parallel polarization reflection images and orthogonal polarization reflection images under the same detection field of view. The illumination source includes illumination source one and illumination source two. The incident angle of illumination source one is set to 20° to 35°; the incident angle of illumination source two is set to 45° to 65°; and the illumination width of illumination source one and illumination source two is not less than the detection field of view width.

[0027] In the above technical solution, the dual-angle, wide-width multi-band illumination and dual-path polarization physical beam splitting acquisition work together to suppress specular highlight interference and improve the coverage integrity of shallow surface anomalies within a wide field of view.

[0028] Furthermore, the transmission acquisition component includes: A transmission light source is positioned above the conveying assembly to emit transmission detection light that penetrates the wet fish fillet to be tested at the transmission detection station. A light-transmitting bearing area is set on the conveying component; A transmission camera, positioned below the conveying assembly, is used to acquire transmission images of the wet fish fillet being tested from below the conveying assembly. The conveying component is equipped with an encoder; the transmission light source is a large-area diffuse narrowband backlight with an emission wavelength of 850±15nm.

[0029] In the above technical solution, the transmission attenuation information inside the fish fillet is obtained through a transmissive diffuse infrared optical path, and the transport displacement state is obtained by using a displacement encoder. This provides motion information for cross-station correspondence and subsequent registration of reflected and transmitted images, thereby improving the stability of transmission imaging and cross-station data correspondence.

[0030] Furthermore, the conveying speed of the conveying component is set to 0.05m / s to 0.60m / s, and the conveying component is equipped with an encoder with a pulse equivalent of 0.05mm / pulse to 1.00mm / pulse; the frame rate of the parallel acquisition camera, the orthogonal acquisition camera, and the transmission camera are all set to 30fps to 300fps, and the single frame exposure time is set to 0.1ms to 10ms.

[0031] The aforementioned conveyor speed, camera frame rate, exposure time, and spatial resolution are matched according to the actual detection conditions and are not arbitrarily combined within their respective ranges. Let the conveyor speed be... The single-frame exposure time is t e The actual size corresponding to a single pixel in the conveying direction is ρ y Under continuous lighting conditions, the following is preferably satisfied: vt e ≤βρ y In the formula, β To allow for motion ambiguity coefficients, it is preferable that the coefficient is no greater than 1; ρ y Determined by camera calibration results. When using pulsed illumination, the effective illumination pulse width is used. t p Replace single frame exposure time t e Appropriate constraints are imposed. By matching the parameters described above, the movement and displacement of the fish fillet along the transport direction during a single effective exposure are limited, reducing the impact of motion blur on the reflected and transmitted images.

[0032] In the above technical solution, by coordinating and matching the conveying speed, displacement detection resolution, camera frame rate and effective exposure time, motion blur caused by rapid conveying is reduced, and stable image and motion information are provided for dynamic registration between the reflection detection station and the transmission detection station.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention sets up reflection and transmission detection stations at intervals along the conveying direction, and combines light-shielding isolation and on-demand staggered acquisition to reduce optical crosstalk between the two stations, thereby improving the acquisition stability of reflection and transmission images. Combined with the internal extinction design of the detection chamber, stray reflections within the system are significantly reduced, ensuring clean backgrounds and high contrast during imaging at each station. Furthermore, the physical constraint of the fish fillet posture by the flattening and thickness-limiting component, along with coordinated matching of conveying speed, camera frame rate, and effective exposure time, reduces motion blur under high-speed assembly line conditions, improving the original imaging quality of reflection and transmission images.

[0034] 2. This invention introduces a long strip-shaped polarization beam splitter extending along the width of the detection field of view. Combined with a multi-band light source with dual illumination angles and a large illumination width, it enables the synchronous and non-destructive acquisition of parallel polarized reflection images and orthogonal polarized reflection images within the same wide transmission field of view. This significantly improves the accuracy of field-of-view overlap and measurement consistency during wide-range multi-channel detection. By fusing parallel and orthogonal polarization components, it not only suppresses specular highlight interference from water film and mucus on the surface of wet fish fillets but also effectively excites and enhances the characteristic contrast of superficial tissue abnormalities such as blood spots, fascia, skin residue, and contaminant adhesion under specific spectra, thereby improving the effective detection coverage of the wide field of view.

[0035] 3. This invention establishes cross-station correspondence based on the motion state of the conveying component, conveying speed, or encoder displacement pulse information. It also combines the reflection-transmission coordinate transformation matrix and local image feature correction to improve the spatial registration accuracy of the corresponding areas of the reflection image and the transmission image, reduce station registration errors caused by conveying speed fluctuations, conveying slippage, attitude deflection, and local deformation, thereby reducing the impact of registration errors on subsequent anomaly detection.

[0036] 4. This invention utilizes transmission images to construct a thickness proxy quantity, and uses this thickness distribution as a benchmark to adaptively compensate for the polarization multi-band reflection characteristics, reducing the coupling effect between the natural thickness variation of the fish fillet and the actual abnormal reflection response. By performing multimodal fusion of the compensated residual with polarization difference, multi-angle reflection difference, transmission anomaly, edge sharpness, and directional inconsistency quantity converted from directional consistency characteristics, and combining it with regional statistical features for category determination, this invention reduces the risk of missing important defects while avoiding misjudgment and rejection of harmless process conditions. Attached Figure Description

[0037] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Many anticipated advantages of the embodiments and other embodiments of the invention will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0038] Figure 1 This is a flowchart of a wet fish fillet detection method according to an embodiment of the present invention; Figure 2 This is a flowchart of reflection image normalization and highlight suppression according to an embodiment of the present invention; Figure 3 This is a flowchart of transmission image normalization thickness compensation according to an embodiment of the present invention; Figure 4This is a flowchart of anomaly enhancement and classification discrimination according to an embodiment of the present invention; Figure 5 This is a front view of a wet fish fillet detection device according to an embodiment of the present invention; Figure 6 This is a schematic diagram of a wet fish fillet detection device according to an embodiment of the present invention; Figure 7 This is a cross-sectional structural diagram of a reflection detection station and a transmission detection station according to an embodiment of the present invention; Figure 8 This is a schematic diagram of a reflection detection station structure according to an embodiment of the present invention; Figure 9 This is a schematic diagram of a transmission detection station structure according to an embodiment of the present invention; Figure 10 This is a flowchart of a measurement method for a wet fish fillet detection device according to an embodiment of the present invention.

[0039] The meanings of the numbers in the diagram are as follows: 100-Conveying assembly, 200-Flattening and thickness limiting assembly, 300-Detection chamber, 400-Reflection acquisition assembly, 500-Transmission acquisition assembly, 600-Processor, 700-Wet fish fillet to be tested, 410-Illumination source one, 420-Illumination source two, 430-Polarization beam splitter, 440-Parallel acquisition camera, 450-Orthogonal acquisition camera, 411-First linear polarization assembly, 421-Second linear polarization assembly, 510-Transmission light source, 520-Transmission camera. Detailed Implementation

[0040] In the following detailed description, reference is made to the accompanying drawings, which form part of the detailed description and are illustrated by specific illustrative embodiments in which the invention may be practiced. In this regard, directional terms such as “top,” “bottom,” “left,” “right,” “up,” “down,” etc., are used with reference to the orientation of the described figures. Because components of the embodiments can be positioned in several different orientations, directional terms are used for illustrative purposes and are by no means limiting. It should be understood that other embodiments may be utilized or logical changes may be made without departing from the scope of the invention. Therefore, the following detailed description should not be taken in a limiting sense, and the scope of the invention is defined by the appended claims.

[0041] This invention proposes a method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission. For example... Figure 1 As shown, the method includes: S1, using a conveying assembly, the wet fish fillet to be tested is sent through a reflection detection station and a transmission detection station set at intervals along the conveying direction, and reflection images and transmission images of the wet fish fillet to be tested are acquired respectively; the reflection images include parallel polarization reflection images and orthogonal polarization reflection images; In some specific embodiments, a conveying assembly carries and transports the wet fish fillet to be tested, allowing it to pass sequentially through a reflection detection station and a transmission detection station. During the reflection acquisition stage, the wet fish fillet enters the reflection detection field of view in a flattened state. A multi-angle, multi-band reflection illumination source emits detection light into the reflection detection field of view. The reflected light from the reflection detection field of view is separated into parallel polarized reflection signals and orthogonal polarized reflection signals by a long strip polarization beam splitter, and parallel polarized reflection images and orthogonal polarized reflection images are output respectively. Subsequently, the wet fish fillet continues to move along the conveying direction with the conveying assembly to the downstream transmission detection station. During the transmission acquisition stage, the transmitted light passes through the wet fish fillet and the light-transmitting carrying area, acquiring a transmission image.

[0042] S2, determine the correspondence between the wet fish fillet to be tested at the reflection detection station and the transmission detection station based on the motion state of the conveying component, and align the corresponding areas in the reflection image and the transmission image with spatial coordinates; In some specific embodiments, spatial coordinate alignment includes, during image registration, setting the center distance between the reflection detection station and the transmission detection station to be L, and determining the positional correspondence between the corresponding areas of the wet fish fillet to be tested between the front and rear stations based on the movement state of the conveying assembly. Under constant speed conveying conditions, based on the time delay... Determine the time delay of transmission acquisition relative to reflection acquisition, where For conveying speed, and will The corresponding time is used as the corresponding time of the transmission image of the corresponding region, where The moment of image acquisition is indicated by the pulse equivalent. q The encoder determines the number of trigger delay pulses based on the center distance between the reflection detection station and the transmission detection station: ;in, This represents a rounding operation; for production lines where the conveyor speed fluctuates, the rounding operation is performed when the encoder's accumulated pulses reach a certain value. At that time, the corresponding transmission image is acquired or retrieved to determine the longitudinal correspondence of the corresponding area in the direction of movement, wherein... The encoder accumulates pulse values ​​when acquiring reflected images.

[0043] Specifically, the conveying speed of the conveying component is set to 0.05m / s to 0.60m / s, and the encoder configured in the conveying component has an encoder pulse equivalent of 0.05mm / pulse to 1.00mm / pulse.

[0044] Furthermore, to improve registration accuracy, during the calibration stage, a coordinate mapping relationship between the coordinates of the reflected image and the coordinates of the transmitted image is established using a calibration plate with grid points or markers. The coordinate mapping is expressed as: ; In the formula, These are the pixel coordinates in the reflected image. These correspond to the pixel coordinates in the transmitted image. This is the reflection-transmission coordinate transformation matrix obtained from calibration.

[0045] In actual testing, first, based on the time delay... or encoder pulse information Determine the initial longitudinal correspondence of the corresponding regions, and then use the reflection-transmission coordinate transformation matrix. The initial spatial coordinate mapping is completed; finally, the outer contour, centroid, leading edge point or boundary features of the wet fish fillet to be tested are used for secondary fine-tuning to correct the registration error caused by conveying slippage, attitude deflection and local deformation, so as to obtain the spatial coordinate alignment result of the corresponding area of ​​the reflected image and the transmitted image.

[0046] S3. Determine the specular suppression coefficient based on the response difference between the parallel polarization reflection image and the orthogonal polarization reflection image. Then, weightedly fuse the parallel polarization reflection image and the orthogonal polarization reflection image based on the specular suppression coefficient to obtain the specular suppression reflection amount. Construct a thickness surrogate quantity to characterize transmission attenuation based on the aligned transmission image. Extract local fiber direction or texture direction features based on the specular suppression reflection amount, and combine it with the normal reflection baseline model constructed using the thickness surrogate quantity. Compare the specular suppression reflection amount with the normal reflection baseline model to obtain the thickness compensation residual. In some specific embodiments, reference is made in the reflection image processing flow. Figure 2 , Figure 2 A flowchart illustrating reflection image normalization and specular suppression according to an embodiment of the present invention is shown. As shown, firstly, the original parallel polarized reflection images and the original orthogonal polarized reflection images at each band and illumination angle are input, and the corresponding reflection reference image and dark field image are acquired simultaneously. Subsequently, the reflection image is normalized and specular suppression is performed on the first... The first band, the first Original reflection image at each illumination angle After normalization, the normalized reflectance image is obtained: ,in, This is a dark field image; Use a whiteboard or a reference area image within the same frame as a reference. The preset positive number is used; through the above calculations, normalized parallel polarization reflection images are output respectively. and normalized orthogonal polarization reflection image .

[0047] Subsequently, the normalized parallel polarization reflection image was further analyzed. and normalized orthogonal polarization reflection image To achieve fusion, construct a specular suppression reflection level. Polarization difference and the difference in reflection from multiple angles .

[0048] Among them, the amount of specular reflection suppression , indicating the first The detection band, the first The effective reflection intensity of the fish fillet obtained after reducing the specular highlights of the water film and mucus on the surface of the wet fillet under a certain illumination angle can be used as the reflection input for subsequent thickness compensation residual and abnormal enhancement index calculation.

[0049] Because the specular reflection at the water film interface on the surface of wet fish fillets has strong polarization-preserving characteristics, the response of the parallel polarization reflection channel is significantly enhanced in the highlight region, while the diffuse reflection of the fish tissue retains some tissue information in the orthogonal polarization channel. After achieving response balance between the two polarization channels, the highlight indicator is used... The formula for characterizing the degree of local specular reflection enhancement of the parallel polarization channel relative to the orthogonal polarization channel is as follows: ; In the formula, For normalized parallel polarization reflectance images, This is a normalized orthogonal polarization reflectance image; It is the balance coefficient for the two polarization channels, used to correct differences in camera response, optical transmittance, and polarization splitting efficiency; Preset to positive numbers; This means that only the positive values ​​within the parentheses will be retained.

[0050] Specifically, when acquiring parallel polarization reflection images and orthogonal polarization reflection images, the corresponding detection light is made to form linearly polarized detection light with a preset incident polarization direction; the parallel polarization reflection image represents the reflection component whose polarization direction is parallel to the preset incident polarization direction; the orthogonal polarization reflection image represents the reflection component whose polarization direction is orthogonal to the preset incident polarization direction. Further, based on the specular indication... Determine the pixel-level specular suppression coefficient : ; In the formula, The positive threshold for highlight suppression is determined based on the statistical distribution of highlight indicators in areas without obvious specular highlights in normal fish fillets. This is a numeric truncation function that restricts input values ​​to the range of 0 to 1. T p The specific value is determined based on the statistical distribution of the highlight index in areas of normal fish fillets without obvious specular highlights, and the specific value is determined based on the calibration results of the actual acquisition system and the statistical results of normal fish fillet samples.

[0051] Based on this, pixel-level fusion is performed on the normalized parallel polarization reflectance image and the normalized orthogonal polarization reflectance image to construct the specular suppression reflectance. : ; When there is no obvious highlight at the pixel location, the highlight suppression coefficient Approaching 0, highlight suppression of reflection. Primarily retains parallel polarization reflection information; when there is significant water film highlight at the pixel location, the highlight suppression coefficient... Increase the amount of specular reflection suppression By gradually increasing the proportion of orthogonally polarized reflection information after channel balancing, the effective reflection characteristics of fish tissue are preserved while suppressing specular highlights. It should be noted that the amount of highlight suppression reflection... The polarization fusion results correspond only to the same band and the same illumination angle. The response differences between different illumination angles are characterized by subsequent multi-angle reflection differences.

[0052] Among them, polarization difference quantity Represented as: ,in, The preset positive number is used. By calculating and fusing the difference between parallel polarization and orthogonal polarization, the changes in the surface tissue structure of the wet fish fillet under test are highlighted, which, together with the multi-angle reflection difference, serve as a multimodal feature source for subsequent calculation of the anomaly enhancement index.

[0053] Among them, the difference in reflection from multiple angles This is used to characterize the difference in reflection response of the same pixel region under illumination at different incident angles. Let the... Under various wavelengths, with an incident angle of 20° to 35°, preferably 25° ± 5°, the corresponding specular suppression reflection amount is: For illumination with an incident angle of 45° to 65°, preferably 55° ± 5°, the corresponding specular reflection suppression is: Then construct: ; In the formula, and The first The amount of highlight suppression reflection in each band under different illumination angles; The value is assumed to be positive. Fascia residue, skin residue, contaminant adhesion, and uneven surface areas are typically more sensitive to lighting angles, thus their corresponding differences in multi-angle reflectance are greater. Larger.

[0054] In some specific embodiments, reference is made in the transmission image processing flow. Figure 3 , Figure 3A flowchart of transmission image normalization thickness compensation according to an embodiment of the present invention is shown. As shown in the figure, the original transmission images for each band are first input. Transmission reference diagram and transmission dark field diagram The image is then normalized and corrected to obtain a normalized transmission image. ,in, The preset value is a positive number. Then, further calculations are performed according to the formula. Construct a surrogate quantity for the thickness of the wet fish fillet to be tested, wherein, In the formula, This is a normalized transmission image; This is the transmittance after numerical truncation; The value is a preset positive number. The thickness surrogate is used to characterize the transmission attenuation changes caused by the geometric thickness of the fish fillet, as well as tissue absorption, scattering, and water content. It is not equivalent to the absolute measurement of the geometric thickness. When the local thickness of the wet fish fillet increases, or when folds or local overlaps occur, or when tissue density increases, the transmission intensity decreases, and the thickness surrogate increases accordingly; conversely, the thickness surrogate is smaller in thin areas. Thus, the thickness surrogate can accurately describe and quantify the contribution of different thickness regions, folds, curls, and local overlaps to changes in the optical signal, realizing a surrogate mechanism that characterizes thickness by transmission.

[0055] Furthermore, the processing procedure establishes a normal reflectance baseline model based on samples from normal fish fillet areas: ; In the formula, The predicted normal reflection baseline model; For thickness of the agent volume, Mapping function for local fiber direction or texture direction features Linear models, piecewise linear models, polynomial models, lookup table models, or calibrated regression models are used to predict the theoretical reflectance benchmarks that should exist at corresponding thickness surrogate amounts and texture direction features in a defect-free background. Based on this, the measured first... The first band, the first Highlight suppression reflection at each illumination angle Point-to-point subtraction of the corresponding normal reflection baseline model The thickness compensation residual under the corresponding working condition is calculated and obtained: ; Through the energy subtraction operation after spatial alignment, the normal reflective brightness changes and pseudo-abnormal interference caused by the thickness fluctuations, non-uniform folding and overlapping of the wet fish fillet under test are effectively weakened and eliminated, making the background flatter and thus significantly highlighting the real surface and subsurface tissue abnormalities such as blood spots, fascia, skin residue and contamination attached to the background.

[0056] Specifically, local fiber orientation or texture orientation features Indicates pixel position The local fiber orientation angle or principal texture orientation angle is used to characterize the main extension direction of local muscle fibers in the wet fish fillet under test. Texture orientation is constructed from the input image using specular suppression of reflectance. This reduces the interference of water film, mucus, and specular highlights on the surface of wet fish fillets on the determination of muscle fiber orientation. Texture orientation input image. Preset bands can be selected and preset lighting angle Under the specular suppression of reflection It can also be a combination of multiple bands. and multiple lighting angle sets The specular suppression reflection amount is obtained by fusing: ; In the formula, This indicates the number of specular suppression reflection images involved in the fusion.

[0057] Furthermore, based on the texture direction of the input image Calculate the local gradients in the horizontal and vertical directions of the image respectively: ; ; In the formula, and These represent the texture directions of the input image at... direction and Local gradient in the direction. Within a local window centered on the scale parameter, Gaussian smoothing kernel The gradient energy is smoothed by convolution to obtain the structure tensor. The three components: ; ; ; In the formula, , and Let these represent the gradient energy term and the cross-correlation term in the structure tensor, respectively. This represents the convolution operation. The corresponding structure tensor is represented as: ; Finally, through the four-quadrant arctangent function The local fiber orientation angle or texture principal orientation angle is calculated as follows: ; ; In the formula, The local principal gradient direction angle is the local fiber direction angle or texture principal direction angle obtained by rotating the local principal gradient direction angle by π / 2 (i.e., local fiber direction or texture direction feature). For local fiber direction or texture direction features, the value range is set to or .

[0058] In normal fish meat areas, the local fiber orientation angle changes continuously along the muscle fiber direction. However, fascia remnants, skin remnants, tissue damage, contaminant adhesion, or local folded edges can cause abrupt changes or deflections in the local fiber orientation angle. Local fiber orientation or texture orientation characteristics serve as a baseline model for normal reflection. The input variables enable the normal reflection baseline model to simultaneously consider the influence of fish fillet thickness and local texture direction on reflection intensity, thereby reducing the probability of being misjudged as an abnormal area due to changes in normal texture direction.

[0059] S4. The thickness compensation residual and the multimodal features extracted from the reflection and transmission images are scaled and fused to obtain the abnormal enhancement index of the wet fish fillet to be tested. The multimodal features include one or more of the following: polarization difference, multi-angle reflection difference, transmission anomaly, edge sharpness feature and orientation consistency feature.

[0060] refer to Figure 4 , Figure 4 A flowchart illustrating anomaly enhancement and classification based on an embodiment of the present invention is shown. As shown in the figure, the thickness compensation residual is first input. Polarization difference Multi-angle reflection difference Abnormal transmission amount Inconsistent direction quantity 1- and edge sharpness features .

[0061] Among them, transmission anomaly This is used to characterize abnormal light obstruction in the interior or shallow layers of a fish fillet. A localized normal thickness background is established based on the thickness proxy and the normal fish fillet area. Then, the transmission anomaly is calculated: ; ; ; In the formula, The local normal thickness background is calculated by taking the median value of the thickness proxy set of all pixels within a local sliding window. The current position of the pixel to be measured Thickness at the agent level; To measure the position of the pixel to be measured (x,y) A preset local sliding window centered on the target area; H(u,v) The position of each neighboring pixel within the local window (u,v) The thickness of the agent at the point satisfies , This is an abnormal amount of transmission. For standard defect-free fish fillets, the thickness within the normal area is the agent quantity. With corresponding local normal thickness background The standard deviation of the residuals between the differences; This indicates the standard deviation calculated within the normal range of standard defect-free fish fillets; The preset value is positive. Bone spurs, hard foreign bodies, local overlap, and areas of abrupt changes in thickness often cause abnormal transmission. Significantly increased.

[0062] Specifically, localized normal thickness background This method provides the local low-frequency baseline thickness of the fish fillet under test, assuming it is free of foreign objects and damage. It is obtained by median filtering of the thickness surrogate within a local window centered at pixel position (x, y). Within the normal area of ​​a standard, defect-free fish fillet, the thickness surrogate of each pixel is statistically analyzed. H(x,y) With corresponding local normal thickness background The difference residuals between them are used to determine the statistical standard deviation of the difference residuals. .

[0063] Edge sharpness is used to characterize the degree of local abrupt changes at the boundaries of candidate anomaly regions. Gradient magnitudes are calculated for both the thickness compensation residual and the transmission anomaly, yielding the reflection edge sharpness. G R (x,y) and the sharpness of the transmission edge G T (x,y) Among them, the sharpness of the reflective edge G R (x,y) It mainly characterizes the boundary abrupt changes in the reflectance response of abnormal areas such as blood spots, fascia, skin residue, tissue damage, and surface contamination; and the sharpness of the transmission edge. GT (x,y) It mainly characterizes the boundary abrupt changes in transmitted light caused by bone spurs, hard foreign bodies, and local overlap. Among them, ; In the formula, To reflect edge sharpness, To compensate for the residual thickness in The first partial derivative in the direction (horizontal direction); To compensate for the residual thickness in The first partial derivative in the direction (perpendicular direction); ; In the formula, For the sharpness of the transmission edge, For transmission anomaly in The first partial derivative in the direction (horizontal direction); For transmission anomaly in The first partial derivative in the direction (perpendicular direction); To adaptively determine the source of edge features when the anomaly category is unknown, an edge gating threshold independent of the first decision threshold used for region classification is set. T G When pixel position (x,y) Transmission anomaly T a (x,y) Greater than or equal to T G Select the sharpness of the transmission edge. G T (x,y) Use the edge sharpness feature at that location; otherwise, select the reflective edge sharpness. G R (x, y) This serves as the edge sharpness feature at that location. This yields a unified edge sharpness feature. G(x,y) This is used for anomaly enhancement index and subsequent regional classification.

[0064] Directional consistency feature This is used to describe the continuity of the texture direction of fish fillet muscle fibers. A texture feature image is constructed based on specular suppression of reflectance, and the gradient and corresponding structure tensor are calculated within a preset local window. ; In the formula, and These represent the input specular suppression reflection amount or thickness compensation residual. direction and First-order partial derivative in the direction, For Gaussian smoothing kernel, Represents the convolution operation. Based on the structure tensor. The main direction is used to obtain the local image gradient direction, and this local image gradient direction is rotated. and according to Periodic normalization yields local fiber orientation or texture orientation features. Local fiber orientation or texture orientation characteristics As input to the normal reflection baseline model, it is used to compensate for the changes in reflectance brightness caused by different muscle fiber orientations in normal fish fillets.

[0065] Let the structure tensor The two eigenvalues ​​are and ,and Then the direction consistency feature is calculated as follows: ; In the formula, Preset positive number. Directional consistency characteristic. Used to describe the continuity and concentration of the muscle fiber texture in fish fillets. Normal fish muscle fiber regions typically exhibit high directional consistency, meaning they have a large directional consistency characteristic; however, tissue damage, contamination, fascia interlacing, or abnormal texture areas can disrupt local continuity, leading to a decrease in directional consistency characteristics.

[0066] Thickness compensation residuals obtained under multiple detection bands and multiple illumination angles and polarization difference The mean values ​​were converged separately to obtain the comprehensive thickness compensation residual. and comprehensive polarization difference For the multi-angle reflection differences obtained under each detection band By performing mean convergence, the comprehensive multi-angle reflection difference is obtained. It is calculated as follows: ; ; ; In the formula, K To detect the number of bands, M This refers to the number of lighting angles.

[0067] Before calculating the anomaly enhancement index, the thickness compensation residual, polarization difference, multi-angle reflection difference, transmission anomaly, edge sharpness feature, and orientation inconsistency are scaled and normalized respectively, and the normalized feature quantities are then weighted and fused; among them, the orientation inconsistency is determined by the orientation consistency feature. The result of the conversion is defined as follows: .

[0068] ; ; In the formula, This is an abnormal enhancement index; This is a normalized value for the absolute value of the comprehensive thickness compensation residual; This is a normalized quantity for the comprehensive polarization difference. This is a normalized quantity that integrates the differences in reflection from multiple angles; This is the normalized value of the transmission anomaly. This is a normalization factor for the edge sharpness feature; For quantities with inconsistent directions, the normalized quantity is used. to The weights are non-negative and can be determined through training with calibrated samples or preset according to the detection task.

[0069] In some specific embodiments, for thickness compensation residuals, polarization differences, and multi-angle reflection differences with multiple detection bands or multiple illumination angles, it is preferable to obtain the comprehensive thickness compensation residuals, comprehensive polarization differences, and comprehensive multi-angle reflection differences through mean convergence calculation, and then perform scale normalization to participate in the calculation of the anomaly enhancement index. Optionally, each channel can also obtain the corresponding comprehensive quantity using maximum absolute value projection or preset weight fusion. Transmission anomalies, edge sharpness features, and directional inconsistencies are directly participated in the anomaly enhancement index calculation after scale normalization. At the same time, the original feature quantities and candidate region statistics corresponding to each band and each illumination angle are retained in the category discrimination branch for subsequent anomaly category determination.

[0070] S5: Extract candidate abnormal regions and calculate region features based on the comparison results of the abnormality enhancement index and the discrimination threshold. Then, perform region type discrimination based on the region features and output the corresponding abnormality category label.

[0071] In some specific embodiments, reference continues to be made to Figure 4 Based on the comparison between the anomaly enhancement index and the discrimination threshold, candidate anomaly regions of the wet fish fillets to be tested are extracted using the discrimination threshold and connected component analysis is performed to form candidate anomaly regions. When the anomaly enhancement index of a pixel... If the value is greater than or equal to the discrimination threshold, it is determined as a candidate point and connected component analysis is performed to generate a candidate abnormal region mask, thereby determining the corresponding candidate abnormal region; otherwise, it is determined as normal background or non-candidate region.

[0072] Specifically, the discrimination threshold Satisfy the following formula: ,in, The mean value of the abnormal enhancement index in the normal fish fillet area; The standard deviation of the abnormal enhancement index in the normal fish fillet area; Here is the confidence coefficient, and The value range is 1.5 to 4.0, preferably 2.0 to 3.0; or, the discrimination threshold. The 90th to 99.5th percentile values ​​of the abnormal enhancement index in the normal fish fillet area were taken.

[0073] Finally, the extracted candidate abnormal regions are classified into different regions and the corresponding abnormal category labels are output.

[0074] Specifically, to achieve region-level classification, spatial calibration and region feature calculation are performed on each extracted candidate anomaly region. Regarding physical size calibration, a candidate anomaly region corresponds to a set of candidate anomaly pixel coordinates within a candidate anomaly region mask. The corresponding physical region area is obtained by multiplying the total number of pixels contained in the set by the actual horizontal and vertical physical dimensions of the pixels, or by mapping based on a planar calibration matrix. The aspect ratio of the candidate anomaly region is calculated based on the aspect ratio of the minimum bounding rectangle of the set. In terms of feature statistics, a parallel decoupling mechanism is adopted between the candidate extraction branch and the category discrimination branch. Specifically, in the candidate extraction branch, the anomaly enhancement index map composed of comprehensive feature quantities is used to extract the candidate anomaly region mask. In the category discrimination branch, within the range defined by the candidate anomaly region mask, the pixel mean values ​​within the region are calculated for thickness proxy quantity, transmission anomaly quantity, comprehensive polarization difference quantity, comprehensive multi-angle reflection difference quantity, orientation consistency feature, and edge sharpness feature, respectively, to obtain the corresponding average thickness proxy quantity, average transmission anomaly quantity, average polarization difference quantity, average multi-angle reflection difference quantity, average orientation consistency feature, and average edge sharpness feature. At the same time, within the range defined by the candidate anomaly region mask, the pixel mean values ​​within the region are calculated for the thickness compensation residual corresponding to each band and each illumination angle, to obtain the average reflection residual value corresponding to each band and each illumination angle, which is used for subsequent anomaly category determination by combining the physical region area and aspect ratio.

[0075] Furthermore, based on regional features, the region type is determined and the corresponding anomaly category label is output, specifically including: If the area of ​​the candidate abnormal region is within the preset area range, the mean value of the reflection residual corresponding to the candidate abnormal region in the 660±10nm band exceeds the normal reflection residual range corresponding to that band, and the aspect ratio of the candidate abnormal region is less than or equal to the preset aspect ratio threshold, then the abnormal category label is determined as a blood spot or ecchymosis type region. If the area of ​​the candidate abnormal region is within the preset area range, the mean values ​​of the reflection residuals of the candidate abnormal region in the 450±10nm and 530±10nm bands both exceed the normal reflection residual range corresponding to the corresponding bands, the average polarization difference of the candidate abnormal region is greater than or equal to the polarization difference threshold, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, the average directional consistency feature is less than or equal to the directional consistency threshold, and the aspect ratio of the candidate abnormal region is greater than the preset aspect ratio threshold, then the abnormal category label is determined as a fascial residual region. If the mean reflection residual values ​​of the candidate abnormal region in the 450±10nm and 530±10nm bands both exceed the normal reflection residual range corresponding to the corresponding bands, the area of ​​the candidate abnormal region is greater than the upper limit of the preset area range, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, and the average directional consistency feature is less than or equal to the directional consistency threshold, then the abnormal category label is determined as a skin residue region. If the area of ​​the candidate abnormal region is within the preset area range, the average transmission abnormality of the candidate abnormal region is greater than or equal to the first judgment threshold, the average edge sharpness feature is greater than or equal to the second judgment threshold, and the aspect ratio of the candidate abnormal region is greater than the preset aspect ratio threshold, then the abnormal category label is determined as a bone spur or hard foreign body region. If the area of ​​the candidate abnormal region is within the preset area range, the average edge sharpness feature of the candidate abnormal region is greater than or equal to the second judgment threshold, the average directional consistency feature is less than or equal to the directional consistency threshold, and the average reflection residual value corresponding to at least two of the four reflection bands of 450±10nm, 530±10nm, 660±10nm and 810±15nm exceeds the normal reflection residual range corresponding to the corresponding band, but the candidate abnormal region does not meet the judgment conditions for blood spot or ecchymosis type regions, and its average transmission abnormality is less than the first judgment threshold, then the abnormal category label is determined as a tissue damage type region. If the average reflection residual value of the candidate abnormal region in at least one of the 450±10nm, 530±10nm and 810±15nm bands exceeds the normal reflection residual range corresponding to the corresponding band, the average polarization difference of the candidate abnormal region is greater than or equal to the polarization difference threshold, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, the average transmission anomaly is less than the first judgment threshold, the area of ​​the candidate abnormal region is within the preset area range, and the aspect ratio of the candidate abnormal region is less than or equal to the preset aspect ratio threshold, then the abnormal category label is determined as a contamination attachment type area. If the average thickness proxy of the candidate abnormal region exceeds the normal thickness proxy range, the average transmission abnormality is greater than or equal to the first judgment threshold, and the mean reflection residual values ​​corresponding to each band of 450±10nm, 530±10nm, 660±10nm and 810±15nm do not exceed the normal reflection residual range corresponding to the corresponding band, and the candidate abnormal region does not meet the judgment conditions of blood spot or ecchymosis area, fascia residue area, skin residue area, bone spur or hard foreign body area, tissue damage area and contamination attachment area, then the abnormal category label is determined as uneven thickness, folded edge or local overlap area. The normal reflection residual interval, polarization difference threshold, multi-angle reflection difference threshold, directional consistency threshold, first judgment threshold, second judgment threshold, edge gating threshold, normal thickness surrogate interval, and preset aspect ratio threshold for each band are determined based on the statistical distribution of normal fish fillet samples on their respective characteristic quantities. The normal reflection residual interval and normal thickness surrogate interval are primarily determined based on the mean, standard deviation, or quantile of the normal fish fillet calibration samples. The contamination adhesion area interval and the judgment threshold related to specific anomaly categories can be determined by combining the characteristic distribution of the corresponding anomaly calibration samples.

[0076] Specifically, the normal ranges and judgment thresholds for each characteristic quantity are determined based on the mean, standard deviation, or quantile values ​​established from normal fish fillet calibration samples, where: The upper limit of the normal reflection residual range and the normal thickness proxy range is set as the normal sample mean plus 1.5 to 4.0 times the standard deviation (preferably 2.0 to 3.0 times), or as the 90% to 99.5% quantile of the normal sample distribution; the lower limit is set as the normal sample mean minus the same multiple of the standard deviation, or as the 0.5% to 10% quantile of the normal sample distribution. The polarization difference threshold, the multi-angle reflection difference threshold, and the first judgment threshold are set to the mean of the corresponding normal sample plus 1.5 to 4.0 times the standard deviation (preferably 2.0 to 3.0 times), or set to the 90% to 99.5% quantile of the normal sample distribution; The directional consistency threshold is set as the mean of the corresponding normal samples minus 1.5 to 4.0 times the standard deviation, or as the 0.5% to 10% quantile. The edge gating threshold is determined based on the statistical distribution of transmission anomalies in normal fish fillet areas and calibration samples with obvious transmission blocking characteristics. This threshold is used to determine the selection criteria for reflective and transmissive edge sharpness. The second judgment threshold is based on the uniform edge sharpness characteristics obtained after gating. G(x,y) The statistical distribution in the normal fish fillet area can be determined by setting it to the mean of normal edge sharpness plus 1.5 to 3.5 times the standard deviation, or by setting it to the preset high quantile value corresponding to the normal sample distribution.

[0077] The physical region area is expressed in mm² after being calibrated by camera pixels to physical size. Set the noise removal area threshold. A n When the physical region area of ​​the candidate connected region is smaller than A n When this occurs, it is treated as an isolated noise region and removed; the preset area range is denoted as [ A min , A max In one embodiment, the area threshold is set from 3 mm² to 200 mm². Each area threshold is determined based on the camera's spatial resolution, the actual size of the contaminant deposits, and statistical results from calibration samples.

[0078] When determining the type of region, the exclusionary rule is applied in the following order: bone spurs or hard foreign bodies, blood spots or ecchymosis, fascia residue, skin residue, tissue damage, contamination, and uneven thickness or overlapping edges. When a candidate abnormal region meets multiple judgment rules at the same time, the confidence level is determined by calculating the distance between the candidate abnormal region and each category in the standardized feature space, and the abnormal category label corresponding to the highest confidence level is output.

[0079] To reduce system computational overhead and improve compatibility for engineering implementation, this invention is directly compatible with existing mature classifiers at the classification decision layer. That is, classification can be achieved using rule-based thresholding, decision trees, support vector machines, random forests, or lightweight convolutional neural network classifiers. Those skilled in the art should understand that regardless of the specific classifier used, the features of the classification input are derived from the uniquely extracted transmission thickness surrogate quantity, thickness compensation residual, polarization difference quantity, multi-angle reflection difference quantity, transmission anomaly quantity, orientation consistency feature, and edge sharpness feature of this invention. The final output includes the location of the abnormal region, the anomaly category label, the boundary contour, and confidence information.

[0080] Continue to refer to Figures 5-7 As a implementation of the above method, in a second aspect, this application provides a wet fish fillet detection device based on the combined measurement of polarization reflection and transmission. Figures 5-7The figures show a front view, a structural schematic diagram, and cross-sectional structural schematic diagrams of the reflection detection station and the transmission detection station according to an embodiment of the present invention. As shown in the figures, the wet fish fillet detection device of this embodiment includes a conveying assembly 100, a flattening and thickness limiting assembly 200, a detection chamber 300, a reflection acquisition assembly 400, a transmission acquisition assembly 500, and a processor 600. The conveying assembly 100 carries and conveys the wet fish fillet 700 to be tested sequentially through the reflection detection station and the transmission detection station. The flattening and thickness limiting assembly 200 is disposed in front of or at the reflection detection station and is used to guide, flatten, hold, and limit the thickness of the wet fish fillet 700 to be tested, thereby reducing the impact of edge curling, folding, local bulging, and posture fluctuations of the wet fish fillet 700 during the conveying process on the stability of optical measurement. The detection chamber 300 surrounds the reflection detection station and the transmission detection station. The inner wall of the detection chamber 300 is made of matte material to block ambient light and reduce stray reflections inside the detection chamber 300. The detection chamber 300 can be replaced by a light shield, matte curtain, or local darkroom control structure. The reflection detection station and the transmission detection station are arranged at intervals along the conveying direction of the conveying assembly 100. The center distance between the reflection detection station and the transmission detection station is 80mm to 300mm, preferably 150mm to 250mm. This center distance is used to avoid posture fluctuations of the wet fish fillet 700 to be tested between the two stations, while facilitating the arrangement of the reflection acquisition assembly 400 and the transmission acquisition assembly 500.

[0081] Specifically, the flattening and thickness limiting component 200 can adopt physical structures that can constrain the flatness and thickness limits of the fish fillet, such as mechanical pressure rollers, flattening flexible scrapers, negative pressure adsorption platforms, or non-contact high-pressure air curtain blowing components.

[0082] Specifically, to reduce optical crosstalk, a light-shielding structure is installed between the reflection detection station and the transmission detection station. The light-shielding structure includes a light-shielding partition, an anti-glare isolation area, or a narrow slit transition channel. Both the reflection acquisition component 400 and the transmission acquisition component 500 are electrically connected to the processor 600, which receives the acquired images and performs regional feature calculations and anomaly category determination.

[0083] In some specific embodiments, reference continues to be made to Figure 8 , Figure 8 A schematic diagram of the reflective detection station structure of a wet fish fillet detection device according to an embodiment of the present invention is shown. Figure 7 and Figure 8The reflection acquisition component 400 in this embodiment specifically includes an illumination source 410, an illumination source 420, a polarizing beam splitter 430, a parallel acquisition camera 440, an orthogonal acquisition camera 450, a first linear polarization component 411, and a second linear polarization component 421. The illumination source 410 and the second illumination source 420 are positioned above the transport component 100 and are used to emit multi-band detection light from different incident directions into the detection field of view of the reflection detection station to illuminate the wet fish fillet 700 to be tested. The first linear polarization component 411 is disposed in the output light path of the illumination source 410, and the second linear polarization component 421 is disposed in the output light path of the illumination source 420. These components convert the multi-band detection light emitted by the corresponding illumination source into linearly polarized detection light with a preset incident polarization direction. The transmission axes of the first linear polarization component 411 and the second linear polarization component 421 are parallel to each other, ensuring that the detection light at the two illumination angles has a consistent polarization reference direction.

[0084] Preferably, illumination source 410 and illumination source 420 are disposed in the same incident plane, and the transmission axes of the first linear polarization component 411 and the second linear polarization component 421 are both perpendicular to the common incident plane to form s-polarized incident detection light. Specifically, the multi-band detection light preferably includes 450±10nm, 530±10nm, 660±10nm, and 810±15nm. Among them, 450±10nm and 530±10nm are used to enhance the color and scattering differences between light-colored fish meat tissue, fascia residue, skin residue, and surface contaminant attachments; 660±10nm is used to enhance the absorption differences related to blood spots or ecchymosis; 810±15nm is used to reduce some visible texture interference and supplement the characterization of tissue density, water content, and near-infrared reflection differences between some contaminant attachments and the fish meat matrix. Each band does not correspond one-to-one with a single abnormality category, but is used in conjunction with polarization differences, multi-angle reflection differences, transmission anomalies, and morphological features for defect classification. 850±15nm transmitted light was used to construct a thickness surrogate and to characterize the abnormal obstruction of transmitted light by bone spurs, hard foreign objects, and local overlap.

[0085] Specifically, the incident angle α of illumination source one 410 is set to 20° to 35°, preferably 25°±5°, to provide relatively stable surface reflection information; the incident angle b of illumination source two 420 is set to 45° to 65°, preferably 55°±5°, to enhance the response differences of water film highlights, surface undulations, fascia residue, and skin residue to the illumination angle. Furthermore, the illumination width of both illumination source one 410 and illumination source two 420 is not less than the detection field of view width, preferably 1.05 to 1.30 times the detection field of view width.

[0086] Specifically, the polarization beam splitter 430 is installed in the reflection acquisition channel of the reflection detection station to receive reflected light from the detection field of view and separate it according to two polarization directions: parallel and orthogonal to the preset incident polarization direction. The reflection component parallel to the preset incident polarization direction is defined as a parallel polarization reflection signal, and the reflection component orthogonal to the preset incident polarization direction is defined as an orthogonal polarization reflection signal. The two polarization separation directions of the polarization beam splitter 430 correspond to the preset incident polarization direction and its orthogonal direction during installation and calibration. A parallel acquisition camera 440 and an orthogonal acquisition camera 450 are respectively installed on the first and second output optical paths of the polarization beam splitter 430 to acquire parallel polarization reflection images and orthogonal polarization reflection images under the same detection field of view. The polarization beam splitter 430 can be a PBS polarization beam-splitting cubic prism, a strip-shaped polarization beam splitter group, a polarization beam-splitting film combination, or other optical beam-splitting structures capable of separating parallel and orthogonal polarization. In a small field-of-view implementation, a PBS polarization beam-splitting cubic prism can be used. In the wide-field transmission detection implementation, the polarization beam splitter 430 preferably adopts a long strip-shaped polarization beam splitter group extending along the width of the detection field of view. Its effective length is not less than the width of the detection field of view, and preferably 1.05 to 1.30 times the width of the detection field of view. For example, when the width of the detection field of view is 500 mm, the effective length of the polarization beam splitter 430 can be set to 520 mm to 650 mm, the effective width can be set to 50 mm to 80 mm, and the thickness can be set to 40 mm to 70 mm. At the same time, parallel polarization and orthogonal polarization can be acquired synchronously by dual cameras, or by a single camera with time-division switching of the optical path, or by directly using a polarization array camera (such as a polarization mosaic camera).

[0087] Furthermore, the polarization beam splitter 430 may include a low-iron optical glass substrate and a polarization beam splitting film layer tilted at 45°. The polarization beam splitting film layer is used to transmit one polarization component to the parallel acquisition camera 440 and reflect the other polarization component to the orthogonal acquisition camera 450. To adapt to the food processing environment, a sealed protective window, a removable waterproof cover, or a purge air path may be provided on the outside of the polarization beam splitter 430 to reduce the impact of water vapor, fog, and oil contamination on imaging stability. Specifically, the tilt angle of the polarization beam splitting film layer of the polarization beam splitter 430 relative to the incident light path is 45°±3°, and the polarization extinction ratio is not less than 100:1, preferably not less than 500:1. The field-of-view overlap error of the two polarized images acquired by the parallel acquisition camera 440 and the orthogonal acquisition camera 450 after calibration is not greater than 2 pixels, preferably not greater than 1 pixel. By limiting the above parameters, the field-of-view consistency, edge region brightness stability, and polarization separation reliability of the two polarized images in wide-span transmission scenarios can be improved.

[0088] In some specific embodiments, reference continues to be made to Figure 9 , Figure 9A schematic diagram of the transmission detection station structure of a wet fish fillet detection device according to an embodiment of the present invention is shown. Figure 7 and Figure 9 It can be seen that the transmission acquisition component 500 is located downstream of the reflection acquisition component 400. The transmission acquisition component 500 specifically includes a transmission light source 510, a transmission camera 520, and a light-transmitting support area disposed on the conveying component 100. The transmission light source 510 is disposed above the conveying component 100 and is used to emit transmission detection light that penetrates the wet fish fillet 700 to be tested towards the transmission detection station. The transmission light source 510 is preferably a large-area diffuse narrow-band backlight, and its emission wavelength is preferably 850±15nm. Optionally, the transmission wavelength can be adjusted in the visible to near-infrared range according to the fish species and defect type. The light-transmitting support area can be a light-transmitting food-grade conveyor belt, a transparent support window, a narrow-slit light-transmitting window, or a partially light-transmitting support plate to reduce the interference of the surface texture of the conveying component 100 on the transmission image. The transmission camera 520 is disposed below the conveying component 100 and is used to acquire transmission images penetrating the wet fish fillet 700 to be tested from below the conveying component 100. To reduce optical crosstalk between reflection and transmission measurements, the reflection detection station and the transmission detection station are spatially isolated in this embodiment. When the crosstalk between the stations exceeds a preset threshold, the reflection illumination and transmission illumination are controlled to be triggered by short pulses in an alternating manner.

[0089] Specifically, the conveying speed of the conveying component 100 is set to 0.05 m / s to 0.60 m / s, preferably 0.10 m / s to 0.30 m / s; the encoder pulse equivalent of the conveying component 100 is 0.05 mm / pulse to 1.00 mm / pulse; the thickness range of the wet fish fillet 700 to be tested after being flattened by the flattening and thickness limiting component 200 is 2 mm to 35 mm, preferably 5 mm to 25 mm; the frame rate of the parallel acquisition camera 440, the orthogonal acquisition camera 450 and the transmission camera 520 are all set to 30 fps to 300 fps, preferably 60 fps to 120 fps, and the single frame exposure time is set to 0.1 ms to 10 ms, preferably 0.5 ms to 5 ms, and the horizontal spatial resolution of the image is 0.05 mm / pixel to 0.50 mm / pixel; the above parameters are used to reduce image blurring caused by conveying motion through collaborative matching, and to provide stable imaging conditions and motion state information for cross-station registration of reflected and transmitted images.

[0090] In actual testing, based on the conveying speed v Single frame exposure time t e Actual size corresponding to a single pixel in the transport direction ρ y Matching motion imaging parameters; under continuous illumination conditions, preferably satisfying: vte ≤βρ y ,in, β To allow for motion blur coefficients, when using pulsed illumination, the effective illumination pulse width should be used. t p Appropriate constraints are applied. The encoder pulse equivalent is denoted as... q The encoder pulse information is mainly used to determine the initial longitudinal correspondence between the reflection detection station and the transmission detection station. Further registration and correction are then performed by combining the reflection-transmission coordinate transformation matrix and local image features of the fish fillet. In some specific embodiments, the wet fish fillet detection device also includes a reference correction component (not shown in the figure). The reference correction component is used to correct for light source intensity drift, camera dark current, residual stray light in the detection chamber, and background differences in the conveyor belt of the conveyor assembly at the reflection and transmission detection stations. During the reflection acquisition stage, the reference correction component synchronously or periodically acquires reflection reference images and reflection dark field images at various wavelengths and illumination angles, and the processor calculates a normalized reflection image. During the transmission acquisition stage, the reference correction component acquires transmission reference images and transmission dark field images, and the processor calculates a normalized transmission image. Specifically, the reference correction component includes a reflection reference area, a transmission reference area, a dark field acquisition unit, and a correction storage unit. The reflection reference area is set at the edge of the detection field of view of the reflection detection station or at a fixed position in the detection chamber. The surface of the reflection reference area is a diffuse reflection white board, a grayscale reference board, or a food-grade diffuse reflection standard sheet. The transmission reference area is set at the edge of the light-transmitting bearing area or an empty area of ​​the transmission detection station. It is used to collect a transmission reference image when there is no wet fish fillet to be tested obstructing the view. The dark field acquisition unit obtains the dark field image by means of a light shield, an electronic shutter, or by turning off the light source. The calibration storage unit is used to store the acquired reference image and dark field image.

[0091] refer to Figure 10 , Figure 10 A flowchart of a measurement method using a wet fish fillet detection device according to an embodiment of the present invention is shown. As shown in the figure, the detection method using the wet fish fillet detection device of this application specifically includes: Step 101: Use the conveying component to send the wet fish fillet to be tested to the reflection detection station, and let it enter the reflection detection field of the reflection detection station in a flattened state under the action of the flattening and thickness limiting component. Step 102: Turn on multi-angle multi-band reflection illumination at the reflection detection station, and receive the reflected light from the detection field of view through a polarization beam splitter. Use a parallel acquisition camera and an orthogonal acquisition camera to acquire parallel polarized reflection images and orthogonal polarized reflection images under the same detection field of view. Step 103: The wet fish fillet to be tested is moved to the transmission detection station by the conveying assembly; Step 104: Turn on the transmission light source at the transmission detection station; when the optical crosstalk between the reflection station and the transmission station exceeds the preset threshold, control the reflection illumination and transmission illumination to use short pulse interleaved triggering, and acquire the transmission image. Step 105: Based on the encoder information, transport displacement information, or timestamp information of the transport component, perform pixel-level spatial coordinate alignment and registration of the corresponding areas of the wet fish fillet to be tested in the reflected image and the transmitted image. Step 106: Normalize and correct the registered reflection and transmission images respectively to eliminate system background noise and system errors caused by uneven spatial distribution of illumination. Step 107: Construct the thickness proxy of the wet fish fillet to be tested based on the normalized transmission image, extract local fiber orientation features based on the high light suppression reflection amount, establish a normal reflection baseline and calculate the thickness compensation residual. Step 108: The thickness compensation residual, polarization difference, multi-angle reflection difference, transmission anomaly, edge sharpness feature and orientation inconsistency are scaled and normalized respectively. The normalized feature values ​​are then weighted and fused to calculate the anomaly enhancement index. Candidate anomaly regions are extracted based on the comparison between the anomaly enhancement index and the discrimination threshold.

[0092] Step 109: Determine the region type of the candidate abnormal regions and output the corresponding abnormal category label or hierarchical auxiliary information.

[0093] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission, characterized in that, The method includes: S1, the wet fish fillet to be tested is sent through a reflection detection station and a transmission detection station set at intervals along the conveying direction using a conveying assembly, and reflection images and transmission images of the wet fish fillet to be tested are acquired respectively; the reflection images include parallel polarization reflection images and orthogonal polarization reflection images; S2, determine the correspondence between the wet fish fillet to be tested at the reflection detection station and the transmission detection station according to the motion state of the conveying component, and align the corresponding areas in the reflection image and the transmission image with spatial coordinates; S3. Determine the specular suppression coefficient based on the response difference between the parallel polarization reflection image and the orthogonal polarization reflection image. Then, weight and fuse the parallel polarization reflection image and the orthogonal polarization reflection image based on the specular suppression coefficient to obtain the specular suppression reflection amount. Construct a thickness surrogate amount to characterize transmission attenuation based on the aligned transmission image. Extract local fiber direction or texture direction features based on the specular suppression reflection amount, and combine it with the normal reflection baseline model constructed using the thickness surrogate amount. Compare the specular suppression reflection amount with the normal reflection baseline model to obtain the thickness compensation residual. S4, scale normalize and fuse the thickness compensation residual and the multimodal features extracted from the reflection image and the transmission image to obtain the abnormal enhancement index of the wet fish fillet to be tested; the multimodal features include one or more of the following: polarization difference, multi-angle reflection difference, transmission anomaly, edge sharpness feature and orientation consistency feature. S5. Based on the comparison result between the anomaly enhancement index and the discrimination threshold, candidate anomaly regions are extracted and region features are calculated. Based on the region features, region type discrimination is performed, and the corresponding anomaly category label is output.

2. The method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission according to claim 1, characterized in that, The alignment of spatial coordinates in step S2 includes: In response to the conveying assembly operating under constant speed conveying conditions, according to time delay Determine the time delay of transmission acquisition relative to reflection acquisition, and... The corresponding time is taken as the corresponding time of the transmission image of the corresponding region, wherein... The distance between the center of the reflection detection station and the transmission detection station is denoted as . For conveying speed, The moment the reflected image was acquired; Using pulse equivalent q The encoder determines the number of trigger delay pulses based on the center distance between the reflection detection station and the transmission detection station: ,in, For rounding operations; in response to the non-constant conveying speed, when the cumulative pulse value of the encoder reaches... At that time, the corresponding transmission image rows are acquired or retrieved to determine the longitudinal correspondence of the corresponding regions in the running direction, wherein, The encoder accumulates pulse values ​​when acquiring reflected images; Based on time delay Or trigger delay pulse number Determine the initial correspondence between the corresponding regions using the reflection-transmission coordinate transformation matrix. The initial spatial coordinate mapping is completed, and the initial spatial coordinate mapping result is corrected according to the outer contour, centroid, leading edge point, or boundary features of the wet fish fillet to be tested, to obtain the spatial coordinate alignment result of the corresponding regions of the reflected image and the transmitted image; wherein, the coordinate mapping is expressed as: ； In the formula, These are the pixel coordinates in the reflected image. These correspond to the pixel coordinates in the transmitted image.

3. The method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission according to claim 1, characterized in that, The high-light suppression reflection amount is constructed by fusing in the following manner: Calculate specular indication based on normalized parallel polarization reflectance images and normalized orthogonal polarization reflectance images: ； In the formula, This refers to the normalized parallel polarization reflection image at the k-th band and the m-th illumination angle. This corresponds to the normalized orthogonal polarization reflection image; This is the balance coefficient for the two polarization channels; The value is set to a positive number to avoid the denominator being zero. This means that only the positive values ​​within the parentheses should be retained. The highlight suppression coefficient is determined based on the highlight indication value: ； In the formula, The positive threshold for highlight suppression is determined based on the statistical distribution of highlight indicators in areas without obvious specular highlights in normal fish fillets. A numeric truncation function that restricts input values ​​to the range of 0 to 1; The parallel polarization reflection image and the orthogonal polarization reflection image are fused pixel-level using the specular suppression coefficient to construct the first... The first band, the first Specular reflection suppression at each illumination angle: 。 4. The method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission according to claim 1, characterized in that, In step S3, the thickness proxy amount is constructed as follows: ； ； In the formula, This is a normalized transmission image; This is the transmittance after numerical truncation; The default value is a positive number, used to avoid zero input in logarithmic operations; The normal reflection baseline model is characterized by the following formula: ； In the formula, For the predicted reflection baseline model, Mapping function for local fiber direction or texture direction features Obtained through training based on corresponding region samples from normal fish fillets; The extraction methods for the local fiber direction or texture direction features include: The texture direction input image is constructed based on the amount of specular suppression reflection under preset bands and preset illumination angles, or by fusing the specular suppression reflection amounts under multiple bands and illumination angles. Calculate the local gradients of the texture direction input image in different directions, and use a Gaussian smoothing kernel to convolve and smooth the gradient energy to construct a structure tensor; The local principal gradient direction angle is determined based on the gradient energy component and cross-correlation component in the structural tensor. The local principal gradient direction angle is then rotated by π / 2 and normalized according to the π period to obtain the local fiber direction or texture direction feature.

5. The method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission according to claim 1, characterized in that, The thickness compensation residual is expressed as: ； in, To compensate for the thickness residual; This corresponds to the amount of specular reflection suppression; This corresponds to the normal reflection baseline model; The polarization difference is expressed as: ； In the formula, This is a measure of polarization difference; Preset to positive numbers; This is a normalized parallel polarization reflection image; This is a normalized orthogonal polarization reflectance image; The multi-angle reflection difference is expressed as: ； In the formula, This represents the difference in reflection from multiple angles. and The first The amount of highlight suppression reflection in each band under different illumination angles; Preset to positive numbers; The thickness compensation residual and polarization difference are averaged and aggregated for each detection band and each illumination angle to obtain the comprehensive thickness compensation residual and comprehensive polarization difference; the multi-angle reflection difference is averaged and aggregated for each detection band to obtain the comprehensive multi-angle reflection difference. The transmission anomaly is constructed in the following manner: ； ； ； In the formula, The local normal thickness background is calculated by taking the median value of the thickness proxy set of all pixels within a local sliding window. The current position of the pixel to be measured Thickness at the agent level; To use the current position of the pixel to be measured (x,y) A preset local sliding window centered on the target area; H(u,v) The position of each neighboring pixel within the local window (u,v) The thickness of the agent at the point satisfies , This is an abnormal amount of transmission. The standard deviation of the residual difference between the thickness proxy amount within the normal area of ​​a standard defect-free fish fillet and the corresponding local normal thickness background; This indicates the standard deviation calculated within the normal range of standard defect-free fish fillets; Preset to positive numbers; The edge sharpness feature is calculated in the following way: The gradient magnitudes of the integrated thickness compensation residual and the transmission anomaly are calculated separately to obtain the reflection edge sharpness. G R (x,y) and the sharpness of the transmission edge G T (x,y) ; ； ； In the formula, To compensate for residual thickness; G R (x,y) For the sharpness of the reflection edge; G T (x,y) For the sharpness of the transmission edge; x , y These are the pixel coordinates of the image; Based on the comparison result between the transmission anomaly and the preset edge gating threshold, an adaptive gating selection is performed between the reflected edge sharpness and the transmission edge sharpness: when the transmission anomaly at the pixel position is greater than or equal to the preset edge gating threshold, the transmission edge sharpness is used as the edge sharpness feature at the pixel position; otherwise, the reflected edge sharpness is used as the edge sharpness feature. The directional consistency feature Extract using the following method: Calculate the local gradient for the specular suppression reflection amount or the thickness compensation residual, and construct the corresponding structure tensor. : ； In the formula, and The input specular suppression reflection amount or the thickness compensation residual are respectively in direction and Gradient of direction, For Gaussian smoothing kernel, This represents the convolution operation; Let the structure tensor The two eigenvalues ​​are and ,and Then the directional consistency feature The calculation is as follows: ； In the formula, Preset to positive numbers; The comprehensive thickness compensation residual, the comprehensive polarization difference, the comprehensive multi-angle reflection difference, the transmission anomaly, the edge sharpness feature, and the direction inconsistency obtained from the direction consistency feature are all scale-normalized to obtain the corresponding normalized feature values. The anomaly enhancement index is then calculated according to the following formula: ； ； In the formula, This is an abnormal enhancement index; This is the normalized value of the absolute value of the comprehensive thickness compensation residual; This is the normalized value of the comprehensive polarization difference; This is the normalized value of the comprehensive multi-angle reflection difference; This is the normalized value of the transmission anomaly. This is a normalization factor for the edge sharpness feature; This is the normalized quantity of the inconsistency in direction; to The weights are non-negative and the sum of the weights is 1. They can be determined through training with calibration samples or pre-set according to the detection task. Norm [·] indicates that the input features are dimensionless according to the statistical range of normal fish fillet samples, so that different physical quantities are mapped to a uniform scale.

6. The method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission according to claim 5, characterized in that, The process of extracting the candidate anomaly regions and calculating the region features includes: Based on the comparison result between the anomaly enhancement index and the preset discrimination threshold, threshold segmentation and connected component analysis are performed to extract candidate anomaly region masks; Based on camera spatial calibration parameters or planar calibration matrix, the total number of pixels contained in the candidate abnormal region mask is converted into the physical region area, and the aspect ratio is calculated based on the minimum bounding rotation rectangle of the candidate abnormal region mask. Within the range defined by the candidate anomaly region mask, the average pixel values ​​within the calculation area are obtained for the corresponding thickness proxy, transmission anomaly, comprehensive polarization difference, comprehensive multi-angle reflection difference, directional consistency feature, and edge sharpness feature, respectively, to obtain the corresponding average thickness proxy, average transmission anomaly, average polarization difference, average multi-angle reflection difference, average directional consistency feature, and average edge sharpness feature; within the range defined by the candidate anomaly region mask, the average pixel values ​​within the calculation area of ​​the thickness compensation residual corresponding to each band and each illumination angle are obtained to obtain the average reflection residual corresponding to each band and each illumination angle.

7. The method for detecting wet fish fillets based on the combined measurement of polarization reflection and transmission according to claim 6, characterized in that, Based on the aforementioned regional features, the region type is determined, and the corresponding anomaly category label is output, including: If the area of ​​the candidate abnormal region is within a preset area range, the mean value of the reflection residual corresponding to the candidate abnormal region in the 660±10nm band exceeds the normal reflection residual range corresponding to that band, and the aspect ratio of the candidate abnormal region is less than or equal to a preset aspect ratio threshold, then the abnormal category label is determined as a blood spot or ecchymosis type region. If the area of ​​the candidate abnormal region is within the preset area range, the mean reflection residual of the candidate abnormal region in the 450±10nm and 530±10nm bands both exceed the normal reflection residual range corresponding to the corresponding bands, the average polarization difference of the candidate abnormal region is greater than or equal to the polarization difference threshold, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, the average directional consistency feature is less than or equal to the directional consistency threshold, and the aspect ratio of the candidate abnormal region is greater than the preset aspect ratio threshold, then the abnormal category label is determined as a fascial residual region. If the mean reflection residual values ​​of the candidate abnormal region in the 450±10nm and 530±10nm bands both exceed the normal reflection residual range corresponding to the corresponding bands, the area of ​​the candidate abnormal region is greater than the upper limit of the preset area range, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, and the average directional consistency feature is less than or equal to the directional consistency threshold, then the abnormal category label is determined as a skin residue region. If the area of ​​the candidate abnormal region is within the preset area range, the average transmission abnormality of the candidate abnormal region is greater than or equal to the first judgment threshold, the average edge sharpness feature is greater than or equal to the second judgment threshold, and the aspect ratio of the candidate abnormal region is greater than the preset aspect ratio threshold, then the abnormal category label is determined as a bone spur or hard foreign body region. If the area of ​​the candidate abnormal region is within the preset area range, the average edge sharpness feature of the candidate abnormal region is greater than or equal to the second judgment threshold, the average directional consistency feature is less than or equal to the directional consistency threshold, and the average reflection residual value corresponding to at least two of the four reflection bands of 450±10nm, 530±10nm, 660±10nm and 810±15nm exceeds the normal reflection residual range corresponding to the corresponding band, but the candidate abnormal region does not meet the judgment conditions of the blood spot or ecchymosis type region, and its average transmission abnormality is less than the first judgment threshold, then the abnormal category label is determined as a tissue damage type region. If the average reflection residual value of the candidate abnormal region in at least one of the 450±10nm, 530±10nm and 810±15nm bands exceeds the normal reflection residual range corresponding to the corresponding band, the average polarization difference of the candidate abnormal region is greater than or equal to the polarization difference threshold, the average multi-angle reflection difference is greater than or equal to the multi-angle reflection difference threshold, the average transmission anomaly is less than the first determination threshold, the area of ​​the candidate abnormal region is within the preset area range, and the aspect ratio of the candidate abnormal region is less than or equal to the preset aspect ratio threshold, then the abnormal category label is determined as a contamination attachment type region. If the average thickness proxy of the candidate abnormal region exceeds the normal thickness proxy range, the average transmission abnormality is greater than or equal to the first judgment threshold, and the mean reflection residual values ​​corresponding to each band of 450±10nm, 530±10nm, 660±10nm and 810±15nm do not exceed the normal reflection residual range corresponding to the corresponding band, and the candidate abnormal region does not meet the judgment conditions of the blood spot or ecchymosis type region, fascia residue type region, skin residue type region, bone spur or hard foreign body type region, tissue damage type region and contamination attachment type region, then the abnormal category label is determined as uneven thickness, folded edge or local overlap type region; The normal reflection residual interval, polarization difference threshold, multi-angle reflection difference threshold, directional consistency threshold, first judgment threshold, second judgment threshold, normal thickness proxy interval, preset area interval, and preset aspect ratio threshold corresponding to each band are determined based on the statistical distribution of normal fish fillet samples and corresponding abnormal calibration samples.

8. A wet fish fillet detection device based on the combined measurement of polarization reflection and transmission, characterized in that, The device includes: The conveying assembly is used to carry and transport the wet fish fillets to be tested through the reflection detection station and the transmission detection station; A flattening and thickness limiting component is disposed in front of or at the reflection detection station, and is used to guide, flatten, press and limit the thickness of the wet fish fillet to be tested. The testing chamber surrounds the reflection testing station and the transmission testing station; A reflection acquisition component is installed at the reflection detection station to acquire the reflection image information of the wet fish fillet to be tested; A transmission acquisition component is disposed at the transmission detection station and spaced apart from the reflection acquisition component, for acquiring transmission images of the wet fish fillet to be tested; The processor, electrically connected to the reflection acquisition component and the transmission acquisition component, is used to receive the reflection image and the transmission image, and to execute the corresponding steps of the wet fish fillet detection method based on the coordinated measurement of polarization reflection and transmission as described in any one of claims 1-7.

9. The wet fish fillet detection device based on the combined measurement of polarization reflection and transmission according to claim 8, characterized in that, The reflection acquisition component includes: The illumination source includes illumination source one and illumination source two, which emit multi-band detection light into the detection field of view of the reflection detection station from different incident directions; illumination source one and illumination source two each include independently controllable light-emitting units corresponding to multiple detection bands; A linear polarization component, including a first linear polarization component and a second linear polarization component, is respectively disposed on the outgoing light paths of the first illumination source and the second illumination source, and is used to convert the corresponding multi-band detection light into linearly polarized detection light with a preset incident polarization direction; the first linear polarization component and the second linear polarization component are configured to enable the two illumination sources to form linearly polarized detection light with the same preset incident polarization direction; A polarization beam splitter, located in the reflection acquisition channel, is used to receive reflected light from the detection field of view and separate the reflected light into parallel polarization reflection signals and orthogonal polarization reflection signals relative to the preset incident polarization direction. Parallel acquisition camera and orthogonal acquisition camera are respectively set on the first and second outgoing optical paths of the polarization beam splitter, and are used to acquire parallel polarization reflection images and orthogonal polarization reflection images under the same detection field of view. The incident angle of the first illumination source is set to 20° to 35°, and the incident angle of the second illumination source is set to 45° to 65°. The processor controls the light-emitting units of different detection bands and the first and second illumination sources to be activated sequentially according to the preset acquisition sequence, and synchronously triggers the parallel acquisition camera and the orthogonal acquisition camera to obtain the parallel polarization reflection image and the orthogonal polarization reflection image corresponding to each detection band and each illumination angle.

10. The wet fish fillet detection device based on the combined measurement of polarization reflection and transmission according to claim 8, characterized in that, The transmission acquisition component includes: A transmission light source is positioned above the conveying assembly to emit transmission detection light that penetrates the wet fish fillet to be tested at the transmission detection station. A light-transmitting bearing area is provided on the conveying component; A transmission camera is positioned below the conveying assembly to acquire a transmission image penetrating the wet fish fillet to be tested from below the conveying assembly. The conveying component is equipped with an encoder; the transmitted light source is a large-area diffuse narrowband backlight with an emission wavelength of 850±15nm.

Images