A visual recognition method for bone and meat distribution of beef segmentation

Abstract

The present application relates to the technical field of image analysis, and discloses a visual recognition method for beef segmentation, which comprises the following steps: capturing spectral image data of a beef part to be segmented to obtain source image data; performing local binary pattern coding on the source image data to obtain a surface texture feature response graph; performing multi-scale wavelet decomposition on the source image data and performing gradient amplitude analysis on the image to obtain a gradient energy tensor; performing adaptive threshold segmentation on a meat quality feature response area and a bone feature response area to obtain an initial distribution marker graph; performing perspective projection transformation on a collection visual angle to obtain a bone space occupation prior mask; performing spatial position consistency verification on the bone feature response area to obtain a spatial distribution feature graph; and performing feature vector concatenation and splicing on the surface texture feature response graph and the spatial distribution feature graph to obtain a part distribution marker graph; and the present application can improve the efficiency of visual recognition of beef segmentation, bone distribution and meat distribution.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to a visual recognition method for bone and meat distribution in beef segmentation. Background Technology

[0002] In the field of visual recognition of bone and meat distribution in beef segmentation, existing technologies rely on a single method for capturing spectral data of beef parts, only acquiring basic image information. They cannot obtain precise source image data through multi-dimensional light field projection and signal conversion, resulting in incomplete and inaccurate foundational data for subsequent feature extraction. This makes it difficult to accurately reflect the actual distribution of bone and meat in beef parts. Furthermore, existing technologies rely solely on single texture or gradient features for feature analysis of image data, failing to achieve multi-feature fusion processing. This results in insufficient feature mining of beef surface texture and internal gradient energy, failing to form a comprehensive feature response system and leading to low recognition accuracy in distinguishing bone and meat features.

[0003] Existing technologies for segmenting bone and meat regions use fixed thresholds, failing to adaptively define segmentation boundaries based on the actual gradient energy distribution of the beef cut. This leads to misclassification of bone and meat feature response areas, resulting in insufficient accuracy in the initial distribution map. Furthermore, existing technologies lack prior verification of bone spatial location, failing to validate the spatial consistency of identified bone feature regions using a standard 3D skeletal framework. This results in redundant areas exceeding the actual spatial footprint of the bones in the initial identification results. Additionally, insufficient region aggregation and homogeneity verification after feature fusion renders the final distribution map inaccurate in distinguishing between bone, muscle, and fat categories, failing to meet the practical application requirements for precise beef segmentation. Summary of the Invention

[0004] This invention provides a visual recognition method for bone and meat distribution in beef segmentation to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a visual recognition method for bone and meat distribution in beef segmentation, comprising:

[0006] Pt.1. Capture the spectral data of the beef part to be segmented to obtain the source image data of the beef part to be segmented;

[0007] Pt.2. Local binary mode encoding is performed on the source image data to obtain the surface texture feature response map of the beef part to be segmented;

[0008] Pt.3. Perform multi-scale wavelet decomposition on the source image data, and perform gradient magnitude analysis on the decomposed image to obtain the gradient energy tensor of the beef part to be segmented.

[0009] Pt.4. Based on the fused gradient energy field constructed by the gradient energy tensor, adaptive threshold segmentation is performed on the meat feature response area and bone feature response area of ​​the beef part to be segmented to obtain the initial distribution label map of the beef part to be segmented.

[0010] Pt.5. Based on a preset standard 3D skeleton framework, the perspective projection transformation of the source image data acquisition viewpoint is performed to obtain the skeletal space occupancy prior mask of the beef part to be segmented.

[0011] Pt.6. Based on the bone space occupancy prior mask, perform spatial position consistency verification on the bone feature response area in the initial distribution marking map to obtain the spatial distribution feature map of the beef part to be segmented.

[0012] Pt.7. The surface texture feature response map and the spatial distribution feature map are concatenated and spliced ​​together, and the concatenated multidimensional feature tensor is aggregated into regions to obtain the part distribution marking map of the beef part to be segmented.

[0013] In a preferred embodiment, the step of capturing spectral data of the beef portion to be segmented to obtain source image data of the beef portion to be segmented includes:

[0014] A structured light field is projected onto the surface of the beef section to be cut, wherein the beef section to be cut is a multi-band structured light illumination sequence;

[0015] Based on the multi-band structured light illumination sequence, the echo beam of the beef part to be segmented is split to obtain a single-wavelength beam of the beef part to be segmented.

[0016] The single-wavelength beam is photoelectrically converted to obtain the analog electrical signal of the beef part to be segmented;

[0017] The analog electrical signal is converted from analog to digital to obtain a digital image frame of the beef part to be segmented;

[0018] The digital image frames are spatially registered according to the acquisition time sequence and angle information to obtain the source image data of the beef part to be segmented.

[0019] In a preferred embodiment, the step of performing local binary pattern encoding on the source image data to obtain the surface texture feature response map of the beef portion to be segmented includes:

[0020] Based on a preset neighborhood sampling topology, the source image data is traversed pixel by pixel to obtain the center pixel of the beef part to be segmented, and the gray value of the center pixel in the neighborhood range is obtained to obtain the neighborhood gray value sequence of the beef part to be segmented.

[0021] The neighboring gray value sequence is compared with the gray value of the center pixel element by element to obtain the binarized encoding sequence of the beef part to be segmented.

[0022] Binary weights are assigned to the binarized encoded sequence to obtain the local binary pattern encoded value of the beef part to be segmented;

[0023] The local binary pattern encoded values ​​are arranged pixel by pixel to construct the local binary pattern encoded map of the beef part to be segmented;

[0024] Multi-channel fusion of the local binary pattern coding map is performed to obtain the surface texture feature response map of the beef part to be segmented.

[0025] In a preferred embodiment, the step of performing multi-scale wavelet decomposition on the source image data and performing gradient magnitude analysis on the decomposed image to obtain the gradient energy tensor of the beef segment to be segmented includes:

[0026] The source image data is subjected to wavelet decomposition filtering to obtain the original wavelet sub-band image sequence of the beef part to be segmented;

[0027] Based on the spatial orientation of the beef part to be segmented, the original wavelet sub-band image sequence is directionally separated to obtain the horizontal high-frequency sub-band image, the vertical high-frequency sub-band image, and the diagonal high-frequency sub-band image of the beef part to be segmented.

[0028] The high-frequency sub-band images in the horizontal direction, the vertical direction, and the diagonal direction are analyzed for directional gradient within the pixel neighborhood to obtain the high-frequency gradient response map of the beef part to be segmented.

[0029] Energy accumulation is performed on the high-frequency gradient response map to obtain a single-scale gradient energy map of the beef part to be segmented;

[0030] The single-scale gradient energy map is stacked to construct the gradient energy tensor of the beef part to be segmented.

[0031] In a preferred embodiment, the fused gradient energy field constructed based on the gradient energy tensor is used to adaptively threshold the meat feature response region and bone feature response region of the beef part to be segmented, resulting in an initial distribution map of the beef part to be segmented, including:

[0032] Spatial coordinate registration is performed on the single-scale gradient energy map in the gradient energy tensor, and the scale is normalized to obtain the multi-scale gradient energy map sequence of the beef part to be segmented.

[0033] The gradient energy response amplitude of the pixel coordinates in the multi-scale gradient energy map sequence is extracted, and the real-time gradient energy response amplitude is accumulated and fused to construct the fused gradient energy field of the beef part to be segmented.

[0034] By performing distribution histogram statistics on the fused gradient energy field, the trough locations and inflection point abrupt changes of the beef portion to be segmented are obtained;

[0035] Based on the trough location and the inflection point abrupt change location, the energy amplitude segmentation boundary is defined to obtain the adaptive segmentation threshold of the beef part to be segmented;

[0036] Based on the adaptive segmentation threshold, the category attribute of the fused gradient energy field is determined, and the pixels that penetrate the adaptive segmentation threshold are assigned as the skeletal feature response points of the beef part to be segmented, and the pixels that do not penetrate the adaptive segmentation threshold are assigned as the meat quality feature response points of the beef part to be segmented.

[0037] Spatial assignment and filling are performed on the skeletal feature response points and the meat quality feature response points to obtain the initial distribution marking map of the beef part to be segmented.

[0038] In a preferred embodiment, the step of performing perspective projection transformation on the acquisition perspective of the source image data based on a preset standard skeletal 3D framework to obtain a priori mask for the skeletal space occupancy of the beef part to be segmented includes:

[0039] Spatial coordinate analysis is performed on the preset standard 3D skeleton framework to obtain the standard 3D point cloud framework of the beef part to be segmented.

[0040] Based on the optical center position parameters and focal plane orientation parameters of the source image data, construct the perspective projection matrix of the beef part to be segmented;

[0041] Based on the perspective projection matrix, the vertex coordinates in the standard skeletal 3D point cloud framework are projected into spatial coordinates to obtain the 2D skeletal contour point set of the beef part to be segmented.

[0042] Based on the topological connection relationship between the vertex coordinates, the adjacent contour points in the two-dimensional skeleton contour point set are closed and filled to obtain the two-dimensional continuous contour surface of the beef part to be segmented.

[0043] The two-dimensional continuous contour surface is mapped onto a blank canvas of the same size as the source image data, and the pixel positions of the two-dimensional continuous contour surface are assigned a state to obtain the skeletal space occupancy prior mask of the beef part to be segmented.

[0044] In a preferred embodiment, the step of performing spatial position consistency verification on the skeletal feature response regions in the initial distribution marker map based on the skeletal spatial occupancy prior mask to obtain the spatial distribution feature map of the beef part to be segmented includes:

[0045] The connected regions of the skeletal feature response points in the initial distribution map are separated and analyzed to obtain the initial set of skeletal response connected regions of the beef part to be segmented;

[0046] Boundary contour tracing is performed on the bone space occupancy prior mask to obtain the bone space boundary of the beef part to be segmented;

[0047] Based on the spatial boundary of the bone occupancy space, the initial bone response connected domain set is spatially overlapped and compared to obtain the spatial consistency determination result of the beef part to be segmented.

[0048] Based on the spatial consistency determination result, redundant filtering is performed on the connected components located outside the boundary of the bone-occupying space in the initial distribution marking map to obtain the bone response region of the beef part to be segmented.

[0049] The skeletal response region is fused with the meat quality feature response region in the initial distribution marker map to obtain the spatial distribution feature map of the beef part to be segmented.

[0050] In a preferred embodiment, the formula for calculating the spatial overlap confidence in the spatial consistency determination result is as follows:

[0051] ;

[0052] In the formula, The confidence level for the spatial overlap. The overlapping area is the region between the connected components in the initial skeleton response connected component set and the boundary of the skeleton's occupied space. This represents the total area of ​​the internal region of the boundary of the space occupied by the skeleton. The shortest Euclidean distance from the centroid of the connected domain to the boundary of the bone's occupied space is given by denoted as ... The preset spatial scale normalization factor, For the pixels in the overlapping region The gradient energy response amplitude in the fused gradient energy field This is the adaptive segmentation threshold for the beef portion to be segmented. The total number of pixels in the overlapping region. The preset spatial overlap weighting coefficient, The preset distance attenuation weighting coefficient, The preset energy consistency weighting coefficient, This refers to the overlapping region between the connected components in the initial skeleton response connected component set and the boundary of the skeleton's occupied space.

[0053] In a preferred embodiment, the step of concatenating and stitching the surface texture feature response map and the spatial distribution feature map into feature vectors, and then performing region aggregation on the concatenated multidimensional feature tensor to obtain a location distribution marker map of the beef part to be segmented, includes:

[0054] The surface texture feature response map and the spatial distribution feature map are rigidly registered to obtain the texture map and spatial map of the beef part to be segmented;

[0055] Based on the local binary pattern encoding value of the texture map and the bone and meat attribute identifiers in the spatial map, the texture map and the spatial map are serially spliced ​​to obtain the joint feature descriptor of the beef part to be segmented.

[0056] The joint feature descriptor is mapped to the original image coordinate matrix of the source image data to construct the initial multidimensional feature tensor of the beef part to be segmented;

[0057] The initial multidimensional feature tensor is labeled with neighborhood connectivity to obtain the initial aggregated connected component set of the beef part to be segmented;

[0058] The initial aggregated connected component set is subjected to a feature homogeneity check within the region to obtain a pure aggregated connected component set of the beef part to be segmented;

[0059] The cluster centers of the pure aggregated connected domain set in the feature space are counted, and based on the membership distance between the cluster centers and the preset bone category prototype, muscle category prototype and fat category prototype, the pure aggregated connected domain set is assigned a bone category label, muscle category label or fat category label to obtain the part distribution label map of the beef part to be segmented.

[0060] In a preferred embodiment, the step of performing regional feature homogeneity verification on the initial aggregated connected component set to obtain a pure aggregated connected component set for the beef part to be segmented includes:

[0061] The joint feature descriptors in the initial aggregated connected component set are evaluated using mean vectors to obtain the feature distribution statistics of the beef part to be segmented.

[0062] Based on the aforementioned feature distribution statistics, the Mahalanobis distance measure is applied to the initial aggregated connected component set to obtain the feature deviation of the beef part to be segmented.

[0063] The feature deviation is compared with a preset deviation tolerance threshold, and the pixels whose feature deviation exceeds the deviation tolerance threshold are assigned as outlier pixels of the beef part to be segmented based on the comparison result.

[0064] Spatial location stripping is performed on the outlier pixels, and spatial weighted interpolation is performed on the empty pixel positions formed after filtering to obtain the optimized pixel set of the beef part to be segmented.

[0065] The connected components of the optimized pixel set are topologically reconstructed to obtain a pure aggregated connected component set of the beef part to be segmented.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. This invention achieves spectral image data capture through multi-band structured light field projection and multi-step signal conversion. It combines local binary mode coding and multi-scale wavelet decomposition to accurately extract surface texture and gradient energy features. Based on the gradient energy tensor, it constructs a fused gradient energy field and completes adaptive threshold segmentation, making the division of bone and meat feature response areas more consistent with the actual tissue distribution of beef. This significantly improves the accuracy of region recognition in the initial distribution map. At the same time, it uses a standard 3D skeletal framework to complete perspective projection transformation to obtain a priori mask for bone space occupancy. It performs spatial position consistency verification on the initially identified bone feature response areas, effectively filtering out redundant recognition areas and making the spatial distribution features of bone and meat more realistic.

[0068] 2. This invention concatenates and stitches feature vectors from surface texture feature response maps and spatial distribution feature maps. It performs deep processing of multidimensional feature tensors through region aggregation, and after feature fusion, performs feature homogeneity verification within the region, removes outlier pixels, and completes pixel interpolation and topological reconstruction. This significantly improves the feature uniformity of the aggregated connected domains. Finally, based on the feature space clustering centers, it accurately labels the categories of bones, muscles, and fat. The resulting part distribution marking map clearly and accurately reflects the tissue distribution of beef parts, providing a highly reliable visual recognition basis for accurate beef segmentation and enhancing the practical application value and feasibility of beef segmentation visual recognition technology. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating a method for visually recognizing the distribution of bones and meat in beef segmentation, according to an embodiment of the present invention.

[0070] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0071] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0072] This application provides a method for visually recognizing the bone and meat distribution in beef segmentation. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for visually recognizing the bone and meat distribution in beef segmentation can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0073] Reference Figure 1 The diagram shown is a flowchart illustrating a method for visually recognizing the distribution of bones and meat in beef segmentation according to an embodiment of the present invention. In this embodiment, the method for visually recognizing the distribution of bones and meat in beef segmentation includes:

[0074] Pt.1. Capture the spectral data of the beef part to be segmented to obtain the source image data of the beef part to be segmented;

[0075] In this embodiment of the invention, the step of capturing spectral data of the beef portion to be segmented to obtain source image data of the beef portion to be segmented includes:

[0076] A structured light field is projected onto the surface of the beef section to be cut, wherein the beef section to be cut is a multi-band structured light illumination sequence;

[0077] Based on the multi-band structured light illumination sequence, the echo beam of the beef part to be segmented is split to obtain a single-wavelength beam of the beef part to be segmented.

[0078] The single-wavelength beam is photoelectrically converted to obtain the analog electrical signal of the beef part to be segmented;

[0079] The analog electrical signal is converted from analog to digital to obtain a digital image frame of the beef part to be segmented;

[0080] The digital image frames are spatially registered according to the acquisition time sequence and angle information to obtain the source image data of the beef part to be segmented.

[0081] A structured light field, consisting of different monochromatic lights within a preset wavelength range combined sequentially at fixed time intervals, is projected onto the surface of the beef section to be cut. This structured light field forms a multi-band structured light illumination sequence covering the entire area to be cut on the surface of the beef section. The wavelength coverage of the illumination sequence matches the optical reflection characteristics of the beef bones and meat, ensuring that the light field can form differentiated light reflection effects on different tissue surfaces of the beef.

[0082] Based on the established multi-band structured light illumination sequence of the beef parts to be segmented, a beam splitter is used to perform band separation processing on the echo beams reflected from the surface of the beef parts to be segmented. According to the difference in the refraction angle of different wavelength beams, the mixed echo beams are separated into single-wavelength beams of the beef parts to be segmented that correspond one-to-one with the wavelengths of the illumination sequence. Each single-wavelength beam maintains an independent propagation path and there is no band mixing.

[0083] A photodiode is used to receive a single-wavelength light beam from the beef section to be cut. The light signal intensity of the single-wavelength light beam is converted into an electrical signal that is linearly related to the light intensity. This electrical signal is a continuously changing analog electrical signal of the beef section to be cut, and the amplitude change of the electrical signal perfectly matches the light intensity change pattern of the single-wavelength light beam.

[0084] An analog-to-digital converter chip is used to discretize the analog electrical signal of the beef part to be segmented. The continuous analog electrical signal is sampled at equal time intervals according to a preset sampling frequency. The amplitude of the analog electrical signal at each sampling point is converted into a binary digital signal. The digital signals of all sampling points are arranged in the sampling time sequence to form a digital image frame of the beef part to be segmented. The pixel resolution of the digital image frame is positively correlated with the sampling frequency.

[0085] The acquisition time sequence data and acquisition angle data recorded synchronously during the acquisition of digital image frames of the beef part to be segmented are retrieved. Using the geometric center of the beef part to be segmented as the spatial registration origin, digital image frames acquired at different times and angles are mapped to the same three-dimensional spatial coordinate system. Pixel offset areas between image frames are filled in, and pixel grayscale values ​​in overlapping areas are averaged. Finally, source image data of the beef part to be segmented with unified spatial coordinates and complete pixel information are formed.

[0086] Pt.2. Local binary mode encoding is performed on the source image data to obtain the surface texture feature response map of the beef part to be segmented;

[0087] In this embodiment of the invention, the step of performing local binary mode encoding on the source image data to obtain the surface texture feature response map of the beef part to be segmented includes:

[0088] Based on a preset neighborhood sampling topology, the source image data is traversed pixel by pixel to obtain the center pixel of the beef part to be segmented, and the gray value of the center pixel in the neighborhood range is obtained to obtain the neighborhood gray value sequence of the beef part to be segmented.

[0089] The neighboring gray value sequence is compared with the gray value of the center pixel element by element to obtain the binarized encoding sequence of the beef part to be segmented.

[0090] Binary weights are assigned to the binarized encoded sequence to obtain the local binary pattern encoded value of the beef part to be segmented;

[0091] The local binary pattern encoded values ​​are arranged pixel by pixel to construct the local binary pattern encoded map of the beef part to be segmented;

[0092] Multi-channel fusion of the local binary pattern coding map is performed to obtain the surface texture feature response map of the beef part to be segmented.

[0093] Based on a pre-defined 3×3 pixel neighborhood sampling topology, the source image data of the beef part to be segmented is scanned pixel by pixel. Each traversed pixel is determined as the center pixel of the beef part to be segmented. At the same time, the gray values ​​of all pixels in the 3×3 neighborhood around the center pixel are extracted and arranged in a clockwise direction to form the neighborhood gray value sequence of the beef part to be segmented.

[0094] The gray value of each position in the neighborhood gray value sequence of the beef part to be segmented is compared with the gray value of the corresponding center pixel one by one. If the gray value of a certain position in the neighborhood gray value sequence is greater than the gray value of the center pixel, the position is assigned a value of 1. If it is less than or equal to the gray value of the center pixel, the position is assigned a value of 0. In this way, a binary encoding sequence of the beef part to be segmented is formed, consisting of 0 and 1.

[0095] Assign a corresponding power of 2 binary weight to each digit in the binary encoding sequence of the beef part to be segmented, in descending order of high digit to low digit. Starting from the first digit of the encoding sequence, assign weights of 2 to the power of 7, 2 to the power of 6, and so on, up to 2 to the power of 0. Multiply each digit by its corresponding weight and sum the results. The resulting value is the local binary pattern encoding value of the beef part to be segmented.

[0096] The local binary pattern encoding value corresponding to each central pixel of the beef part to be segmented is precisely mapped to the original coordinate position of the central pixel in the source image data. All encoding values ​​are arranged completely according to the pixel arrangement of the source image data to construct the local binary pattern encoding map of the beef part to be segmented.

[0097] By superimposing channel pixel values, multi-channel fusion processing is performed on the local binary pattern coding map of the beef part to be segmented. The coding values ​​of the same pixel coordinate position in each channel of the map are averaged and the calculated average value is used as the pixel value of that pixel coordinate position after fusion. After completing the fusion calculation of all pixel positions, the surface texture feature response map of the beef part to be segmented is formed.

[0098] Pt.3. Perform multi-scale wavelet decomposition on the source image data, and perform gradient magnitude analysis on the decomposed image to obtain the gradient energy tensor of the beef part to be segmented.

[0099] In this embodiment of the invention, the step of performing multi-scale wavelet decomposition on the source image data and performing gradient magnitude analysis on the decomposed image to obtain the gradient energy tensor of the beef part to be segmented includes:

[0100] The source image data is subjected to wavelet decomposition filtering to obtain the original wavelet sub-band image sequence of the beef part to be segmented;

[0101] Based on the spatial orientation of the beef part to be segmented, the original wavelet sub-band image sequence is directionally separated to obtain the horizontal high-frequency sub-band image, the vertical high-frequency sub-band image, and the diagonal high-frequency sub-band image of the beef part to be segmented.

[0102] The high-frequency sub-band images in the horizontal direction, the vertical direction, and the diagonal direction are analyzed for directional gradient within the pixel neighborhood to obtain the high-frequency gradient response map of the beef part to be segmented.

[0103] Energy accumulation is performed on the high-frequency gradient response map to obtain a single-scale gradient energy map of the beef part to be segmented;

[0104] The single-scale gradient energy map is stacked to construct the gradient energy tensor of the beef part to be segmented.

[0105] A multi-level wavelet decomposition filtering operation is performed on the source image data of the beef part to be segmented using a preset wavelet basis function. The source image data is decomposed at scale level by scale in power of 2. Each level of decomposition retains the image detail information and contour information at the corresponding scale. The sub-band images obtained from all scales are arranged in order of increasing decomposition scale to form the original wavelet sub-band image sequence of the beef part to be segmented.

[0106] Based on the three fixed spatial directions of horizontal, vertical, and diagonal of the beef part to be segmented, directional separation processing is performed on the original wavelet sub-band image sequence of the beef part to be segmented. The gray-level change information in the horizontal direction is extracted to form a high-frequency sub-band image in the horizontal direction, the gray-level change information in the vertical direction is extracted to form a high-frequency sub-band image in the vertical direction, and the gray-level change information in the diagonal direction is extracted to form a high-frequency sub-band image in the diagonal direction. The high-frequency sub-band images in all three directions completely preserve the image edge and texture gradient information in the corresponding direction.

[0107] Using a 3×3 pixel neighborhood template, directional gradient analysis is performed on the horizontal, vertical, and diagonal high-frequency sub-band images of the beef part to be segmented. The difference in grayscale value change of each pixel in the corresponding directional neighborhood is calculated, and this difference is used as the gradient magnitude of the pixel and assigned to the corresponding pixel. After calculating the gradient magnitude of all pixels, they are integrated to form the high-frequency gradient response map of the beef part to be segmented.

[0108] A region-by-region energy accumulation operation is performed on the high-frequency gradient response map of the beef part to be segmented. Each energy accumulation unit is 16×16 pixels. The sum of squares of the gradient magnitudes of all pixels in each unit is calculated. The calculation result is used as the gradient energy value of the unit and assigned to all pixels in the unit. After the energy calculation of all accumulation units is completed, a single-scale gradient energy map of the beef part to be segmented is formed.

[0109] The single-scale gradient energy maps obtained at different decomposition scales of the beef part to be segmented are stacked in three-dimensional space in order of increasing decomposition scale, so that each layer of single-scale gradient energy map corresponds to one dimension of the gradient energy tensor. The pixel coordinates of each layer of map are one-to-one and the spatial coordinates are kept consistent, thus constructing the gradient energy tensor of the beef part to be segmented containing multi-scale gradient energy information.

[0110] Pt.4. Based on the fused gradient energy field constructed by the gradient energy tensor, adaptive threshold segmentation is performed on the meat feature response area and bone feature response area of ​​the beef part to be segmented to obtain the initial distribution label map of the beef part to be segmented.

[0111] In this embodiment of the invention, the fused gradient energy field constructed based on the gradient energy tensor is used to adaptively threshold the meat feature response region and bone feature response region of the beef part to be segmented, to obtain an initial distribution map of the beef part to be segmented, including:

[0112] Spatial coordinate registration is performed on the single-scale gradient energy map in the gradient energy tensor, and the scale is normalized to obtain the multi-scale gradient energy map sequence of the beef part to be segmented.

[0113] The gradient energy response amplitude of the pixel coordinates in the multi-scale gradient energy map sequence is extracted, and the real-time gradient energy response amplitude is accumulated and fused to construct the fused gradient energy field of the beef part to be segmented.

[0114] By performing distribution histogram statistics on the fused gradient energy field, the trough locations and inflection point abrupt changes of the beef portion to be segmented are obtained;

[0115] Based on the trough location and the inflection point abrupt change location, the energy amplitude segmentation boundary is defined to obtain the adaptive segmentation threshold of the beef part to be segmented;

[0116] Based on the adaptive segmentation threshold, the category attribute of the fused gradient energy field is determined, and the pixels that penetrate the adaptive segmentation threshold are assigned as the skeletal feature response points of the beef part to be segmented, and the pixels that do not penetrate the adaptive segmentation threshold are assigned as the meat quality feature response points of the beef part to be segmented.

[0117] Spatial assignment and filling are performed on the skeletal feature response points and the meat quality feature response points to obtain the initial distribution marking map of the beef part to be segmented.

[0118] Using the geometric center of the beef part to be segmented as the origin of a unified spatial coordinate system, the pixel coordinates of the single-scale gradient energy maps of each layer in the gradient energy tensor are mapped to this coordinate system to complete spatial coordinate registration. Then, the pixel energy values ​​of each single-scale gradient energy map are mapped to the gray value range of 0 to 255 to complete scale normalization. The pixel energy values ​​and gray values ​​of each map have a linear correspondence, and finally a multi-scale gradient energy map sequence of the beef part to be segmented with unified coordinates and consistent scale is formed.

[0119] Gradient energy response amplitudes at corresponding positions in the multi-scale gradient energy map sequence of the beef part to be segmented are extracted one by one according to pixel coordinates. The gradient energy response amplitudes of all maps at the same pixel coordinate position are accumulated bit by bit, and the accumulated result is used as the fusion energy value at that pixel coordinate position. After the amplitudes of all pixel coordinate positions are accumulated, the fusion gradient energy field of the beef part to be segmented is formed.

[0120] The fusion energy values ​​of all pixels in the fusion gradient energy field of the segmented beef part are statistically analyzed. The statistical intervals are divided according to the numerical range of the fusion energy values. The number of pixels in each interval is counted and a distribution histogram is drawn. The position where the difference in the number of pixels between adjacent statistical intervals turns from negative to positive is extracted from the histogram as the trough region. The position where the number of pixels changes abruptly with the change of energy value is extracted as the inflection point.

[0121] The fusion energy value corresponding to the trough region of the beef part to be segmented is used as the basic limit. The basic limit is then corrected by combining the fusion energy value corresponding to the inflection point mutation position. The corrected fusion energy value is the energy amplitude segmentation limit. This limit is the adaptive segmentation threshold of the beef part to be segmented. The threshold value is the median value between the energy values ​​corresponding to the trough region and the inflection point mutation position.

[0122] The fusion energy value of each pixel in the fusion gradient energy field of the beef part to be segmented is compared with the adaptive segmentation threshold. Pixels with fusion energy values ​​greater than the adaptive segmentation threshold are determined to have penetrated the adaptive segmentation threshold and are directly assigned as skeletal feature response points of the beef part to be segmented. Pixels with fusion energy values ​​less than or equal to the adaptive segmentation threshold are determined to have not penetrated the adaptive segmentation threshold and are directly assigned as meat quality feature response points of the beef part to be segmented.

[0123] Based on the pixel coordinate distribution of the source image data of the beef parts to be segmented, all skeletal feature response points and meat feature response points are restored to their corresponding spatial positions. Connected regions formed by adjacent feature response points of the same type are filled with solid colors. Skeletal feature response areas and meat feature response areas are distinguished by different fill colors. After all areas are filled, an initial distribution mark map of the beef parts to be segmented is formed.

[0124] Pt.5. Based on a preset standard 3D skeleton framework, the perspective projection transformation of the source image data acquisition viewpoint is performed to obtain the skeletal space occupancy prior mask of the beef part to be segmented.

[0125] In this embodiment of the invention, the step of performing perspective projection transformation on the acquisition perspective of the source image data based on a preset standard skeletal 3D framework to obtain a priori mask for the skeletal space occupancy of the beef part to be segmented includes:

[0126] Spatial coordinate analysis is performed on the preset standard 3D skeleton framework to obtain the standard 3D point cloud framework of the beef part to be segmented.

[0127] Based on the optical center position parameters and focal plane orientation parameters of the source image data, construct the perspective projection matrix of the beef part to be segmented;

[0128] Based on the perspective projection matrix, the vertex coordinates in the standard skeletal 3D point cloud framework are projected into spatial coordinates to obtain the 2D skeletal contour point set of the beef part to be segmented.

[0129] Based on the topological connection relationship between the vertex coordinates, the adjacent contour points in the two-dimensional skeleton contour point set are closed and filled to obtain the two-dimensional continuous contour surface of the beef part to be segmented.

[0130] The two-dimensional continuous contour surface is mapped onto a blank canvas of the same size as the source image data, and the pixel positions of the two-dimensional continuous contour surface are assigned a state to obtain the skeletal space occupancy prior mask of the beef part to be segmented.

[0131] A full-dimensional spatial coordinate analysis is performed on the pre-defined standard skeletal 3D framework suitable for beef segmentation. The spatial coordinates of the vertices of all skeletal structures in the framework are extracted. The coordinates of each vertex are arranged in an orderly manner according to the physiological structure and topological relationship of the beef skeleton. All the vertex coordinates together constitute the standard skeletal 3D point cloud framework of the beef part to be segmented. The coordinate accuracy of this framework matches the pixel resolution of the source image data.

[0132] The optical center position parameters and focal plane orientation parameters recorded in real time during the acquisition of source image data of the beef part to be segmented are retrieved. The optical center is used as the projection origin and the normal vector of the focal plane is used as the main projection direction. The matrix elements are assigned and constructed in combination with the three-axis direction parameters of the spatial coordinate system to form a perspective projection matrix of the beef part to be segmented that can realize the transformation from three-dimensional spatial coordinates to two-dimensional planar coordinates. The dimension of the matrix is ​​adapted to the coordinate dimension of the standard skeletal three-dimensional point cloud framework.

[0133] The coordinates of each vertex in the standard 3D point cloud framework of the beef part to be segmented are substituted into the perspective projection matrix one by one to perform spatial coordinate transformation calculation, mapping the vertex coordinates in 3D space to the 2D imaging plane corresponding to the source image data. All the transformed 2D coordinate points are arranged according to the original topological relationship to form a 2D skeletal contour point set of the beef part to be segmented. The coordinates of the points in the point set correspond one-to-one with the pixel coordinates of the source image data.

[0134] Based on the original topological connection relationship of vertex coordinates in the standard 3D point cloud framework of the beef part to be segmented, the adjacent contour points of each contour point in the 2D skeletal contour point set are identified. The adjacent contour points are connected sequentially according to the direction of the skeletal contour to form a closed skeletal contour line. The area enclosed by the closed contour line is filled at the pixel level to form a 2D continuous contour surface of the beef part to be segmented. The filled area completely covers the 2D projection range of the beef skeleton.

[0135] Create a blank canvas with pixel dimensions identical to the source image data of the beef part to be segmented. The pixel resolution and coordinate system of the canvas are consistent with the source image data. Map the two-dimensional continuous contour surface to the blank canvas according to the coordinate correspondence. Assign the pixel positions within the range of the two-dimensional continuous contour surface as valid bone occupancy states, and assign the pixel positions outside the range of the two-dimensional continuous contour surface as invalid bone occupancy states. After completing the state assignment of all pixel positions, the bone space occupancy prior mask of the beef part to be segmented is obtained.

[0136] Pt.6. Based on the bone space occupancy prior mask, perform spatial position consistency verification on the bone feature response area in the initial distribution marking map to obtain the spatial distribution feature map of the beef part to be segmented.

[0137] In this embodiment of the invention, the step of performing spatial position consistency verification on the skeletal feature response regions in the initial distribution marker map based on the skeletal spatial occupancy prior mask to obtain the spatial distribution feature map of the beef part to be segmented includes:

[0138] The connected regions of the skeletal feature response points in the initial distribution map are separated and analyzed to obtain the initial set of skeletal response connected regions of the beef part to be segmented;

[0139] Boundary contour tracing is performed on the bone space occupancy prior mask to obtain the bone space boundary of the beef part to be segmented;

[0140] Based on the spatial boundary of the bone occupancy space, the initial bone response connected domain set is spatially overlapped and compared to obtain the spatial consistency determination result of the beef part to be segmented.

[0141] Based on the spatial consistency determination result, redundant filtering is performed on the connected components located outside the boundary of the bone-occupying space in the initial distribution marking map to obtain the bone response region of the beef part to be segmented.

[0142] The skeletal response region is fused with the meat quality feature response region in the initial distribution marker map to obtain the spatial distribution feature map of the beef part to be segmented.

[0143] The formula for calculating the spatial overlap confidence in the spatial consistency determination result is as follows:

[0144] ;

[0145] In the formula, The confidence level for the spatial overlap. The overlapping area is the region between the connected components in the initial skeleton response connected component set and the boundary of the skeleton's occupied space. This represents the total area of ​​the internal region of the boundary of the space occupied by the skeleton. The shortest Euclidean distance from the centroid of the connected domain to the boundary of the bone's occupied space is given by denoted as ... The preset spatial scale normalization factor, For the pixels in the overlapping region The gradient energy response amplitude in the fused gradient energy field This is the adaptive segmentation threshold for the beef portion to be segmented. The total number of pixels in the overlapping region. The preset spatial overlap weighting coefficient, The preset distance attenuation weighting coefficient, The preset energy consistency weighting coefficient, This refers to the overlapping region between the connected components in the initial skeleton response connected component set and the boundary of the skeleton's occupied space.

[0146] The initial distribution map of the beef part to be segmented is used to detect the connectivity of pixel regions. All connected regions formed by skeletal feature response points are identified using the 8-neighborhood connectivity rule. Boundary boxes are marked and regions are numbered for each independent connected region. All marked and numbered skeletal feature response connected regions are integrated to obtain the initial skeletal response connected region set of the beef part to be segmented.

[0147] Pixel-level contour extraction is performed on the prior mask of the skeletal space occupancy of the beef part to be segmented. A boundary tracing algorithm is used to continuously traverse along the edges of the effective state pixels of the skeletal occupancy in the mask. The coordinates of all edge pixels during the traversal are recorded and connected in sequence to form a closed contour line. This closed contour line is the boundary of the skeletal space occupancy of the beef part to be segmented.

[0148] The pixel coordinates of each connected region in the initial skeletal response connected domain set of the beef part to be segmented are compared point by point with the pixel coordinates of the boundary of the skeletal space. The spatial overlap state, overlap range and relative position of each connected region and the boundary of the skeletal space are determined. The comparison results of all connected regions are summarized and organized to obtain the spatial consistency judgment result of the beef part to be segmented.

[0149] Based on the spatial consistency determination results of the beef parts to be segmented, connected regions whose initial skeletal response connected domains are completely located outside the boundary of the skeletal space are selected. All skeletal feature response points corresponding to such connected regions are directly removed. Connected regions located inside the boundary of the skeletal space and those overlapping with the boundary are retained. The retained regions are integrated to form the skeletal response regions of the beef parts to be segmented.

[0150] The skeletal response region of the beef part to be segmented is restored to the initial distribution map according to the original pixel coordinates. All pixel information of the meat quality feature response region that was not removed in the initial distribution map is retained. The adjacent boundary between the skeletal response region and the meat quality feature response region is smoothed at the pixel level to ensure that there is no pixel discontinuity in the boundary transition between the two regions. After the processing is completed, the spatial distribution feature map of the beef part to be segmented is obtained.

[0151] The spatial overlap confidence values ​​were directly extracted from the relevant processing results of the bone space occupancy consistency verification. The spatial overlap area is the statistical value of the number of pixels in the actual overlapping area between the connected domains in the initial bone response connected domain set and the boundary of the bone space occupancy. The total area of ​​the area inside the boundary of the bone space occupancy is the statistical value of the total number of pixels in the area enclosed by the boundary of the bone space occupancy.

[0152] The shortest Euclidean distance from the centroid of the connected domain to the boundary of the bone space is the minimum value obtained by calculating the distance between the centroid coordinates of the initial bone response connected domain and the pixel coordinates of the bone space boundary. The spatial scale normalization factor is a fixed value pre-set according to the actual application scenario of beef segmentation, which is used to normalize the distance value.

[0153] The gradient energy response amplitude of pixels in the overlapping region is the actual energy amplitude value of the corresponding pixel coordinate position in the fused gradient energy field. The adaptive segmentation threshold is the energy amplitude segmentation boundary value determined based on the trough location and inflection point change location of the fused gradient energy field. The total number of pixels in the overlapping region is the statistical value of the total number of pixels in the overlapping region of the initial skeleton response connected domain and the skeleton occupied space boundary.

[0154] The spatial overlap weighting coefficient, distance attenuation weighting coefficient, and energy consistency weighting coefficient are all fixed values ​​preset based on the visual recognition characteristics of beef bones and meat, and are used to assign weights to the results of different calculation parts.

[0155] The significance of this numerical calculation lies in comprehensively quantifying the spatial overlap between the initial skeletal response connected region and the boundary of the skeletal space. The ratio of the overlapping area to the total area within the skeletal space directly reflects the coverage of the connected region within the skeletal space. Attenuation processing of the shortest distance from the centroid of the connected region to the boundary of the skeletal space reflects the spatial fit between the connected region and the boundary; the closer the distance, the higher the quantization result. The mean absolute value of the difference between the gradient energy response amplitude of pixels within the overlapping region and the adaptive segmentation threshold reflects the consistency between the energy features of pixels within the overlapping region and the skeletal features; the smaller the difference, the higher the quantization result. The three quantization results are weighted and superimposed to obtain a comprehensive quantization value, which serves as the core basis for determining whether the initial skeletal response connected region and the skeletal space have spatial consistency. The magnitude of the value directly reflects the credibility of the connected region as a true skeletal feature response area, providing a precise and quantifiable criterion for subsequent filtering of redundant connected regions, ensuring the accuracy of skeletal response region recognition.

[0156] Pt.7. The surface texture feature response map and the spatial distribution feature map are concatenated and spliced ​​together, and the concatenated multidimensional feature tensor is aggregated into regions to obtain the part distribution marking map of the beef part to be segmented.

[0157] In this embodiment of the invention, the step of concatenating and splicing the surface texture feature response map and the spatial distribution feature map into feature vectors, and then performing region aggregation on the concatenated multidimensional feature tensor to obtain the location distribution marker map of the beef part to be segmented, includes:

[0158] The surface texture feature response map and the spatial distribution feature map are rigidly registered to obtain the texture map and spatial map of the beef part to be segmented;

[0159] Based on the local binary pattern encoding value of the texture map and the bone and meat attribute identifiers in the spatial map, the texture map and the spatial map are serially spliced ​​to obtain the joint feature descriptor of the beef part to be segmented.

[0160] The joint feature descriptor is mapped to the original image coordinate matrix of the source image data to construct the initial multidimensional feature tensor of the beef part to be segmented;

[0161] The initial multidimensional feature tensor is labeled with neighborhood connectivity to obtain the initial aggregated connected component set of the beef part to be segmented;

[0162] The initial aggregated connected component set is subjected to a feature homogeneity check within the region to obtain a pure aggregated connected component set of the beef part to be segmented;

[0163] The cluster centers of the pure aggregated connected domain set in the feature space are counted, and based on the membership distance between the cluster centers and the preset bone category prototype, muscle category prototype and fat category prototype, the pure aggregated connected domain set is assigned a bone category label, muscle category label or fat category label to obtain the part distribution label map of the beef part to be segmented.

[0164] The step of performing regional feature homogeneity verification on the initial aggregated connected component set to obtain a pure aggregated connected component set for the beef part to be segmented includes:

[0165] The joint feature descriptors in the initial aggregated connected component set are evaluated using mean vectors to obtain the feature distribution statistics of the beef part to be segmented.

[0166] Based on the aforementioned feature distribution statistics, the Mahalanobis distance measure is applied to the initial aggregated connected component set to obtain the feature deviation of the beef part to be segmented.

[0167] The feature deviation is compared with a preset deviation tolerance threshold, and the pixels whose feature deviation exceeds the deviation tolerance threshold are assigned as outlier pixels of the beef part to be segmented based on the comparison result.

[0168] Spatial location stripping is performed on the outlier pixels, and spatial weighted interpolation is performed on the empty pixel positions formed after filtering to obtain the optimized pixel set of the beef part to be segmented.

[0169] The connected components of the optimized pixel set are topologically reconstructed to obtain a pure aggregated connected component set of the beef part to be segmented.

[0170] Using the pixel coordinate system of the source image data of the beef part to be segmented as a unified benchmark, a rigid registration operation is performed on the surface texture feature response map and the spatial distribution feature map. Through coordinate transformation by translation and rotation, the pixel coordinates of the two maps are made to correspond completely one-to-one. During the registration process, the pixel ratio and feature information of the maps are kept undistorted. After the registration is completed, the texture map and spatial map of the beef part to be segmented are obtained.

[0171] The local binary pattern encoding value corresponding to each pixel in the texture map of the beef part to be segmented is extracted. At the same time, the bone and meat quality attribute identifiers corresponding to each pixel in the spatial map are extracted. The local binary pattern encoding value and bone and meat quality attribute identifiers at the same pixel position are concatenated in a fixed order. Each pixel position forms a set of independent feature data. The feature data of all pixel positions together constitute the joint feature descriptor of the beef part to be segmented.

[0172] The original image coordinate matrix of the source image data of the beef part to be segmented is retrieved. Each set of feature data in the joint feature descriptor is accurately mapped to the corresponding position in the original image coordinate matrix according to the pixel coordinates, so that each pixel position in the coordinate matrix carries the corresponding joint feature data. Based on the coordinate matrix with feature data, the initial multidimensional feature tensor of the beef part to be segmented is constructed.

[0173] The 4-neighborhood connectivity rule is used to detect and label the connectivity of the pixel neighborhoods of the initial multidimensional feature tensor of the beef part to be segmented. Pixels with similar feature data and adjacent spatial positions are grouped into the same connected region. Each independent connected region is assigned a unique region identifier. All the independent connected regions that have been labeled are integrated to form the initial aggregated connected region set of the beef part to be segmented.

[0174] For each connected region in the initial aggregated connected region set of the beef part to be segmented, feature homogeneity verification is performed within the region. First, the mean and dispersion of the joint feature data of all pixels in the region are calculated. Pixels with dispersion exceeding the preset feature homogeneity threshold are identified as abnormal pixels and removed. Then, pixel completion processing is performed on the region after removing abnormal pixels. All connected regions that have completed verification and optimization are integrated to form a pure aggregated connected region set of the beef part to be segmented.

[0175] Statistical analysis is performed on the joint feature data of each connected region in the pure aggregated connected region set of the beef to be segmented. The cluster center corresponding to each connected region in the feature space is calculated. Preset bone category prototypes, muscle category prototypes and fat category prototypes are retrieved, and the membership distance from each cluster center to the three prototypes is calculated. The connected region is assigned the label corresponding to the category with the smallest membership distance, that is, bone category label, muscle category label or fat category label. All connected regions with completed label assignment are integrated according to coordinates to form the part distribution label map of the beef to be segmented.

[0176] Extract the joint feature descriptors corresponding to all pixels in each connected region of the initial aggregated connected region set of the beef part to be segmented. Calculate the average value of the joint feature descriptors in each connected region according to their dimensions. Combine the average values ​​of each dimension according to the original dimensional order of the joint feature descriptors to form the feature mean vector corresponding to each connected region. The feature mean vectors of all connected regions together constitute the feature distribution statistics of the beef part to be segmented.

[0177] Based on the feature distribution statistics of the beef parts to be segmented, Mahalanobis distance is used to measure the joint feature descriptor corresponding to each pixel in each connected region of the initial aggregated connected region set. This measure measures the degree of feature difference between the joint feature descriptor of each pixel and the feature mean vector of the connected region. The quantification result of this degree of difference is directly assigned as the feature deviation of the corresponding pixel. The feature deviation of all pixels together constitutes the feature deviation of the beef parts to be segmented.

[0178] A pre-set deviation tolerance threshold is retrieved. This threshold is set to a fixed value based on the characteristic distribution of beef bones, muscles, and fat. The feature deviation of each pixel in the beef part to be segmented is compared with the deviation tolerance threshold one by one. Pixels with feature deviation values ​​greater than the deviation tolerance threshold are determined to have penetrated the deviation tolerance threshold and are directly assigned as outlier pixels in the beef part to be segmented.

[0179] Spatial coordinates are located for outlier pixels in the beef segment to be segmented. Pixel information and spatial coordinates of all outlier pixels in the initial aggregated connected component set are directly removed. For the resulting empty pixel positions, spatial weighted interpolation is performed using feature data from effective pixels within a 3×3 neighborhood centered on the empty pixel. Weighting coefficients are assigned based on the distance between neighboring pixels and the empty pixel; the closer the distance, the higher the weighting coefficient. The weighted feature data is calculated and assigned to the empty pixel. After interpolating all empty pixels, the optimized pixel set for the beef segment to be segmented is obtained.

[0180] The optimized pixel set of the beef parts to be segmented is re-detected with 4-neighborhood connectivity. The connected regions are re-divided according to the similarity of feature data and the adjacency of spatial location. Each new independent connected region is assigned a unique region identifier. At the same time, the boundaries of each connected region are topologically regularized to ensure that the region boundaries are continuous and unbroken. After all the identified and regularized connected regions are integrated, a pure aggregated connected region set of the beef parts to be segmented is obtained.

[0181] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0182] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A visual recognition method for bone and meat distribution in beef cutting, characterized in that, The method includes: Pt.

1. Capture the spectral data of the beef part to be segmented to obtain the source image data of the beef part to be segmented; Pt.

2. Local binary mode encoding is performed on the source image data to obtain the surface texture feature response map of the beef part to be segmented; Pt.

3. Perform multi-scale wavelet decomposition on the source image data, and perform gradient magnitude analysis on the decomposed image to obtain the gradient energy tensor of the beef part to be segmented. Pt.

4. Based on the fused gradient energy field constructed by the gradient energy tensor, adaptive threshold segmentation is performed on the meat feature response area and bone feature response area of ​​the beef part to be segmented to obtain the initial distribution label map of the beef part to be segmented. Pt.

5. Based on a preset standard 3D skeleton framework, the perspective projection transformation of the source image data acquisition viewpoint is performed to obtain the skeletal space occupancy prior mask of the beef part to be segmented. Pt.

6. Based on the bone space occupancy prior mask, perform spatial position consistency verification on the bone feature response area in the initial distribution marking map to obtain the spatial distribution feature map of the beef part to be segmented. Pt.

7. The surface texture feature response map and the spatial distribution feature map are concatenated and spliced ​​together, and the concatenated multidimensional feature tensor is aggregated into regions to obtain the part distribution marking map of the beef part to be segmented.

2. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 1, characterized in that, The step of capturing spectral data of the beef portion to be segmented to obtain source image data of the beef portion to be segmented includes: A structured light field is projected onto the surface of the beef section to be cut, wherein the beef section to be cut is a multi-band structured light illumination sequence; Based on the multi-band structured light illumination sequence, the echo beam of the beef part to be segmented is split to obtain a single-wavelength beam of the beef part to be segmented. The single-wavelength beam is photoelectrically converted to obtain the analog electrical signal of the beef part to be segmented; The analog electrical signal is converted from analog to digital to obtain a digital image frame of the beef part to be segmented; The digital image frames are spatially registered according to the acquisition time sequence and angle information to obtain the source image data of the beef part to be segmented.

3. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 1, characterized in that, The step of performing local binary mode encoding on the source image data to obtain the surface texture feature response map of the beef part to be segmented includes: Based on a preset neighborhood sampling topology, the source image data is traversed pixel by pixel to obtain the center pixel of the beef part to be segmented, and the gray value of the center pixel in the neighborhood range is obtained to obtain the neighborhood gray value sequence of the beef part to be segmented. The neighboring gray value sequence is compared with the gray value of the center pixel element by element to obtain the binarized encoding sequence of the beef part to be segmented. Binary weights are assigned to the binarized encoded sequence to obtain the local binary pattern encoded value of the beef part to be segmented; The local binary pattern encoded values ​​are arranged pixel by pixel to construct the local binary pattern encoded map of the beef part to be segmented; Multi-channel fusion of the local binary pattern coding map is performed to obtain the surface texture feature response map of the beef part to be segmented.

4. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 1, characterized in that, The step of performing multi-scale wavelet decomposition on the source image data and analyzing the gradient magnitude of the decomposed image to obtain the gradient energy tensor of the beef segment to be segmented includes: The source image data is subjected to wavelet decomposition filtering to obtain the original wavelet sub-band image sequence of the beef part to be segmented; Based on the spatial orientation of the beef part to be segmented, the original wavelet sub-band image sequence is directionally separated to obtain the horizontal high-frequency sub-band image, the vertical high-frequency sub-band image, and the diagonal high-frequency sub-band image of the beef part to be segmented. The high-frequency sub-band images in the horizontal direction, the vertical direction, and the diagonal direction are analyzed for directional gradient within the pixel neighborhood to obtain the high-frequency gradient response map of the beef part to be segmented. Energy accumulation is performed on the high-frequency gradient response map to obtain a single-scale gradient energy map of the beef part to be segmented; The single-scale gradient energy map is stacked to construct the gradient energy tensor of the beef part to be segmented.

5. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 1, characterized in that, The fused gradient energy field constructed based on the gradient energy tensor is used to adaptively threshold the meat feature response region and bone feature response region of the beef part to be segmented, resulting in an initial distribution map of the beef part to be segmented, including: Spatial coordinate registration is performed on the single-scale gradient energy map in the gradient energy tensor, and the scale is normalized to obtain the multi-scale gradient energy map sequence of the beef part to be segmented. The gradient energy response amplitude of the pixel coordinates in the multi-scale gradient energy map sequence is extracted, and the real-time gradient energy response amplitude is accumulated and fused to construct the fused gradient energy field of the beef part to be segmented. By performing distribution histogram statistics on the fused gradient energy field, the trough locations and inflection point abrupt changes of the beef portion to be segmented are obtained; Based on the trough location and the inflection point abrupt change location, the energy amplitude segmentation boundary is defined to obtain the adaptive segmentation threshold of the beef part to be segmented; Based on the adaptive segmentation threshold, the category attribute of the fused gradient energy field is determined, and the pixels that penetrate the adaptive segmentation threshold are assigned as the skeletal feature response points of the beef part to be segmented, and the pixels that do not penetrate the adaptive segmentation threshold are assigned as the meat quality feature response points of the beef part to be segmented. Spatial assignment and filling are performed on the skeletal feature response points and the meat quality feature response points to obtain the initial distribution marking map of the beef part to be segmented.

6. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 1, characterized in that, The method, based on a preset standard skeletal 3D framework, performs perspective projection transformation on the acquisition perspective of the source image data to obtain a priori mask for the skeletal space occupancy of the beef part to be segmented, including: Spatial coordinate analysis is performed on the preset standard 3D skeleton framework to obtain the standard 3D point cloud framework of the beef part to be segmented. Based on the optical center position parameters and focal plane orientation parameters of the source image data, construct the perspective projection matrix of the beef part to be segmented; Based on the perspective projection matrix, the vertex coordinates in the standard skeletal 3D point cloud framework are projected into spatial coordinates to obtain the 2D skeletal contour point set of the beef part to be segmented. Based on the topological connection relationship between the vertex coordinates, the adjacent contour points in the two-dimensional skeleton contour point set are closed and filled to obtain the two-dimensional continuous contour surface of the beef part to be segmented. The two-dimensional continuous contour surface is mapped onto a blank canvas of the same size as the source image data, and the pixel positions of the two-dimensional continuous contour surface are assigned a state to obtain the skeletal space occupancy prior mask of the beef part to be segmented.

7. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 1, characterized in that, The step of performing spatial position consistency verification on the skeletal feature response regions in the initial distribution marker map based on the skeletal spatial occupancy prior mask to obtain the spatial distribution feature map of the beef part to be segmented includes: The connected regions of the skeletal feature response points in the initial distribution map are separated and analyzed to obtain the initial set of skeletal response connected regions of the beef part to be segmented; Boundary contour tracing is performed on the bone space occupancy prior mask to obtain the bone space boundary of the beef part to be segmented; Based on the spatial boundary of the bone occupancy space, the initial bone response connected domain set is spatially overlapped and compared to obtain the spatial consistency determination result of the beef part to be segmented. Based on the spatial consistency determination result, redundant filtering is performed on the connected components located outside the boundary of the bone-occupying space in the initial distribution marking map to obtain the bone response region of the beef part to be segmented. The skeletal response region is fused with the meat quality feature response region in the initial distribution marker map to obtain the spatial distribution feature map of the beef part to be segmented.

8. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 7, characterized in that, The formula for calculating the spatial overlap confidence in the spatial consistency determination result is as follows: ； In the formula, The confidence level for the spatial overlap. The overlapping area is the region between the connected components in the initial skeleton response connected component set and the boundary of the skeleton's occupied space. This represents the total area of ​​the internal region of the boundary of the space occupied by the skeleton. The shortest Euclidean distance from the centroid of the connected domain to the boundary of the bone's occupied space is given by denoted as ... The preset spatial scale normalization factor, For the pixels in the overlapping region The gradient energy response amplitude in the fused gradient energy field This is the adaptive segmentation threshold for the beef portion to be segmented. The total number of pixels in the overlapping region. The preset spatial overlap weighting coefficient, The preset distance attenuation weighting coefficient, The preset energy consistency weighting coefficient, This refers to the overlapping region between the connected components in the initial skeleton response connected component set and the boundary of the skeleton's occupied space.

9. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 1, characterized in that, The process of concatenating and splicing the surface texture feature response map and the spatial distribution feature map into feature vectors, and then performing region aggregation on the concatenated multidimensional feature tensor to obtain the location distribution marker map of the beef part to be segmented, includes: The surface texture feature response map and the spatial distribution feature map are rigidly registered to obtain the texture map and spatial map of the beef part to be segmented; Based on the local binary pattern encoding value of the texture map and the bone and meat attribute identifiers in the spatial map, the texture map and the spatial map are serially spliced ​​to obtain the joint feature descriptor of the beef part to be segmented. The joint feature descriptor is mapped to the original image coordinate matrix of the source image data to construct the initial multidimensional feature tensor of the beef part to be segmented; The initial multidimensional feature tensor is labeled with neighborhood connectivity to obtain the initial aggregated connected component set of the beef part to be segmented; The initial aggregated connected component set is subjected to a feature homogeneity check within the region to obtain a pure aggregated connected component set of the beef part to be segmented; The cluster centers of the pure aggregated connected domain set in the feature space are counted, and based on the membership distance between the cluster centers and the preset bone category prototype, muscle category prototype and fat category prototype, the pure aggregated connected domain set is assigned a bone category label, muscle category label or fat category label to obtain the part distribution label map of the beef part to be segmented.

10. The method for visual recognition of bone and meat distribution in beef segmentation as described in claim 9, characterized in that, The step of performing regional feature homogeneity verification on the initial aggregated connected component set to obtain a pure aggregated connected component set for the beef part to be segmented includes: The joint feature descriptors in the initial aggregated connected component set are evaluated using mean vectors to obtain the feature distribution statistics of the beef part to be segmented. Based on the aforementioned feature distribution statistics, the Mahalanobis distance measure is applied to the initial aggregated connected component set to obtain the feature deviation of the beef part to be segmented. The feature deviation is compared with a preset deviation tolerance threshold, and the pixels whose feature deviation exceeds the deviation tolerance threshold are assigned as outlier pixels of the beef part to be segmented based on the comparison result. Spatial location stripping is performed on the outlier pixels, and spatial weighted interpolation is performed on the empty pixel positions formed after filtering to obtain the optimized pixel set of the beef part to be segmented. The connected components of the optimized pixel set are topologically reconstructed to obtain a pure aggregated connected component set of the beef part to be segmented.

Images