Cattle face key morphological parameter measuring system and method based on multi-view images

Abstract

This invention relates to the field of morphological measurement technology, specifically to a system and method for measuring key morphological parameters of a cow's face based on multi-view images. The system includes a data synchronization construction module, a region segmentation and discrimination module, a distortion correction construction module, a pose prediction and fusion module, and a structural parameter measurement module. In this invention, multi-view image data is combined with timestamps, vibration frequency values, and sound wave reflection information for parameter alignment, achieving synchronous integration of image information and external physical states. Confidence is dynamically assessed based on image channels and structural region grayscale values ​​to identify abnormal frames. Image spatial accuracy compensation is achieved by combining pixel offset correction and structural boundary mapping. Spatial direction vectors are constructed using the polarization angle values ​​of the cow's eye region and superimposed with historical behavior to predict pose trajectories. Through multi-view measurement, the stability and accuracy of cow face morphological parameter measurement are effectively improved under conditions of image complexity, pose variation, and structural occlusion.

Description

Technical Field

[0001] This invention relates to the field of morphological measurement technology, and in particular to a system and method for measuring key morphological parameters of a cow's face based on multi-view images. Background Technology

[0002] Morphological measurement technology involves the identification, measurement, and analysis of the external morphological features of organisms. It primarily utilizes optical imaging, image processing, and 3D modeling to extract the geometric structure and dimensional parameters of specific parts of a target object. It is widely applied in biometrics, animal health assessment, breed selection, and behavioral monitoring, and has particularly high application value and research significance in animal individual identification and structural parameter modeling. Traditional bovine facial key morphological parameter measurement systems refer to systems that identify features and extract morphological parameters from bovine facial structures. These systems are mainly used to obtain key geometric indicators such as nasal bridge width, interocular distance, and forehead curvature. They typically employ single-angle image acquisition for measurement, relying heavily on edge extraction and fixed-point measurement of two-dimensional images. Key point positions are obtained through manual annotation or simple image segmentation methods, and then the corresponding length or angle parameters are determined based on geometric calculations. However, limitations imposed by image angle, lighting variations, and occlusion issues lead to unstable identification points and significant parameter errors, making it difficult to comprehensively and accurately reflect the true three-dimensional morphological features of the bovine face.

[0003] Traditional systems for measuring key morphological parameters of cow faces rely on single-angle images for structural feature extraction. Due to limitations in camera angle, the acquired images cannot fully represent the three-dimensional structure of the cow's face. Edge extraction and fixed-point measurement methods for two-dimensional images are highly sensitive to image quality. Under conditions of lighting changes or partial occlusion, the identification position of key points is unstable, and the measurement error increases significantly. Manual annotation and simple image segmentation methods suffer from strong subjectivity and poor consistency, resulting in insufficient accuracy in measuring key structural parameters. In particular, they cannot reliably reflect the true morphological features of the cow's face in complex environments. Summary of the Invention

[0004] To address the technical problems existing in the prior art, embodiments of the present invention provide a system and method for determining key morphological parameters of a cow's face based on multi-view images. The technical solution is as follows: On the one hand, a system for determining key morphological parameters of a cow's face based on multi-view images is provided. This system includes: The data synchronization module obtains the image frame number, image acquisition timestamp, and image channel data from the cattle channel camera, and performs matching and alignment with the triaxial vibration frequency value of the inertial sensor and the sound wave reflection time value of the ultrasonic array to generate a multi-parameter alignment relationship group. The region segmentation and discrimination module calculates the confidence level of the cow face structure region in the image based on the image channel data in the multi-parameter alignment relationship group, marks abnormal path image frames, and outputs path type labels; The distortion correction construction module extracts the abnormal path image frame number from the path type label, calculates the offset displacement value of each pixel in the vibration direction, and constructs the image structure boundary displacement range map by combining the sound wave reflection time value, and outputs the pixel offset correction image group. The attitude prediction fusion module collects the image coordinate information of the bull's eye region in the pixel offset correction image group, extracts the polarization angle value of the bull's eye region in each frame image, calculates the spatial direction distribution vector of the bull's eye, and performs temporal interval superposition analysis in combination with historical behavior to generate a bull's eye attitude trajectory prediction sequence. The structural parameter determination module extracts the image coordinates of the nose tip, the corners of both eyes and the maxilla based on the predicted sequence of the bovine eye posture trajectory, summarizes the distribution relationship of key structural points in the multi-view image frames, and outputs the key morphological parameter identification record of the bovine face in the multi-view image.

[0005] As a further aspect of the present invention, the abnormal path image frame specifically refers to an image frame in which the confidence level of the cow face structure region is greater than a set dynamic confidence threshold.

[0006] As a further aspect of the present invention, the multi-parameter alignment relationship group includes image frame number matching relationship, image acquisition timestamp mapping relationship, vibration frequency value and image frame comparison table, and sound wave reflection time value and image frame comparison table. The path type label includes normal image path label, abnormal image path label, and low confidence path label. The pixel offset correction image group includes pixel vector remapping map, structural boundary displacement range map, and vibration direction offset correction map. The bull's eye posture trajectory prediction sequence includes bull's eye polarization angle change curve, direction distribution vector time series map, and behavior trajectory reconstruction map. The key morphological parameter identification record of the bull's face includes nasal tip point spatial position parameter, eye corner point relative distance parameter, and maxillary point posture angle parameter.

[0007] As a further aspect of the present invention, the data synchronization construction module includes: The data stream receiving submodule collects the image streams from the visible light cameras on both sides of the cattle channel, including the image frame number, image acquisition timestamp, and image channel data. It also collects the triaxial vibration frequency values ​​output by the inertial sensor and the sound wave reflection time values ​​recorded by the ultrasonic array within the time period corresponding to the image timestamp. Data pairs with time differences exceeding the maximum interval threshold between image frames are removed, and image frames and sensing time series sets are generated. The multi-source data pairing submodule extracts data groups based on the same timestamp according to the image frame number, triaxial vibration frequency value and sound wave reflection time value in the image frame and sensing time series set. It performs interval averaging on the vibration frequency value and sound wave reflection time value to construct the binding relationship between the image frame number and the sensor parameters of the time period, and generates the image frame and multi-source sensing data binding set. The parameter alignment generation submodule establishes an index mapping for the image frame number based on the image frame number and the bound vibration frequency value and sound wave reflection time value in the image frame and multi-source sensing data binding set, establishes a matching relationship between the image frame and the parameters, and generates a multi-parameter alignment relationship group.

[0008] As a further aspect of the present invention, the region segmentation and discrimination module includes: The image grayscale extraction submodule extracts the pixel matrix of the red, green, blue and infrared channels in each frame image based on the image channel data in the multi-parameter alignment relationship group, performs channel normalization processing on the four-channel pixel matrix, and generates an image channel grayscale value matrix set. The confidence calculation submodule locates the coordinate range corresponding to the cow face structure region in the image based on the gray value of each frame in the image channel gray value matrix set, calculates the confidence based on the gray value distribution of the region's red, green, blue and infrared channels, and simultaneously calculates the average confidence of the cow face region in historical image frames and constructs a dynamic confidence threshold to generate structural region confidence data. The path type labeling submodule determines whether the confidence of the current frame is higher than the dynamic confidence threshold based on the confidence data of the structural region. If the condition is met, the image frame is labeled as an abnormal path image frame, and the corresponding label number and discrimination result under the abnormal state are output to generate a path type label.

[0009] As a further aspect of the present invention, the distortion correction construction module includes: The principal point angle construction submodule extracts the image position coordinates of all pixels in each image based on the image frame number marked as an abnormal path in the path type label, collects the coordinates of the image center point as the principal point position, calculates the vector angle value between each pixel and the principal point, and indexes and binds the image frame number with the corresponding angle data to generate a pixel angle relationship matrix. The vibration offset calculation submodule reads the angle data in the pixel angle relationship matrix and the vibration frequency value in the multi-parameter alignment relationship group, constructs a projection vector relationship model based on the angle between the pixel vector direction and the vibration direction, calculates the unit displacement value and total displacement value of each pixel in the vibration direction, establishes a displacement data structure according to the image frame number, and generates a vibration direction offset displacement set. The pixel coordinate remapping submodule performs coordinate remapping between the original pixel position and the direction of the offset value within the image frame based on the offset value of each pixel in the vibration direction offset displacement set. At the same time, it constructs a response delay region map of the pixel boundary structure by combining the sound wave reflection time value, and superimposes the offset mapping and the boundary region range to generate a pixel offset correction image group.

[0010] As a further aspect of the present invention, the attitude prediction fusion module includes: The polarization angle extraction submodule, based on the pixel offset correction image group, collects the image coordinate information of the bullseye region in the image, obtains the polarization channel value of the corresponding pixel position, filters the polarization angle data of each pixel in the bullseye region, and constructs a mapping index between the frame number and the polarization angle of the bullseye region to generate a set of polarization angle values ​​for the bullseye region. The direction vector construction submodule is based on the polarization angle distribution in each frame of the bullseye region polarization angle value set. According to the directionality definition corresponding to the polarization angle, it performs vector superposition operation on the polarization angle values ​​of the bullseye region in each frame image, calculates the spatial direction main vector corresponding to the current frame, establishes the index relationship between the frame number and the direction vector, and generates a spatial direction distribution vector sequence. The trajectory sequence generation submodule reads the individual cow number corresponding to the image frame number based on the direction information corresponding to each frame in the spatial direction distribution vector sequence, and obtains the direction record interval corresponding to the number in the historical behavior database. It then performs time-series matching between the current spatial direction vector sequence and the historical direction interval to generate a cow eye posture trajectory prediction sequence.

[0011] As a further aspect of the present invention, the structural parameter measurement module includes: The structural point extraction submodule collects the image coordinates of the nasal tip, the corners of the eyes and the maxillary point in the corresponding image based on the frame number and image position index of each image in the bovine eye posture trajectory prediction sequence. It then sets a coordinate template for key points in the bovine face structure anatomical region and summarizes the structural point coordinate information according to the image frame number to generate a set of key structural point coordinates. The spatial difference recognition submodule extracts the position of structural points in each frame of the key structural point coordinate set, calculates the Euclidean distance difference between the coordinate distances from the tip of the nose to the corners of the eyes and from the corners of the eyes to the maxilla in any two image frames, constructs the relative position difference matrix of the key structural points, and generates a spatial difference distribution map of structural points. The morphological parameter generation submodule classifies the spatial relationship of key points under the same viewpoint sequence image frame number based on the spatial difference data between each pair of structural points in the spatial difference distribution map of the structural points, calculates the relative offset values ​​of structural points in the horizontal, vertical and radial directions in the multi-view images, constructs the morphological measurement index of key regions, and generates key morphological parameter identification records for the cow face.

[0012] A method for determining key morphological parameters of a cow's face based on multi-view images, wherein the method is performed based on the aforementioned system for determining key morphological parameters of a cow's face based on multi-view images, and includes the following steps: S1: Obtain the image frame number, image acquisition timestamp, and image channel data from the cattle channel camera, and match and align them with the triaxial vibration frequency value of the inertial sensor and the sound wave reflection time value of the ultrasonic array to generate a multi-parameter alignment relationship group; S2: Based on the image channel data in the multi-parameter alignment relationship group, calculate the confidence of the cow face structure region in the image, mark the abnormal path image frames, and output the path type label; S3: Extract the abnormal path image frame number from the path type label, calculate the offset displacement value of each pixel in the vibration direction, combine the sound wave reflection time value to construct the image structure boundary displacement range map, and output the pixel offset correction image group. S4: Collect the image coordinate information of the bull's eye region in the pixel offset correction image group, extract the polarization angle value of the bull's eye region in each frame image, calculate the spatial direction distribution vector of the bull's eye, combine it with historical behavior to perform time interval superposition analysis, and generate a bull's eye posture trajectory prediction sequence. S5: Based on the predicted sequence of bovine eye posture trajectory, extract the image coordinates of the nasal tip, the corners of both eyes and the maxillary point, summarize the distribution relationship of key structural points in the multi-view image frames, and output the key morphological parameter identification record of the bovine face in the multi-view image.

[0013] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By introducing multi-view image data and combining it with timestamps, vibration frequency values, and sound wave reflection information for parameter alignment, the synchronous integration of image information and external physical state is achieved. Confidence is dynamically evaluated and abnormal frames are identified based on the grayscale values ​​of image channels and structural regions. Image spatial accuracy compensation is completed by combining pixel offset correction and structural boundary mapping. Spatial direction vectors are constructed using the polarization angle values ​​of the bull's eye region and superimposed with historical behavior to achieve posture trajectory prediction. Spatial relationship parameters of the nasal tip, eye corners, and maxillary point are extracted through the distribution analysis of key structural points in multi-view images. This effectively improves the stability and accuracy of bull face morphological parameter measurement under conditions of image complexity, posture changes, and structural occlusion. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1This is a schematic diagram of the bovine face key morphological parameter measurement system based on multi-view images provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the data synchronization construction module of the present invention; Figure 4 This is a flowchart of the region segmentation and discrimination module of the present invention; Figure 5 This is a flowchart of the distortion correction construction module of the present invention; Figure 6 This is a flowchart of the attitude prediction and fusion module of the present invention; Figure 7 This is a flowchart of the structural parameter measurement module of the present invention; Figure 8 This is a flowchart of a method for determining key morphological parameters of a cow's face based on multi-view images, provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0018] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0019] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] This invention provides a system for determining key morphological parameters of a cow's face based on multi-view images, such as... Figure 1-2 The system diagram shown includes: The data synchronization construction module obtains the image frame number, image acquisition timestamp and image channel data of the visible light camera on both sides of the cattle passage. At the same time, it obtains the three-axis vibration frequency value output by the inertial sensor and the sound wave reflection time value recorded by the ultrasonic array. It matches and aligns the image frame with the vibration frequency value and the sound wave reflection time value to generate a multi-parameter alignment relationship group. The region segmentation and discrimination module extracts the grayscale values ​​of the red, green, blue and infrared channels in each frame image based on the image channel data in the multi-parameter alignment relationship group, calculates the confidence of the cow face structure region in the image, compares whether the confidence of the current frame region is higher than the set confidence dynamic threshold (determined by the average confidence of the cow face structure region in historical frames), and if it is higher, it is marked as an abnormal path image frame and outputs the path type label; The distortion correction construction module extracts the abnormal path image frame number from the path type label, establishes a principal point vector angle relationship model for each image pixel, calculates the offset displacement value of each pixel in the vibration direction, remaps the offset pixel coordinates to the original image position, constructs an image structure boundary displacement range map by combining the sound wave reflection time value, and outputs a pixel offset correction image group. The attitude prediction fusion module collects the image coordinate information of the bull's eye region in the pixel offset correction image group, extracts the polarization angle value of the bull's eye region in each frame image, calculates the spatial direction distribution vector of the bull's eye, reads the historical behavior database corresponding to the individual cattle number, performs temporal interval superposition analysis with the current direction distribution vector, and generates a bull's eye attitude trajectory prediction sequence. The structural parameter determination module extracts the image coordinates of the nasal tip, the corners of both eyes and the maxilla based on the predicted sequence of the bovine eye posture trajectory, identifies the relative spatial differences of key structural points in different image frames, summarizes the distribution relationship of key structural points in multi-view image frames, and outputs the key morphological parameter identification record of the bovine face in multi-view images.

[0022] The multi-parameter alignment relationship group includes image frame number matching relationship, image acquisition timestamp mapping relationship, vibration frequency value and image frame comparison table, and sound wave reflection time value and image frame comparison table. The path type label includes normal image path label, abnormal image path label, and low confidence path label. The pixel offset correction image group includes pixel vector remapping map, structural boundary displacement range map, and vibration direction offset correction map. The bull's eye posture trajectory prediction sequence includes bull's eye polarization angle change curve, direction distribution vector time series map, and behavior trajectory reconstruction map. The key morphological parameter identification record of the bull's face includes the spatial position parameter of the nasal tip, the relative distance parameter between the two eye corners, and the posture angle parameter of the maxillary point.

[0023] Specifically, such as Figure 2 , 3 As shown, the data synchronization building module includes: The data stream receiving submodule collects the image streams from the visible light cameras on both sides of the cattle channel, including the image frame number, image acquisition timestamp, and image channel data. It also collects the triaxial vibration frequency values ​​output by the inertial sensor and the sound wave reflection time values ​​recorded by the ultrasonic array within the time period corresponding to the image timestamp. Data pairs with time differences exceeding the maximum interval threshold between image frames are removed, and image frames and sensing time series sets are generated. The data stream receiving submodule acquires image streams from visible light cameras on both sides of the cattle passage. First, a basic signal pattern recognition mechanism is introduced, and the acquisition parameters of the visible light cameras are configured, setting the resolution to [resolution value missing]. The pixel count and frame rate were set to 60fps, corresponding to an inter-frame time interval of 16.67ms. Simultaneously, the inertial measurement unit (IMU) deployed on the tunnel floor and the ultrasonic array on the side wall were activated. The sampling frequency of the inertial sensor was set to 200Hz, i.e., a sampling interval of 5ms, and the sound wave emission frequency of the ultrasonic array was set to 40kHz. Subsequently, clock synchronization calibration was performed on each data source, using the Network Time Protocol (NTP) to control the time error of all acquisition devices to within 1ms.

[0024] During the acquisition process, the camera outputs a continuous stream of image frames, and the frame number of each frame is extracted sequentially. (e.g., frame number #10245) and image acquisition timestamp (For example ms). Meanwhile, read from the buffer within the time interval. All sensor data within, of which It is set to half of the image frame interval, i.e. ms. Within this time window, extract the triaxial vibration frequency values ​​from the inertial sensor data stream. And extract the sound wave reflection time value from the ultrasonic array data stream. For each set of sensor data, calculate its recording timestamp. Image timestamp absolute difference .

[0025] Based on time-series matching pattern recognition logic, data filtering logic is executed, and if the calculated... If the time interval is less than ms, the sensor data set is considered asynchronous and discarded; if If the time difference is ms, then the data pair is retained. As shown in Table 1, the original data acquisition situation under a specific frame number is recorded. For frame #10245, two sets of sensor data with a time difference of 2ms and 4ms are retained, and a set of data with a time difference of 12ms is removed. Finally, the filtered valid data are rearranged in chronological order to generate image frames and sensor time series sets.

[0026] Table 1. Raw Data Acquisition Table of Sensors and Image Frames The multi-source data pairing submodule extracts data groups based on the same timestamp, according to the image frame number, triaxial vibration frequency value and sound wave reflection time value in the image frame and sensing time series set. It performs interval averaging on the vibration frequency value and sound wave reflection time value, constructs the binding relationship between the image frame number and the sensor parameters of the time period, and generates the image frame and multi-source sensing data binding set. The multi-source data pairing submodule first iterates through each independent image frame number in the sequence set based on the image frame number, triaxial vibration frequency value, and sound wave reflection time value in the image frame and sensor time series set. It then aggregates multiple sets of sensor data retained under the same frame number, using a statistical feature pattern recognition algorithm. Taking frame number #10245 in Table 1 as an example, this frame corresponds to two sets of valid sensor data: vibration frequencies of 45.2Hz and 45.5Hz, and sound wave reflection times of 1250µs and 1255µs. The arithmetic mean calculation logic is then invoked to perform interval averaging on the parameters within this time period, calculating the average vibration frequency value. and mean sound wave reflection time .

[0027] Substituting the numerical values, the average vibration frequency corresponding to frame #10245 is calculated as follows: Hz, average sound wave reflection time is us. Then, a data binding structure is constructed, linking the calculated average parameters with the corresponding image frame numbers. Perform one-to-one association. For frames in the sequence that contain only a single set of sensor data (such as frame #10246), directly use this set of data as the binding parameter without averaging. Perform the above operation on all frames in sequence to ensure that each image frame has a unique sensor parameter description, eliminate one-to-many data redundancy caused by inconsistent sampling frequencies, and finally store all image frame data bound to unique environmental parameters into a structured database to generate an image frame and multi-source sensing data binding set.

[0028] The parameter alignment generation submodule establishes an index mapping for the image frame numbers and the bound vibration frequency value and sound wave reflection time value in the image frame and multi-source sensing data binding set, establishes the matching relationship between the image frames and the parameters, and generates a multi-parameter alignment relationship group. The parameter alignment generation submodule, based on the image frame number and the bound vibration frequency and sound wave reflection time values ​​in the image frame and multi-source sensing data binding set, first initializes a hash index table (HashMap), setting the image frame number as the key and the corresponding average vibration frequency value, sound wave reflection time value, and original image data storage address as values. It then performs a traversal operation, reading each record in the binding set; for example, reading frame #10245, extracting its bound vibration frequency of 45.35Hz and sound wave reflection time of 1252.5us, and writing them to the corresponding positions in the index table.

[0029] Subsequently, a temporal matching relationship between parameters is established, and a pattern recognition strategy for sequence anomaly detection is used to check the continuity of parameter changes between adjacent frame numbers in the index table. If the rate of change of vibration frequency between adjacent frames (such as #10245 and #10246) exceeds a preset abrupt change threshold (e.g., set to 10%), the index at that point is marked for subsequent verification. Here, the rate of change between #10245 (45.35Hz) and #10246 (46.1Hz) is... If the percentage is below the 10% threshold, it conforms to the continuous pattern of normal environmental changes and is determined to be a normal continuous sequence. After the indexing of all frames is completed, a metadata file containing the complete retrieval path and parameter mapping relationship is output to ensure that subsequent modules can quickly call the corresponding environmental parameters through the frame number and generate multi-parameter alignment relationship groups.

[0030] Specifically, such as Figure 2 , 4 As shown, the region segmentation and discrimination module includes: The image grayscale extraction submodule extracts the pixel matrix of the red, green, blue and infrared channels in each frame image based on the image channel data in the multi-parameter alignment relationship group, performs channel normalization processing on the four-channel pixel matrix, and generates an image channel grayscale value matrix set. The image grayscale extraction submodule, based on the image channel data in the multi-parameter alignment group, first reads the raw RAW data of each frame and separates it into four independent single-channel pixel matrices: red (R), green (G), blue (B), and near-infrared (NIR). Each matrix has a dimension of... The pixel bit depth is 8 bits, meaning the value range is... Max-min normalization is performed on each of the four channel matrices to establish a unified feature space for subsequent grayscale pattern recognition.

[0031] Calling formula ,in The current pixel value. and These are the minimum and maximum values ​​of all pixels in that channel. For example, in the red channel, a certain pixel... The minimum value of the entire frame is 10, and the maximum value is 240. Therefore, the normalized grayscale value is... This operation is performed on each pixel in the R, G, B, and NIR channels one by one, yielding four values ​​within a certain range. The floating-point grayscale value matrix between the two is used to eliminate the influence of light intensity differences on the image numerical distribution, ensuring that images acquired at different time periods have a unified numerical benchmark. Finally, the four normalized matrices are merged and stored to generate an image channel grayscale value matrix set.

[0032] The confidence calculation submodule locates the coordinate range corresponding to the cow face structure region in the image based on the gray value of each frame in the image channel gray value matrix set, calculates the confidence based on the gray value distribution of the region's red, green, blue and infrared channels, and simultaneously calculates the average confidence of the cow face region in historical image frames and constructs a dynamic confidence threshold to generate structural region confidence data. The confidence calculation submodule first loads a preset cow face structure region coordinate template based on the gray values ​​of each frame in the image channel gray value matrix set. This template is based on the general proportions of a cow's face and covers the central region of the image. Pixel range. Extract the normalized grayscale matrix of the four channels (R, G, B, NIR) within this region, and calculate the mean grayscale value of each channel within the region. With variance (in ).

[0033] Based on the principle of spatial feature pattern recognition, and in conjunction with preset channel weight coefficients (set... Calculate the confidence score of the cow face region in the current frame. The calculation formula is: ,in The standard grayscale average value for a cow's face is set to 0.5. If a certain frame region... ,but Simultaneously, the confidence scores of the previous 500 frames are read, and their arithmetic mean is calculated as the dynamic confidence threshold. Assuming the historical average is 0.85, the calculated current frame confidence score of 0.947 is associated with and stored with the dynamic threshold to generate structural region confidence data.

[0034] The path type labeling submodule determines whether the confidence of the current frame is higher than the dynamic confidence threshold based on the structural region confidence data. If the condition is met, the image frame is labeled as an abnormal path image frame, and the corresponding label number and discrimination result under the abnormal state are output to generate path type labels. The path type tagging submodule, based on the structural region confidence data and combined with behavioral pattern recognition technology, first reads the confidence score of the current frame. With dynamic confidence threshold Execute the comparison and judgment logic. The judgment condition is set as follows: if... ,in The floating tolerance coefficient is set to 0.05, which means the judgment threshold is... .

[0035] Comparative findings If the confidence level meets the high confidence condition, it indicates that the image clearly captures the cow's face structure and is located in the expected area, thus determining it as a forward path or a valid capture. Conversely, if the confidence level is below the threshold, it may be a lateral or backward path. In this embodiment, the focus is on screening frames that meet the abnormally high confidence level (representing abnormal behavior such as cows standing still or approaching the camera directly). By identifying this specific behavior pattern, if the judgment result is true, the frame is marked as an abnormal path image frame, assigned a label number \(L_{type}=1\) (0 represents normal, 1 represents abnormal), and the judgment result is written to the status register. For frame \#10245, the label number 1 and the judgment description of abnormal - facing directly are output, generating a path type label.

[0036] Specifically, such as Figure 2 , 5 As shown, the distortion correction construction module includes: The principal point angle construction submodule extracts the image position coordinates of all pixels in each image based on the image frame number marked as an abnormal path in the path type label, collects the coordinates of the image center point as the principal point position, calculates the vector angle value between each pixel and the principal point, and indexes and binds the image frame number with the corresponding angle data to generate a pixel angle relationship matrix. The principal point angle construction submodule, based on the image frame number marked as an abnormal path in the path type label, first filters out the image frame with label number 1 (e.g., frame #10245), and extracts that frame. All pixel coordinates at resolution Define the coordinates of the principal point at the center of the image as follows: .

[0037] Calculate its spatial position vector relative to the principal point pixel by pixel. It also uses inverse trigonometric functions to calculate the angle between each pixel vector and the horizontal axis. Construct the basic geometric distribution pattern. The calculation formula is as follows: The results are expressed in radians and range from 1 to 10. For example, for pixels ,calculate ,but (i.e., 45 degrees); for pixels , ,but (i.e., 135 degrees). Traverse 3,686,400 pixels in the image to generate an angle matrix with the same dimensions as the original image, and bind this matrix to the frame number #10245 to generate a pixel angle relationship matrix.

[0038] The vibration offset calculation submodule reads the angle data in the pixel angle relationship matrix and the vibration frequency value in the multi-parameter alignment relationship group. Based on the angle between the pixel vector direction and the vibration direction, it constructs a projection vector relationship model, calculates the unit displacement value and total displacement value of each pixel in the vibration direction, establishes a displacement data structure according to the image frame number, and generates a vibration direction offset displacement set. The vibration offset calculation submodule reads the angle data from the pixel angle relationship matrix and the vibration frequency value from the multi-parameter alignment relationship group. First, it obtains the vibration frequency corresponding to frame #10245. Hz, set the vibration direction to vertical (i.e., 90 degrees or Based on the projection vector relationship model, a physical vibration disturbance mode is simulated, and the displacement of each pixel under the influence of vibration is calculated. A unit vibration influence coefficient is defined. Pixels / Hz, calculate the total displacement of each pixel in the vibration direction. The calculation formula is: .

[0039] A sine function is used here because the vibration direction is vertical, and the pixel vector needs to be projected onto the vertical axis. For example, for the aforementioned pixel... (angle Its offset value is calculated as follows: Pixels; for pixels located on the horizontal axis (angle 0), the offset value is 0. Traverse all angle data in the matrix, calculate the corresponding offset value, and establish a displacement data structure according to pixel coordinates, as shown in Table 2. Table 2 records the calculation results for some typical pixels, generating a set of vibration direction offset displacements. Table 2: Pixel Vibration Offset Calculation Results (Partial) The pixel coordinate remapping submodule performs coordinate remapping between the original pixel position and the direction of the offset value in the image frame based on the offset value of each pixel in the vibration direction offset displacement set. At the same time, it constructs a response delay region map of the pixel boundary structure by combining the sound wave reflection time value, and superimposes the offset mapping and the boundary region range to generate a pixel offset correction image group. The pixel coordinate remapping submodule first creates a blank target image matrix based on the offset values ​​of each pixel in the vibration direction offset displacement set. For each pixel in the original image... Read its corresponding offset value Calculate the corrected target coordinates Because vibration causes pixel displacement in the vertical direction, the correction process is a reverse operation, i.e. , In pixels For example, if its offset value is 1.60 pixels, then the remapping coordinates are... The floating-point coordinates are mapped back to an integer pixel grid using bilinear interpolation.

[0040] At the same time, combined with the sound wave reflection time value us constructs a response latency region map based on the speed of sound. m / s, calculate the distance the sound wave travels Meters. Based on camera parameters, this physical distance is mapped to a ring-shaped boundary region on the image (e.g., radius 1). (pixels), applying region validity pattern recognition logic, overlaying the offset-corrected image with the response delay region map, retaining only the valid pixels within the region, removing invalid data with severe edge distortion, and generating a pixel offset-corrected image group.

[0041] Specifically, such as Figure 2 , 6 As shown, the pose prediction fusion module includes: The polarization angle extraction submodule is based on the pixel offset correction image group. It collects the image coordinate information of the bull's eye region in the image, obtains the polarization channel value of the corresponding pixel position, filters the polarization angle data of each pixel in the bull's eye region, and constructs a mapping index between the frame number and the polarization angle of the bull's eye region to generate a set of polarization angle values ​​for the bull's eye region. The polarization angle extraction submodule, based on pixel offset corrected image groups, utilizes an image feature pattern recognition algorithm to first locate the bullseye region in the corrected image and obtain the center coordinates of the left eye region. and the center coordinates of the right eye region Define the region of interest (ROI) as a circular area with a radius of 50 pixels, centered at the eye center. Extract the polarization channel (P channel) values ​​of all pixels within this region. According to the principle of polarization imaging, there is a mapping relationship between the polarization value and the normal angle of the object's surface.

[0042] Calling the polarization angle analysis function ,in For Stokes vector components, in this embodiment, it is simplified to directly reading the preprocessed polarization angle data. Assuming a pixel's polarization value is 120°, the corresponding polarization angle is resolved as follows: Pixels with a polarization angle confidence score greater than 0.8 within the selected area are filtered to remove noise interference. The filtered polarization angle data is then processed. Create an index with the current frame number #10245 to generate a set of polarization angle values ​​for the bullseye region.

[0043] The direction vector construction submodule is based on the polarization angle distribution in each frame in the polarization angle value set of the bullseye region. According to the directionality definition corresponding to the polarization angle, it performs vector superposition operation on the polarization angle value of the bullseye region in each frame image, calculates the spatial direction main vector corresponding to the current frame, and establishes the index relationship between the frame number and the direction vector to generate a spatial direction distribution vector sequence. The direction vector construction submodule is based on the intra-frame polarization angle distribution in the bullseye region polarization angle value set. First, it selects... Statistical analysis was performed on the polarization angle data, and an angle histogram was constructed by identifying statistical distribution patterns. to The area is divided into 18 bins. The number of pixels falling into each bin is counted, and the peak bin with the highest number of pixels is identified. The center angle of this bin is taken as the principal polarization angle. .

[0044] For example, the peak range is ,but Based on the definition of directionality corresponding to the polarization angle, a unit direction vector is constructed. Substitute have to This vector represents the spatial direction of the bull's gaze. This process is performed on each frame to obtain a time-varying sequence of direction vectors, and frame number #10245 is then assigned to the vector. The index relationships are stored in the database. For a continuous frame sequence, the dot product of adjacent frame vectors is calculated to verify directional continuity, ensure the smoothness of the vector field, and generate a spatially oriented vector sequence.

[0045] The trajectory sequence generation submodule reads the individual cow number corresponding to the image frame number based on the direction information of each frame in the spatial direction distribution vector sequence, and obtains the direction record interval corresponding to the number in the historical behavior database. It then performs temporal matching between the current spatial direction vector sequence and the historical direction interval to generate a cow eye posture trajectory prediction sequence. The trajectory sequence generation submodule first reads the individual cow number (e.g., ID: C-2025-X99) associated with image frame number #10245 based on the direction information corresponding to each frame in the spatial direction distribution vector sequence. It then accesses the historical behavior database to retrieve all passageway behavior records of this individual in the past 30 days and extracts its historical direction vector interval data.

[0046] Employing the Dynamic Time Warping (DTW) algorithm and integrating temporal trajectory pattern recognition technology, the currently generated spatial direction vector sequence is... With standard trajectory templates in historical records Perform temporal matching and calculate the similarity distance between sequences. If the matching distance is less than a preset threshold (e.g., ...), If the distance is too large, the current posture trajectory is considered to conform to the existing behavioral patterns of the cow. If the distance is too large, it is marked as a novel trajectory. Based on the matching results, the head movement trend of the cow in the next few frames is predicted, and the predicted posture coordinate sequence is bound to the current frame number to generate a cow eye posture trajectory prediction sequence.

[0047] Specifically, such as Figure 2 , 7 As shown, the structural parameter measurement module includes: The structural point extraction submodule collects the image coordinates of the nasal tip, the corners of the eyes and the maxillary point in the corresponding image based on the frame number and image position index of each image in the bovine eye posture trajectory prediction sequence. It then sets a coordinate template for key points in the bovine face structure anatomical region and summarizes the structural point coordinate information according to the image frame number to generate a set of key structural point coordinates. The structural point extraction submodule, based on the frame number and image position index of each image in the bull's eye pose trajectory prediction sequence, first loads a coordinate template for the anatomical region of the bull's face structure into each frame of the corrected image. This template is designed based on biological topological pattern recognition and defines the relative topological relationships of the nasal tip, the inner corners of both eyes (left inner corner and right inner corner), and the maxillary point under standard emmetropic conditions.

[0048] Template matching technology is used to search for corresponding feature points in the image and collect the coordinates of the nose tip. Left corner of the eye Right corner of the eye and maxillary point For example, in frame #10245, the following was captured: , , , Perform a geometric consistency check on the extracted coordinates to verify whether the line connecting the eyes is approximately horizontal (absolute slope value). The system also checks whether the tip of the nose is located below the line connecting the eyes, eliminates misidentified points that do not conform to anatomical logic, and finally summarizes all valid coordinates in the continuous sequence to generate a set of coordinates for key structural points.

[0049] The spatial difference recognition submodule extracts the location of structural points in each frame of images from the key structural point coordinate set. It calculates the Euclidean distance difference between the coordinate distances from the tip of the nose to the corners of the eyes and from the corners of the eyes to the maxilla in any two image frames, constructs the relative position difference matrix of key structural points, and generates a spatial difference distribution map of structural points. The spatial difference recognition submodule extracts the location of structural points in each frame of the key structural point coordinate set. First, it selects two image frames at different time points, for example, frame #10245 (time point). ) and frame #10265 (time) Calculate the Euclidean distance between keypoints within each frame to construct a distance feature vector. For frame #10245, calculate the distance from the tip of the nose to the left corner of the eye. Pixels, calculating the distance from the corners of both eyes to the maxillary point. .

[0050] Then the difference between the corresponding distance features between the two frames is calculated. Assuming the corresponding distance in frame #10265 is 185.50 pixels, then Pixels. Construct a The relative position difference matrix captures and quantifies the subtle deformation patterns during facial movement, and this matrix intuitively reflects the projection differences caused by changes in viewing angle. As shown in Table 3, all calculated difference data are mapped to grayscale maps to generate a spatial difference distribution map of structural points.

[0051] Table 3 Calculation Table of Spatial Distance Differences of Key Structural Points The morphological parameter generation submodule classifies the spatial relationship of key points under the same viewpoint sequence image frame number based on the spatial difference data between each pair of structural points in the spatial difference distribution map of structural points, calculates the relative offset values ​​of structural points in the horizontal, vertical and radial directions in multi-view images, constructs key region morphological measurement indexes, and generates key morphological parameter identification records for the cow face. The morphological parameter generation submodule, based on the spatial difference data between pairs of structural points in the spatial difference distribution map, first categorizes the data under the same viewpoint sequence according to the image frame number, and identifies the stable frame interval with the smallest difference value. Within this interval, it calculates the mean relative offset of the structural points in the horizontal direction (X-axis), vertical direction (Y-axis), and radial direction (Z-axis depth estimation).

[0052] Using formula Calculate the rate of morphological change, where The reference distance (such as the distance between the eyes). For the first The spatial difference of each feature point. Assuming the interocular distance is 200 pixels and the average spatial difference of the nose tip is 5.22 pixels, then the morphological change index is... Based on this indicator, a key area morphological measurement system is constructed, and a long-term biological growth and development pattern recognition model is invoked. The calculation results are compared with the standard parameters in the model. If the rate of change is within the preset growth range, the animal's biological characteristic profile is updated, and a detailed record containing morphological indices of various dimensions is finally output, generating a record of key morphological parameters of the animal's face.

[0053] Please see Figure 8The method for determining key morphological parameters of a cow's face based on multi-view images is executed based on the aforementioned system for determining key morphological parameters of a cow's face based on multi-view images, and includes the following steps: S1: Obtain the image frame number, image acquisition timestamp, and image channel data from the cattle channel camera, and match and align them with the triaxial vibration frequency value of the inertial sensor and the sound wave reflection time value of the ultrasonic array to generate a multi-parameter alignment relationship group; S2: Based on the image channel data in the multi-parameter alignment relationship group, calculate the confidence of the cow face structure region in the image, mark the abnormal path image frames, and output the path type label; S3: Extract the abnormal path image frame number from the path type label, calculate the offset displacement value of each pixel in the vibration direction, combine the sound wave reflection time value to construct the image structure boundary displacement range map, and output the pixel offset correction image group. S4: Collect the image coordinate information of the bull's eye region in the pixel offset correction image group, extract the polarization angle value of the bull's eye region in each frame image, calculate the spatial direction distribution vector of the bull's eye, combine it with historical behavior to perform time-series interval superposition analysis, and generate a bull's eye attitude trajectory prediction sequence. S5: Based on the predicted sequence of bovine eye posture trajectories, extract the image coordinates of the nasal tip, the corners of both eyes, and the maxillary point. Summarize the distribution relationships of key structural points in multi-view image frames and output the key morphological parameter identification record of the bovine face in the multi-view images. The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A system for determining key morphological parameters of a cow's face based on multi-view images, characterized in that, include: The data synchronization module obtains the image frame number, image acquisition timestamp, and image channel data from the cattle channel camera, and performs matching and alignment with the triaxial vibration frequency value of the inertial sensor and the sound wave reflection time value of the ultrasonic array to generate a multi-parameter alignment relationship group. The region segmentation and discrimination module calculates the confidence level of the cow face structure region in the image based on the image channel data in the multi-parameter alignment relationship group, marks abnormal path image frames, and outputs path type labels; The distortion correction construction module extracts the abnormal path image frame number from the path type label, calculates the offset displacement value of each pixel in the vibration direction, and constructs the image structure boundary displacement range map by combining the sound wave reflection time value, and outputs the pixel offset correction image group. The attitude prediction fusion module collects the image coordinate information of the bull's eye region in the pixel offset correction image group, extracts the polarization angle value of the bull's eye region in each frame image, calculates the spatial direction distribution vector of the bull's eye, and performs temporal interval superposition analysis in combination with historical behavior to generate a bull's eye attitude trajectory prediction sequence. The structural parameter determination module extracts the image coordinates of the nose tip, the corners of both eyes and the maxilla based on the predicted sequence of the bovine eye posture trajectory, summarizes the distribution relationship of key structural points in the multi-view image frames, and outputs the key morphological parameter identification record of the bovine face in the multi-view image.

2. The system for determining key morphological parameters of a cow's face based on multi-view images according to claim 1, characterized in that: The abnormal path image frame specifically refers to the image frame in which the confidence level of the cow face structure region is greater than a set dynamic confidence threshold.

3. The system for determining key morphological parameters of a cow's face based on multi-view images according to claim 1, characterized in that: The multi-parameter alignment relationship group includes image frame number matching relationship, image acquisition timestamp mapping relationship, vibration frequency value and image frame comparison table, and sound wave reflection time value and image frame comparison table. The path type label includes normal image path label, abnormal image path label, and low confidence path label. The pixel offset correction image group includes pixel vector remapping map, structural boundary displacement range map, and vibration direction offset correction map. The bull's eye posture trajectory prediction sequence includes bull's eye polarization angle change curve, direction distribution vector time series map, and behavior trajectory reconstruction map. The key morphological parameter identification record of the bull's face includes nasal tip point spatial position parameter, eye corner point relative distance parameter, and maxillary point posture angle parameter.

4. The system for determining key morphological parameters of a cow's face based on multi-view images according to claim 1, characterized in that, The data synchronization construction module includes: The data stream receiving submodule collects the image streams from the visible light cameras on both sides of the cattle channel, including the image frame number, image acquisition timestamp, and image channel data. It also collects the triaxial vibration frequency values ​​output by the inertial sensor and the sound wave reflection time values ​​recorded by the ultrasonic array within the time period corresponding to the image timestamp. Data pairs with time differences exceeding the maximum interval threshold between image frames are removed, and image frames and sensing time series sets are generated. The multi-source data pairing submodule extracts data groups based on the same timestamp according to the image frame number, triaxial vibration frequency value and sound wave reflection time value in the image frame and sensing time series set. It performs interval averaging on the vibration frequency value and sound wave reflection time value to construct the binding relationship between the image frame number and the sensor parameters of the time period, and generates the image frame and multi-source sensing data binding set. The parameter alignment generation submodule establishes an index mapping for the image frame number based on the image frame number and the bound vibration frequency value and sound wave reflection time value in the image frame and multi-source sensing data binding set, establishes a matching relationship between the image frame and the parameters, and generates a multi-parameter alignment relationship group.

5. The system for determining key morphological parameters of a cow's face based on multi-view images according to claim 1, characterized in that, The region segmentation and discrimination module includes: The image grayscale extraction submodule extracts the pixel matrix of the red, green, blue and infrared channels in each frame image based on the image channel data in the multi-parameter alignment relationship group, performs channel normalization processing on the four-channel pixel matrix, and generates an image channel grayscale value matrix set. The confidence calculation submodule locates the coordinate range corresponding to the cow face structure region in the image based on the gray value of each frame in the image channel gray value matrix set, calculates the confidence based on the gray value distribution of the region's red, green, blue and infrared channels, and simultaneously calculates the average confidence of the cow face region in historical image frames and constructs a dynamic confidence threshold to generate structural region confidence data. The path type labeling submodule determines whether the confidence of the current frame is higher than the dynamic confidence threshold based on the confidence data of the structural region. If the condition is met, the image frame is labeled as an abnormal path image frame, and the corresponding label number and discrimination result under the abnormal state are output to generate a path type label.

6. The system for determining key morphological parameters of a cow's face based on multi-view images according to claim 1, characterized in that, The distortion correction construction module includes: The principal point angle construction submodule extracts the image position coordinates of all pixels in each image based on the image frame number marked as an abnormal path in the path type label, collects the coordinates of the image center point as the principal point position, calculates the vector angle value between each pixel and the principal point, and indexes and binds the image frame number with the corresponding angle data to generate a pixel angle relationship matrix. The vibration offset calculation submodule reads the angle data in the pixel angle relationship matrix and the vibration frequency value in the multi-parameter alignment relationship group, constructs a projection vector relationship model based on the angle between the pixel vector direction and the vibration direction, calculates the unit displacement value and total displacement value of each pixel in the vibration direction, establishes a displacement data structure according to the image frame number, and generates a vibration direction offset displacement set. The pixel coordinate remapping submodule performs coordinate remapping between the original pixel position and the direction of the offset value within the image frame based on the offset value of each pixel in the vibration direction offset displacement set. At the same time, it constructs a response delay region map of the pixel boundary structure by combining the sound wave reflection time value, and superimposes the offset mapping and the boundary region range to generate a pixel offset correction image group.

7. The system for determining key morphological parameters of a cow's face based on multi-view images according to claim 1, characterized in that, The attitude prediction fusion module includes: The polarization angle extraction submodule, based on the pixel offset correction image group, collects the image coordinate information of the bullseye region in the image, obtains the polarization channel value of the corresponding pixel position, filters the polarization angle data of each pixel in the bullseye region, and constructs a mapping index between the frame number and the polarization angle of the bullseye region to generate a set of polarization angle values ​​for the bullseye region. The direction vector construction submodule is based on the polarization angle distribution in each frame of the bullseye region polarization angle value set. According to the directionality definition corresponding to the polarization angle, it performs vector superposition operation on the polarization angle values ​​of the bullseye region in each frame image, calculates the spatial direction main vector corresponding to the current frame, establishes the index relationship between the frame number and the direction vector, and generates a spatial direction distribution vector sequence. The trajectory sequence generation submodule reads the individual cow number corresponding to the image frame number based on the direction information corresponding to each frame in the spatial direction distribution vector sequence, and obtains the direction record interval corresponding to the number in the historical behavior database. It then performs time-series matching between the current spatial direction vector sequence and the historical direction interval to generate a cow eye posture trajectory prediction sequence.

8. The system for determining key morphological parameters of a cow's face based on multi-view images according to claim 1, characterized in that, The structural parameter measurement module includes: The structural point extraction submodule collects the image coordinates of the nasal tip, the corners of the eyes and the maxillary point in the corresponding image based on the frame number and image position index of each image in the bovine eye posture trajectory prediction sequence. It then sets a coordinate template for key points in the bovine face structure anatomical region and summarizes the structural point coordinate information according to the image frame number to generate a set of key structural point coordinates. The spatial difference recognition submodule extracts the position of structural points in each frame of the key structural point coordinate set, calculates the Euclidean distance difference between the coordinate distances from the tip of the nose to the corners of the eyes and from the corners of the eyes to the maxilla in any two image frames, constructs the relative position difference matrix of the key structural points, and generates a spatial difference distribution map of structural points. The morphological parameter generation submodule classifies the spatial relationship of key points under the same viewpoint sequence image frame number based on the spatial difference data between each pair of structural points in the spatial difference distribution map of the structural points, calculates the relative offset values ​​of structural points in the horizontal, vertical and radial directions in the multi-view images, constructs the morphological measurement index of key regions, and generates key morphological parameter identification records for the cow face.

9. A method for determining key morphological parameters of a cow's face based on multi-view images, characterized in that, The system for determining key morphological parameters of a cow's face based on multi-view images, according to any one of claims 1-8, comprises the following steps: S1: Obtain the image frame number, image acquisition timestamp, and image channel data from the cattle channel camera, and match and align them with the triaxial vibration frequency value of the inertial sensor and the sound wave reflection time value of the ultrasonic array to generate a multi-parameter alignment relationship group; S2: Based on the image channel data in the multi-parameter alignment relationship group, calculate the confidence level of the cow face structure region in the image, mark the abnormal path image frames, and output the path type label; S3: Extract the abnormal path image frame number from the path type label, calculate the offset displacement value of each pixel in the vibration direction, combine the sound wave reflection time value to construct the image structure boundary displacement range map, and output the pixel offset correction image group. S4: Collect the image coordinate information of the bull's eye region in the pixel offset correction image group, extract the polarization angle value of the bull's eye region in each frame image, calculate the spatial direction distribution vector of the bull's eye, and perform temporal interval superposition analysis in combination with historical behavior to generate a bull's eye posture trajectory prediction sequence; S5: Based on the bull's eye posture trajectory prediction sequence, extract the image coordinates of the nose tip, the corners of both eyes and the maxillary point, summarize the distribution relationship of key structural points in the multi-view image frames, and output the key morphological parameter identification record of the bull's face in the multi-view image.

Images