A lightweight, high-precision 3D model reconstruction and body size measurement system and method for Anhui white geese

Abstract

This invention relates to the fields of livestock and poultry breeding management and 3D computer vision technology, and discloses a lightweight 3D high-precision model reconstruction and body size measurement system and method for Anhui West White Goose. The method includes steps such as multi-view RGB image acquisition, image preprocessing, feature extraction and matching, camera pose estimation, sparse-dense reconstruction, Gaussian splash modeling, mesh wrapping, key feature point localization, body size parameter calculation, and error correction. By employing innovative techniques such as anisotropic Gaussian modeling, multi-resolution optimization, multi-view consistency constraints, and multi-parameter linear regression correction, non-contact, high-efficiency, and high-precision body size measurement is achieved, with a measurement accuracy within ±0.1cm and a measurement efficiency improvement of over 90% compared to traditional methods.

Description

Technical Field

[0001] This invention relates to the fields of livestock and poultry breeding management and three-dimensional computer vision technology, and more specifically, to a lightweight three-dimensional high-precision model reconstruction and body size measurement system and method for Anhui white geese. Background Technology

[0002] The Anhui White Goose is a distinctive breed of livestock and poultry listed in the National List of Protected Livestock and Poultry Genetic Resources. Its body size parameters (such as height, length, and chest circumference) are crucial for precise feed formulation, optimized breeding cycles, and product quality assessment. Currently, livestock and poultry body size measurement primarily employs contact measurement methods, directly measuring various animal parameters using manual measuring tools. However, this method has several major drawbacks: first, it is inefficient, requiring manual operation of each animal individually, which is extremely labor-intensive in large-scale farming scenarios and difficult to meet the measurement needs of large-scale farms; second, it easily triggers stress responses in animals, as direct contact and fixation negatively impact animal welfare; and third, the consistency of manual operation is difficult to guarantee, resulting in significant measurement errors, typically exceeding ±0.5 cm, which fails to meet the demands of precision farming management. With the rapid penetration of smart agriculture into traditional livestock and poultry farming, traditional contact measurement methods have become a bottleneck restricting the digital development of livestock and poultry farming. Summary of the Invention

[0003] This invention provides a lightweight, high-precision three-dimensional model reconstruction and body size measurement system and method for Anhui white geese, solving the technical problems of low efficiency, easy to induce stress response, and large measurement error of traditional contact measurement in related technologies.

[0004] This invention provides a lightweight, high-precision three-dimensional model reconstruction and body size measurement method for Anhui white geese, including:

[0005] Multi-view RGB images of geese were captured by deploying synchronous high-definition cameras at multiple locations in the passageway and using hardware-triggered synchronous shooting to ensure the temporal consistency of the multi-view images.

[0006] The acquired multi-view RGB images are preprocessed, including image denoising using a two-dimensional Gaussian convolution kernel, distortion correction based on camera intrinsic parameter calibration, brightness normalization, and image size unification.

[0007] Local feature points in images from different viewpoints are extracted using feature extraction algorithms, feature descriptors are constructed and cross-viewpoint matching is performed to obtain the initial feature correspondence between different viewpoints.

[0008] Camera pose estimation is performed using feature matching results. The coordinates of the sparse point cloud of camera extrinsic participants are optimized by the bundle adjustment method to generate the initial sparse 3D point cloud on the surface of the goose.

[0009] Based on camera pose and intrinsic parameters, multi-view stereo matching is performed on multi-view images. The matching cost is calculated using pixel similarity calculation method. Through depth propagation, depth optimization and depth map fusion processing, a dense 3D point cloud covering the surface of the goose is generated.

[0010] Gaussian splashing technology is applied to model and optimize dense point clouds. A three-dimensional Gaussian function is used to characterize the points in the point cloud and generate a three-dimensional Gaussian distribution to represent the surface details of the goose.

[0011] The Gaussian splash model is meshed, and the Gaussian representation is converted into a triangular mesh model through surface reconstruction. A structured three-dimensional mesh model is generated using an isosurface extraction algorithm.

[0012] Based on prior knowledge of goose anatomy, a semantic segmentation algorithm is used to identify key anatomical feature points of geese in a 3D mesh model. The feature point coordinates are refined through clustering and curvature analysis to obtain high-precision 3D coordinates of key feature points.

[0013] Based on the three-dimensional coordinates of the identified key feature points, spatial geometric calculation methods are used to obtain body size parameters, including core body size parameters such as body height, body length, and chest circumference.

[0014] The predicted body size parameters of the model are corrected by a multi-parameter linear regression equation. The regression equation covers multiple error sources such as feature point positioning deviation, feature matching error, and grid curvature. The least squares method is used to train the fit, so as to achieve high-precision body size measurement.

[0015] Furthermore, image preprocessing includes:

[0016] The original image is convolved pixel by pixel using Gaussian weights that are higher at the center than at the periphery to eliminate noise in the image.

[0017] By using the focal length parameter and principal point coordinates in the camera intrinsic parameter matrix, the image distortion introduced by the lens is compensated through radial and tangential distortion coefficients;

[0018] Linearly transform the pixel values ​​of the image to a uniform brightness target range;

[0019] Images are scaled using coordinate mapping to achieve a uniform resolution across all images.

[0020] Furthermore, the feature extraction and matching employs algorithms including extremum detection of key points in multi-scale scale space, construction of feature descriptors, and cross-view feature point correspondence.

[0021] Furthermore, the Gaussian splash modeling adopts anisotropic Gaussian parameter settings, where the scaling parameter in the major axis direction is used to represent the scale in the feather direction, and the scaling parameter in the minor axis direction is used to represent the scale perpendicular to the feather direction. Through parameter combination, the complex structural morphology is accurately restored.

[0022] Furthermore, Gaussian splash modeling employs a multi-resolution optimization strategy, which involves iterative optimization in multiple stages: global structural optimization in the low-resolution stage; local geometric refinement in the medium-resolution stage; and detail enhancement in the high-resolution stage, thereby improving model accuracy through progressive optimization.

[0023] Furthermore, a multi-view consistency constraint is introduced in the Gaussian splash optimization process, including a multi-view reprojection loss constraint and an adjacent geometric normal smoothing constraint. The combination of constraint conditions generates a three-dimensional model with small geometric discontinuities.

[0024] Furthermore, the linear regression equation corrects the initially predicted body size parameters by using feature point positioning deviation, feature matching error, and grid curvature factor, and the regression coefficients are determined by least squares fitting.

[0025] Furthermore, the mesh wrapping process estimates the point cloud normal vector through neighborhood covariance analysis, constructs a vector field using the normal vector, solves the surface reconstruction equation to obtain the inner and outer indicator functions, and extracts the triangular mesh from the isosurface of the indicator functions.

[0026] Furthermore, key feature point localization employs semantic segmentation to identify dissected feature points, clusters candidate feature points using a clustering algorithm, and refines coordinates using local curvature calculation.

[0027] This invention provides a lightweight, high-precision 3D model reconstruction and body size measurement system for Anhui white geese, comprising:

[0028] The image acquisition unit includes a synchronous high-definition camera, a hardware triggering device, and a visual edge computing device;

[0029] The image preprocessing unit performs filtering, distortion correction, brightness adjustment, and size unification processing.

[0030] The feature matching unit is used to extract feature points and perform cross-view matching.

[0031] The pose estimation unit is used to calculate camera extrinsic parameters and sparse point cloud reconstruction.

[0032] Stereo matching units are used to generate dense point clouds;

[0033] Gaussian splash modeling unit, used for high-fidelity modeling of point clouds;

[0034] Mesh-wrapping cells are used to generate triangular mesh models;

[0035] Feature point recognition unit, used to locate anatomical feature points;

[0036] Body size calculation unit, used to calculate body size parameters;

[0037] The error correction unit is used to perform multi-parameter linear regression correction.

[0038] The beneficial effects of this invention are as follows:

[0039] This invention employs innovative techniques such as multi-view synchronous RGB image acquisition, image preprocessing, feature extraction and matching, camera pose estimation, sparse-dense reconstruction, Gaussian splash fine modeling, mesh wrapping, semantic segmentation feature point localization, and multi-parameter linear regression error correction to achieve a non-contact, fully automated high-precision body size measurement solution. This effectively solves three fundamental technical problems inherent in traditional contact measurement methods.

[0040] First, the non-contact measurement method completely solves the measurement efficiency problem. Geese only need to pass through the collection channel without manual operation. The entire process is automated, which improves measurement efficiency by more than 90% compared to the traditional manual measurement method of one goose at a time, and fully meets the batch testing needs of large-scale farms.

[0041] Secondly, by completely avoiding direct human contact, the problem of stress response is effectively solved. The non-contact measurement method eliminates the animal's response to direct stimuli during the measurement process, effectively reducing stress response, improving animal welfare, and having no negative impact on animal husbandry results;

[0042] Third, the measurement accuracy problem was effectively solved through multi-level technical means such as Gaussian splashing fine modeling, multi-view consistency constraints, and multi-parameter linear regression correction. The measurement accuracy was further corrected from the basic ±0.2cm (corresponding to feature point positioning error <0.1cm) to within ±0.1cm, which is a significant improvement compared to the ±0.5cm accuracy of the traditional method. This fully meets the requirements of large-scale farms for body size measurement accuracy and provides a reliable data foundation for precision breeding management. Attached Figure Description

[0043] Figure 1 This is a flowchart of the lightweight three-dimensional high-precision model reconstruction and body size measurement method for western Anhui white geese of the present invention.

[0044] Figure 2 This is a schematic diagram of camera position estimation according to the present invention;

[0045] Figure 3 This invention relates to sparse and dense point clouds;

[0046] Figure 4 This is the Gaussian splash reconstruction effect of the present invention;

[0047] Figure 5 This is the prototype of the goose-Gauss ball of the present invention;

[0048] Figure 6 This is the mesh wrapping model of the present invention;

[0049] Figure 7 This refers to the calculation of the body size parameters in the upper part of this invention;

[0050] Figure 8 This relates to the calculation of body size parameters in the lower half of this invention. Detailed Implementation

[0051] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0052] At least one embodiment of the present invention discloses a lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size, such as... Figure 1 As shown, it includes the following steps:

[0053] Step 1: Collect multi-view RGB images of geese to establish the visual data foundation for the 3D model. Utilize 5 to 8 synchronized high-definition cameras (resolution no less than 20 megapixels) positioned in front, behind, left, and right of the goose's passageway. Employ hardware-triggered synchronous shooting to ensure temporal consistency of the multi-view images, generating a multi-view image sequence of the geese.

[0054] Step 2: Preprocess the acquired multi-view RGB images to eliminate the impact of environmental factors and equipment errors on image quality. Image denoising is performed using a two-dimensional Gaussian convolution kernel, distortion correction is performed based on camera intrinsic parameter calibration, brightness normalization is applied, and image size is standardized to obtain a preprocessed, normalized image sequence.

[0055] Image denoising is performed using a two-dimensional Gaussian convolution kernel filtering method, where the Gaussian kernel function is:

[0056]

[0057] in: (Standard deviation): Controls the "width" of the Gaussian distribution, determining the degree of smoothness; The larger the value, the flatter the kernel function and the stronger the smoothing effect; The smaller the value, the "sharper" the kernel function and the weaker the smoothing effect; : Position coordinates relative to the center of the core; Normalization coefficients ensure that the total integral of the kernel function is 1. Then, perform discrete convolution:

[0058]

[0059] in: It is the original image. It is a smooth image. These are pixel coordinates; kernel function parameters: G is the Gaussian kernel function; dimensions: ,in For kernel radius characteristics: the weight values ​​are obtained by discretizing the Gaussian function, satisfying... (Normalization conditions), where Gaussian kernel ( ):

[0060]

[0061] It is the kernel offset, representing the position relative to the center of the convolution kernel, with a range of: , where k is the kernel radius; iterate through all positions within the convolution kernel. Center point: when At time t, the corresponding position is the center of the convolution kernel; k is the kernel radius, a positive integer, the distance from the center of the convolution kernel to the edge; its relationship with the kernel size is: convolution kernel size Typical values: (3×3 cores) (5×5 cores) (7×7 cores); G These are weight values, following a Gaussian distribution, with the largest weight at the center and decreasing towards the edges; all G... The sum is 1, ensuring that the average brightness of the image remains unchanged. Distortion correction is performed based on camera intrinsic parameter calibration. Camera intrinsic parameter matrix:

[0062]

[0063] in Focal length parameter:

[0064]

[0065] f is the actual focal length (in millimeters); W and H are the image sensor dimensions (in millimeters). Pixel size (unit: millimeters / pixel), when At that time, the pixels are square (in most modern cameras); when At that time, the pixels are rectangular (in some industrial cameras); Principal point coordinates, ideally located at the exact center of the image:

[0066]

[0067] Due to manufacturing errors, the principal point is usually slightly off-center, resulting in the distortion coefficient vector:

[0068]

[0069] Caused by lens curvature, resulting in a radial shift of image points, coefficient: (generally Mathematical model for fisheye lenses:

[0070]

[0071]

[0072] in Normalized image coordinates (centered on the principal point); Radial distance to the principal point. Distortion morphology: Barrel distortion. Image edges curve inwards, commonly seen in wide-angle lenses; pincushion distortion: Outward bending of image edges is common in telephoto lenses; tangential distortion is caused by the lens and image sensor not being parallel. (Coefficient:) Mathematical model:

[0073]

[0074]

[0075] Brightness normalization processing:

[0076]

[0077] in For input pixel values, To input the minimum value, Enter the maximum value; To achieve the minimum target value; The target maximum value; This is for outputting pixel values.

[0078] Uniform size (scaling):

[0079]

[0080] For grayscale images: For color images: , The width of the input image; The height of the input image; The dimensions of the output image are: (Grayscale) or (color); To output the image width; This represents the height of the output image. (x, y) represents the coordinate range of the output image. .

[0081] Step 3: Use the Deep-Image-Matching algorithm to extract local feature points from images from each viewpoint, construct feature descriptors and perform cross-viewpoint matching to obtain the initial feature correspondence between each viewpoint.

[0082] The Deep-Image-Matching algorithm is a conventional deep learning feature extraction and matching algorithm. Its input is a multi-view RGB image sequence, and its output is the correspondence of feature points between each viewpoint. This algorithm uses a deep learning network to extract local feature points and their feature descriptors that are robust to scale, rotation, and illumination changes in each viewpoint image. Then, it performs feature matching between viewpoints based on the similarity of the feature descriptors, obtaining feature correspondence pairs between multiple viewpoints. These feature correspondences are used for camera pose estimation in step 4 and stereo matching in step 5.

[0083] Step 4: Use COLMAP to complete camera pose estimation and sparse point cloud reconstruction. Optimize the coordinates of the sparse point cloud with external camera involvement using the bundle adjustment method to generate the initial sparse 3D point cloud and camera pose parameters of the goose surface.

[0084] The COLMAP algorithm is a conventional structure and motion recovery algorithm. Its inputs include the feature correspondences between different viewpoints (obtained in step 3) and the camera intrinsic parameter matrix. The outputs are the extrinsic parameters (rotation matrix and translation vector) of each camera and the sparse coordinates of the 3D point cloud. The bundle adjustment algorithm is a conventional nonlinear optimization algorithm. Its inputs are the initial camera extrinsic parameter estimates, 3D point coordinates, and feature point correspondences. It optimizes all camera extrinsic parameters and 3D point coordinates by minimizing the reprojection error. The outputs are the optimized camera pose and sparse point cloud coordinates, so that the 3D points projected from each camera viewpoint are as consistent as possible with the observed feature point positions.

[0085] Feature point detection algorithm: Constructing SIFT (Scale-Invariant Feature Transform) using Gaussian difference scale space:

[0086]

[0087] Extreme point localization: 3D neighborhood (spatial) in scale space Scale σ) extreme value search:

[0088]

[0089] Precise location of key points (3D quadratic function fitting):

[0090]

[0091] Hessian matrix response calculation SURF (Speeded-Up Robust Features):

[0092]

[0093] Hessian determinant approximation calculation (accelerated by integral image):

[0094]

[0095] in for The second-order partial derivative of the direction is the image in Curvature of direction:

[0096]

[0097] Second-order partial derivative of the direction:

[0098]

[0099] In isotropic image regions: In the edge region: the value is large in one direction and small in the other direction; in the corner region: the value is large in both directions. For mixed partial derivatives:

[0100]

[0101] According to Schwarz's theorem: (The reason for the symmetry of the Hessian matrix) is that the value is smaller at the edges and larger at corners or speckled features. For rotationally symmetric features, .

[0102] Figure 2 This is a schematic diagram for camera position estimation.

[0103] Step 5: Based on the estimated camera pose and intrinsic parameters, perform distortion correction and cropping on the multi-view images, perform multi-view stereo matching, calculate the pixel matching cost using the normalized cross-correlation method, and generate a dense 3D point cloud covering the surface of the goose through processes such as depth propagation, depth optimization and depth map fusion.

[0104] The Normalized Cross-Correlation (NCC) algorithm is a conventional similarity calculation algorithm. Its input consists of the pixel neighborhood in the reference image and the corresponding neighborhood projected from the target image based on a depth assumption. The output is a similarity index (0 for a perfect match, 1 for a complete mismatch). The depth propagation algorithm is a conventional optimization strategy that progressively improves the depth estimate of each pixel through spatial and view propagation. It compares the depth values ​​of neighboring pixels with the current pixel; if the depth value of a neighboring pixel reduces the matching cost, the depth of the current pixel is updated. The depth optimization algorithm is a conventional probabilistic optimization process that assumes the depth value of each pixel follows a specific probability distribution and uses an MRF (Markov Random Field) model to constrain the depth values ​​of neighboring pixels to maintain a smooth change. The depth map fusion algorithm is a conventional multi-view ensemble algorithm. Its input consists of depth maps from multiple viewpoints, and the output is a consistent 3D point cloud. It transforms the depth maps from each viewpoint to a global coordinate system and uses voxelization or an octree structure to store and fuse duplicate 3D points.

[0105] Matching cost calculation: Pixel matching cost is calculated between image pairs based on normalized cross-correlation (NCC) or photometric consistency.

[0106]

[0107] Where: measures the matching quality of pixel p under the depth assumption d, 0: perfect match ( ); 1: Completely irrelevant ( ); 2: Perfectly negative correlation ( ), Optimization goal: Find ways to make C Minimum depth d; p is the pixel position of the reference image; d is the assumed depth value. ;I The reference image is pixel brightness type, where q is the pixel position in the neighborhood. It is usually converted to a grayscale image and then histogram equalization is performed. For the target image pixel brightness, To be based on depth The corresponding position after projection, due to The coordinates are typically not integers, requiring bilinear interpolation. Then, depth propagation and optimization, depth map filtering, and depth map fusion are performed. Depth propagation and optimization: Through iterative random initialization, spatial propagation, view propagation, and depth normal optimization, the depth value of each pixel is progressively optimized. Depth map filtering: Outliers are removed based on geometric consistency checks and depth continuity constraints, retaining reliable depth estimates. Depth map fusion: Multi-view depth maps are uniformly transformed to a global coordinate system, and duplicate points are fused using voxel hashing or an octree structure to generate a globally dense point cloud.

[0108] Figure 3 For sparse and dense point clouds.

[0109] Step 6: Apply Gaussian splashing technology to model and optimize the dense point cloud. Use a three-dimensional Gaussian function to represent each point in the point cloud, generate a large number of three-dimensional Gaussian distributions to represent the surface details of the goose, and obtain a high-fidelity three-dimensional geometric model.

[0110] The aforementioned Gaussian splash model is a very large-scale model, whose constituent units include assigning a three-dimensional Gaussian function to each point cloud point. Each Gaussian function consists of the following parameters: mean vector. The covariance matrix represents the center position of the Gaussian sphere. By scaling the vector and rotation matrix Combining the results, we get: ,in A diagonal matrix; color parameters Represents RGB color values; opacity parameter The transparency of the Gaussian sphere is controlled. In this model, the training mode is unsupervised learning, and all Gaussian parameters are optimized to reconstruct the 3D surface of the goose using given multi-view images and camera parameters.

[0111] The training optimization strategy for the Gaussian splash model employs the gradient descent algorithm, starting with randomly initialized Gaussian parameters and gradually optimizing the model parameters through multiple training iterations. The model's loss function consists of multiple components, including multi-view reprojection loss and normal consistency loss. The multi-view reprojection loss constrains the reconstruction quality by comparing the differences between the Gaussian model's projection results and the actual observed images from each viewpoint. The normal consistency loss avoids geometric discontinuities by constraining the smooth change of the normal directions of adjacent Gaussian spheres.

[0112] 3D Gaussian function:

[0113]

[0114] Where X is the input vector (random variable) and D is a column vector:

[0115]

[0116] For data points used to calculate probability density, the dimension D of the feature is: One-dimensional Gaussian distribution (normal distribution) Two-dimensional Gaussian distribution (elliptical contour lines); : High-dimensional Gaussian distribution. In image processing: ( In point cloud processing: ( In the feature space: ; Mean vector:

[0117]

[0118] The location with the highest probability density:

[0119]

[0120] Each dimension has its own mean. If all dimensions are independent and identically distributed, then . is the covariance matrix. Symmetric positive definite matrix:

[0121]

[0122] For feather detail enhancement strategies, anisotropic scaling is used to target the slender structure of feathers:

[0123]

[0124] in Long axis, The short axis is used to simulate the shape of a feather. Coarse-to-fine strategy: Stage 1 (low resolution, global optimization): Resolution: Iterations: 10000; Stage 2 (Medium resolution, local optimization): Resolution: Iteration: 15000; Stage 3 (High Resolution, Enhanced Detail): Resolution: Iteration: Multi-view reprojection loss:

[0125]

[0126] Normal consistency loss (estimating normal from Gaussian distribution):

[0127]

[0128] in: The multi-view reprojection loss measures the consistency between the reconstructed model and the actual observations across all views. The optimization objective is to improve the accuracy of the reconstructed model across multiple views by minimizing this loss. K represents the total number of views used for reconstruction: small-scale: K=310 (sparse view reconstruction); medium-scale: K=1050 (multi-view reconstruction); large-scale: K>50 (dense view or video sequence). More views result in stronger constraints and more stable reconstruction, but the computational cost increases linearly with K. For the first Predicted reconstructed values ​​from each perspective; For the first The actual observation value from each perspective.

[0129] Figure 4 This is a reconstruction effect using Gaussian splashing.

[0130] Figure 5 It is the prototype of the Goose Gauss sphere.

[0131] Step 7: Mesh the Gaussian splash model, convert the Gaussian representation into a triangular mesh model through Poisson surface reconstruction, and extract the mesh from the isosurface of the indicator function using the moving cube algorithm to generate a structured 3D mesh model (obj format).

[0132] The Poisson surface reconstruction is a conventional implicit surface reconstruction algorithm. Its input is a point cloud and its normal vector, and its output is an indicator function value. This indicator function is 0.5 on the large surface, tends to 1 inside, and tends to 0 outside. The moving cube algorithm is a conventional isosurface extraction algorithm. Its input is the indicator function value obtained by solving the Poisson equation, and its output is a triangular mesh model. By traversing the cube elements in the discretized space, the intersection of the cube elements with the isosurface is determined based on the indicator function value at each vertex, and the corresponding triangular patches are generated.

[0133] Poisson surface reconstruction process, point cloud normal estimation: For each Gaussian splash point, its normal vector is calculated through local neighborhood covariance analysis.

[0134]

[0135] Where k is the number of neighborhood points. Let the three-dimensional coordinates of the j-th neighboring point be:

[0136]

[0137] It is a three-dimensional column vector. For the "centroid" or "average point" of the local neighborhood:

[0138]

[0139] satisfy:

[0140]

[0141] The indicator function is obtained by solving the Poisson equation:

[0142]

[0143] in The vector field is constructed from the point cloud and its normal vectors:

[0144]

[0145] Isosurface extraction: using the moving cube algorithm from Extracting triangular meshes from isosurfaces:

[0146]

[0147] Adaptive resolution control: To balance computational efficiency and model accuracy, an adaptive mesh resolution strategy is adopted, with local mesh edge length as follows:

[0148]

[0149] High curvature region ( (Large), L Small (fine grid); flat area Small), Large (coarse grid).

[0150] The base resolution is used to adjust the degree to which curvature changes affect the mesh size; For local curvature, local point cloud covariance analysis is used. in These are the eigenvalues ​​of the covariance matrix, which can be converted into the obj model for easier measurement.

[0151] Figure 6 The model is wrapped in a mesh.

[0152] Step 8: Combining prior knowledge of goose anatomy, use semantic segmentation algorithms to identify key anatomical feature points such as the top of the goose's head, shoulder armor, end of the keel, and tibia joint in the 3D mesh model. Refine the feature point coordinates through Euclidean distance clustering and local curvature analysis to obtain high-precision 3D coordinates of key feature points.

[0153] The aforementioned semantic segmentation network is a convolutional neural network. Its input consists of geometric features such as spatial coordinates, local curvature, and normal vectors of each vertex in a 3D mesh model, as well as feature vectors composed of color and texture information extracted from a Gaussian splatter model. The output is the probability distribution of the anatomical structure category corresponding to each vertex. The network adopts an encoder-decoder architecture. The encoder gradually extracts the local and global geometric features of the vertices through multiple convolutional and pooling layers, while the decoder gradually restores the spatial resolution and fuses multi-scale features through upsampling and skip connections. In this invention, the semantic segmentation network adopts a supervised learning training mode and is trained using a 3D goose model dataset with anatomical feature point annotations. The training optimization strategy adopts the Adam optimization algorithm. The loss function is a weighted combination of cross-entropy loss and Dice loss, used to balance the recognition accuracy among multiple anatomical structure categories. The network output undergoes thresholding and post-processing to obtain pixel-level predictions for each anatomical feature point category. Then, the predicted pixels of the same category are aggregated using the Euclidean distance clustering algorithm, and the cluster center is obtained as the preliminary coordinate estimate of the feature point. Finally, the preliminary coordinates are refined using local curvature information to obtain the high-precision final coordinates of the feature points.

[0154] The Euclidean distance clustering is a conventional clustering algorithm. Its input is a set of feature points of the same category (identified by a semantic segmentation network), and its output is the coordinates of the cluster center. This algorithm calculates the Euclidean distance between points and groups points that are close together into a cluster. The aggregation method is as follows: if the Euclidean distance between two points is less than a predetermined distance threshold, they are considered to belong to the same cluster. Finally, the average coordinates of all points within each cluster are calculated as the representative point of that cluster.

[0155] Step 9: Based on the 3D coordinates of the identified key feature points, obtain the body size parameters using spatial geometric calculation methods. Body height is the distance between the highest point of the head feature point in the vertical axis and the ground; body length is the straight-line distance between the feature points at the base of the neck and the base of the tail; chest circumference is the circumference of the circle obtained by fitting the contour point at the thickest part of the chest using the least squares method. The core body size parameters of the goose are calculated.

[0156] The least squares method is a conventional curve fitting algorithm applied to bust circumference calculation. Its input is the set of contour points at the widest point of the bust, and its output is the parameters (center and radius) of the fitted circle. This algorithm finds the optimal center coordinates by minimizing the sum of the squared distances from all contour points to the fitted circle. and radius This minimizes the deviation between the circle and the actual contour points. The bust parameter is defined as the circumference of the fitted circle, i.e. .

[0157] It should be noted that the image preprocessing operations in step 2 can employ different processing sequences, but the integrity of image denoising, distortion correction, brightness normalization, and size unification should be ensured. This preprocessing step eliminates the influence of factors such as changes in ambient lighting, camera distortion, and resolution differences, ensuring the stability and reliability of subsequent feature matching algorithms.

[0158] It should be noted that in the multi-view stereo matching process in step 5, depth map filtering uses geometric consistency checks and depth continuity constraints to remove outliers and retain reliable depth estimates verified from multiple perspectives. Depth map fusion uses an octree or voxel hash structure to store and fuse duplicate 3D points. The resulting dense point cloud typically contains millions of points, providing sufficient geometric information for Gaussian splash modeling.

[0159] It should be noted that in the mesh wrapping process in step 7, the point cloud normal vector is estimated by calculating the neighborhood covariance matrix of each Gaussian splash point. These normal vectors are then used to construct a vector field, and the Poisson equation is solved to obtain the indicator function, thus accurately distinguishing between the interior and exterior of the model. The adaptive resolution control strategy dynamically adjusts the mesh edge length based on the local curvature of the point cloud, using a fine mesh in high-curvature regions (such as details like feathers) and a coarse mesh in flat regions, achieving a balance between computational efficiency and model accuracy.

[0160] In this embodiment of the application, in order to further improve the ability of the Gaussian splash model to represent complex structures (especially feathers), an anisotropic Gaussian modeling strategy is adopted in step 6, and the anisotropic scaling parameter of the Gaussian distribution is set to... ,in The scaling of the major axis indicates the direction of the feathers. This parameter represents the scaling of the short axis perpendicular to the feather direction. This parameter setting accurately reproduces the slender shape of the feather and can capture complex structures such as feathers more accurately than an isotropic Gaussian distribution.

[0161] In this embodiment of the application, in order to further optimize the detail representation of the Gaussian splash model, a multi-resolution optimization strategy is adopted, which is carried out in three stages of iterative optimization: the first stage adopts The resolution is globally optimized, iterating 10,000 times to quickly converge the global structure; the second stage adopts... The resolution is locally optimized, iterating 15,000 times to refine the local geometry; the third stage adopts... The resolution was enhanced through 5000 iterations to improve the expressiveness of details such as feathers. This progressive optimization, from coarse to fine, ensured both the stability of model convergence and significantly improved the accuracy of detail reproduction.

[0162] In this embodiment, to overcome the geometric inconsistency problem caused by different viewpoints, a multi-view consistency constraint is introduced in the Gaussian splash optimization process in step 6. The multi-view reprojection loss of all viewpoints is summed to constrain the consistency between the projection results of the Gaussian model under all training viewpoints and the actual observed images. At the same time, a normal consistency loss is introduced to constrain the normal direction of adjacent Gaussian spheres to maintain a smooth change, avoiding geometric discontinuities or wavy distortions, thereby generating a more stable and accurate 3D model.

[0163] In this embodiment, to further improve the accuracy of body size parameter measurement, a linear regression error correction step is added after step 9 to establish a precise mapping relationship between model predicted values, error influencing factors, and true values. The linear regression model is a conventional multivariate linear regression model. Its input layer contains multiple error influencing factors, including feature point localization bias (measuring the difference between the coordinates of the output points of the semantic segmentation network and the refined coordinates), feature matching error (measuring the reliability of the feature correspondence obtained by the Deep-Image-Matching algorithm, obtained through the matching confidence index), mesh curvature (characterizing the local geometric characteristics of the 3D mesh model), and other multi-dimensional error sources. The output layer consists of corrected body size parameters, including corrected body length, corrected chest circumference, and other core measurement parameters. The input feature vector of the aforementioned linear regression model is represented as follows: ,in For the first One error influencing factor; the output is: ,in For the intercept term, For the first The regression coefficients of each feature. The training method of the linear regression model adopts a supervised learning mode, using a training dataset containing measured scale annotations to train the model; the training optimization strategy is the least squares method, which solves for the optimal regression coefficients by minimizing the mean square error between the predicted and the true values; the loss function is... ,in The number of training samples. For the first The true body size value of each sample. These are the predicted values; model performance is measured by the coefficient of determination. Assessment, in which The average of the true values, when This indicates that the model has sufficient fitting accuracy and can effectively correct measurement errors.

[0164] Figure 7 Calculate the body size parameters for the upper part;

[0165] Figure 7 Calculate the body size parameters for the lower half.

[0166] This implementation method solves key problems of traditional contact measurement, such as low efficiency, easy to induce stress response, and large measurement error, by adopting innovative technologies such as multi-view synchronous RGB image acquisition, Gaussian splash fine modeling, AI intelligent feature point positioning, and multi-parameter linear regression error correction.

[0167] First, compared to the traditional contact measurement method that requires manual operation on each goose, this implementation method uses non-contact measurement. The geese can complete the body size data collection simply by passing through the channel, without the need for direct human contact or fixation. This significantly reduces the stress on the animals during the measurement process, improves animal welfare, and increases the measurement efficiency by more than 90%, meeting the batch testing needs of large-scale farms.

[0168] Secondly, by employing 5 to 8 simultaneous cameras for multi-view imaging, unobstructed imaging of the goose's full outline was ensured, providing sufficient geometric constraints for subsequent 3D reconstruction. The combination of hardware-synchronized triggering and visual edge computing devices guaranteed the temporal consistency and acquisition stability of the multi-view images.

[0169] Furthermore, compared to traditional mesh-based 3D modeling, which is prone to distortion in complex structures such as feathers, this implementation uses Gaussian splash modeling technology to represent the surface of the goose with a large number of 3D Gaussian distributions. In particular, anisotropic scaling parameters are used for the feathers to accurately reproduce their slender shape. Combined with multi-resolution optimization and multi-view consistency constraints, the surface smoothness and detail reproduction of the model are significantly improved, providing a high-precision geometric foundation for subsequent key feature point localization.

[0170] Finally, based on the three-dimensional coordinates of the identified key feature points, body size parameters were obtained through spatial geometric calculations, with the feature point positioning accuracy error controlled within 0.1 cm. A multi-parameter linear regression correction equation was introduced, and by fusing multiple error factors such as feature point deviation, matching error, and curvature, the measurement accuracy was further improved, correcting the basic accuracy from ±0.2 cm to within ±0.1 cm, fully meeting the body size measurement accuracy requirements of large-scale farms. Simultaneously, the entire measurement process was fully automated, from data acquisition and model reconstruction to parameter calculation and error correction, significantly reducing manpower input and providing effective technical support for smart livestock farming.

[0171] The present invention provides a lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size, and the following implementation methods also exist.

[0172] Step 1: Obtain the multi-source error factor vector of the current measurement sample.

[0173] Obtain intermediate quality indicators generated by the current measurement sample at each processing stage, and generate a multi-source error factor vector based on the intermediate quality indicators.

[0174] The aforementioned intermediate quality indicators are quantifiable evaluation parameters output from each processing stage in the body size measurement process, including the inlier ratio in the feature matching stage, the reprojection error in the sparse reconstruction stage, the convergence loss value in the Gaussian splashing stage, and the confidence score in the semantic segmentation stage.

[0175] The aforementioned feature matching inlier ratio refers to the ratio of the number of matching point pairs that meet the geometric consistency constraint to the total number of matching point pairs during the image feature point matching process. This ratio reflects the image acquisition quality and the reliability of feature extraction.

[0176] The aforementioned reprojection error of sparse reconstruction refers to the average pixel distance between the projected position and the actual detection position after a point in three-dimensional space is projected back onto the two-dimensional image plane. This error reflects the geometric accuracy of the three-dimensional reconstruction.

[0177] The aforementioned convergence loss value of Gaussian splash refers to the loss function value at the end of the 3D Gaussian splash rendering optimization process, which reflects the fitting quality of the surface reconstruction.

[0178] The aforementioned confidence score for semantic segmentation refers to the mean probability output of the semantic segmentation network for classifying the target region. This score reflects the degree of certainty in target recognition.

[0179] The aforementioned multi-source error factor vector is constructed by arranging the above intermediate quality indicators in a predetermined order to form a one-dimensional vector, with each element in the vector corresponding to the quality evaluation results of different processing stages.

[0180] In this embodiment of the application, in order to eliminate the differences in the dimensions and numerical ranges of different quality indicators, each intermediate quality indicator is normalized before constructing the multi-source error factor vector. The min-max normalization method is used to map each indicator to... Interval.

[0181] Step 2: Diagnose the error sources of the multi-source error factor vector and generate error source diagnosis results.

[0182] The multi-source error factor vector is input into the error source diagnosis network for processing. Based on the combination pattern of the error factors, the dominant error type of the current sample is determined, the contribution weight of each error source is output, and the error source diagnosis result is generated.

[0183] The aforementioned error source diagnosis network is a trained multilayer perceptron network. The input layer of this network receives a vector of multi-source error factors, the hidden layer performs a nonlinear mapping of the correlation between error factors, and the output layer generates a contribution weight vector corresponding to the number of predefined error types.

[0184] The number of input layer nodes in the aforementioned multilayer perceptron network is equal to the dimension of the multi-source error factor vector, and the number of output layer nodes is equal to the number of predefined error types. The output layer uses the softmax activation function, which normalizes the sum of each element in the output contribution weight vector to 1.

[0185] The aforementioned multilayer perceptron network is trained in the following manner:

[0186] Input data preparation: Use the multi-source error factor vector of historical samples as input and the error type distribution vector of the sample as label;

[0187] Loss function calculation: The cross-entropy loss function is used to calculate the difference between the network output and the label;

[0188] Parameter optimization: The Adam optimization algorithm is used to update the network parameters and minimize the loss function;

[0189] Iterative training: Repeat the above steps until the loss function converges.

[0190] The error type labeling of the aforementioned training samples is determined by comparing the deviation between the measured value before correction and the manually calibrated true value, combined with the degree of abnormality of the quality indicators at each processing stage.

[0191] The numerical values ​​of each element in the aforementioned contribution weight vector represent the degree of contribution of the corresponding error type to the measurement error of the current sample.

[0192] The aforementioned predefined error types include image acquisition error, geometric reconstruction error, surface fitting error, and target recognition error, corresponding to errors that may be introduced in the four processing stages: feature matching, sparse reconstruction, Gaussian splashing, and semantic segmentation, respectively. The error source diagnosis network determines the dominant error source of the current sample by learning the mapping relationship between error factor combination patterns and error types.

[0193] In this embodiment, the error source diagnosis network further includes an attention mechanism layer, which is located between the hidden layer and the output layer. This layer is used to dynamically adjust the influence weights of different error factors on the diagnosis results, so that the network can focus on the factor features that contribute significantly to the error of the current sample.

[0194] Step 3: Integrate the error source diagnosis results with the classification regression coefficients to generate dynamically adjusted regression coefficients.

[0195] Based on the error source diagnosis results, the regression coefficient group that matches the current error pattern is retrieved from the pre-stored classification regression coefficient library. The multiple regression coefficients are weighted and fused by contribution weight to generate dynamically adjusted regression coefficients.

[0196] The aforementioned classification and regression coefficient library is a pre-built data storage structure that stores regression coefficient sets trained for different error types. Each set of regression coefficients corresponds to a dominant error type, obtained by regression fitting on historical samples with the same dominant error characteristics.

[0197] The aforementioned weighted fusion calculation method involves multiplying the regression coefficient group corresponding to each error type by the corresponding contribution weight in the error source diagnosis results, and then summing the products to obtain dynamically adjusted regression coefficients. This fusion method enables the final regression coefficients to comprehensively reflect the actual contribution of each error source in the current sample.

[0198] The formula for calculating the aforementioned dynamically adjusted regression coefficient is as follows:

[0199]

[0200] in, This is a dynamically adjusted vector of regression coefficients. The number of predefined error types, The first error source diagnosis result Contribution weights for different error types For the first in the classification regression coefficient library The regression coefficient vector corresponding to each error type.

[0201] In this embodiment of the application, in order to avoid the regression coefficients of a single error type having too much influence on the fusion result under extreme contribution weights, the contribution weights are smoothed before weighted fusion, and the concentration of the weight distribution is controlled by a softmax function with temperature coefficient adjustment.

[0202] Step 4: Correct the initial body size measurements and generate the corrected body size parameters.

[0203] A correction equation specific to the current sample is constructed using dynamically adjusted regression coefficients. The initial body size measurements and error factor vector are substituted into the correction equation for calculation, generating the targeted corrected body size parameters.

[0204] The aforementioned modified equation is in the form of a multiple linear regression equation. The independent variables of the equation include the initial body size measurement value and each error factor in the multi-source error factor vector. The dependent variable of the equation is the modified body size parameter, and the coefficients of the equation are dynamically adjusted regression coefficients.

[0205] The formula for calculating the aforementioned corrected equation is as follows:

[0206]

[0207] in, These are the corrected body size parameters. These are the initial body size measurements. The first in the multi-source error factor vector One error factor, The number of error factors, For the intercept term, to These are the coefficient components in the dynamically adjusted regression coefficient vector.

[0208] The aforementioned body size parameters include multiple measurement items such as body length, chest circumference, and neck length. The above correction calculation process is performed for each measurement item to generate the corrected value corresponding to each measurement item.

[0209] In this embodiment, after generating the corrected body size parameters, a rationality check is performed on the correction result. The rationality check includes determining whether the corrected value is within a preset physiologically reasonable range and whether the correction magnitude exceeds a preset maximum correction threshold. When the correction result fails the rationality check, an anomaly flag is output and the initial measurement value is retained.

[0210] Step 5: Incrementally update the classification and regression coefficient library to generate an adaptive evolution correction model.

[0211] Based on step 4, the following steps are also included: storing the error factor vector of the current sample and the body size parameters before and after correction into the online learning cache; triggering incremental updates of the classification and regression coefficient library when the number of samples in the cache reaches a preset threshold; continuously optimizing the classification and regression coefficients through online learning to generate an adaptive evolution correction model.

[0212] The aforementioned online learning cache is a storage area used to temporarily store recent measurement sample data. Each record in the cache contains the multi-source error factor vector of the sample, the error source diagnosis result, the initial body size measurement value, and the corrected body size parameters.

[0213] The aforementioned incremental update is triggered when the number of cached samples reaches a preset threshold, which is set based on the system's storage capacity and update frequency requirements.

[0214] The aforementioned incremental update execution process is as follows:

[0215] Extract sample data from the cache and assign the samples to the corresponding error type groups according to the error source diagnosis results;

[0216] Regression coefficients were refitted for samples within each group, and new regression coefficients were calculated using the least squares method.

[0217] The newly fitted regression coefficients are then weighted and averaged with the original regression coefficients. The calculation formula is as follows:

[0218]

[0219] in, These are the updated regression coefficients. The newly fitted regression coefficients are... These are the original regression coefficients. The learning rate parameter;

[0220] Write the updated regression coefficients into the corresponding entries in the categorical regression coefficient library.

[0221] In this embodiment, to ensure the quality of sample data used for incremental updates, samples are screened for validity before being stored in the online learning cache. The validity screening criteria include: the corrected body size parameters passing a reasonableness check, and the actual body size values ​​of the current sample subsequently receiving manual verification. Only samples that pass the validity screening are stored in the cache and used for incremental updates, avoiding interference from abnormal samples in the regression coefficient optimization process.

[0222] In this embodiment, an online fine-tuning step for the error source diagnosis network is also included. When the accumulated number of valid samples reaches the network fine-tuning threshold, the error source diagnosis network is incrementally trained using the cached sample data to update the network parameters, enabling the error diagnosis capability to adapt to the long-term changing trend of the measurement environment.

[0223] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A lightweight, high-precision three-dimensional model reconstruction and body size measurement method for Anhui white geese, characterized in that, Includes the following steps: Multi-view RGB images of geese were captured by deploying synchronous high-definition cameras at multiple locations in the passageway and using hardware-triggered synchronous shooting to ensure the temporal consistency of the multi-view images. The acquired multi-view RGB images are preprocessed, including image denoising using a two-dimensional Gaussian convolution kernel, distortion correction based on camera intrinsic parameter calibration, brightness normalization, and image size unification. Local feature points in images from different viewpoints are extracted using feature extraction algorithms, feature descriptors are constructed and cross-viewpoint matching is performed to obtain the initial feature correspondence between different viewpoints. Camera pose estimation is performed using feature matching results. The coordinates of the sparse point cloud of camera extrinsic participants are optimized by the bundle adjustment method to generate the initial sparse 3D point cloud on the surface of the goose. Based on camera pose and intrinsic parameters, multi-view stereo matching is performed on multi-view images. The matching cost is calculated using pixel similarity calculation method. Through depth propagation, depth optimization and depth map fusion processing, a dense 3D point cloud covering the surface of the goose is generated. Gaussian splashing technology is applied to model and optimize dense point clouds. A three-dimensional Gaussian function is used to characterize the points in the point cloud and generate a three-dimensional Gaussian distribution to represent the surface details of the goose. The Gaussian splash model is meshed, and the Gaussian representation is converted into a triangular mesh model through surface reconstruction. The isosurface extraction algorithm is then used to generate a structured 3D mesh model. Based on prior knowledge of goose anatomy, a semantic segmentation algorithm is used to identify key anatomical feature points of geese in a 3D mesh model. The feature point coordinates are refined through clustering and curvature analysis to obtain high-precision 3D coordinates of key feature points. Based on the three-dimensional coordinates of the identified key feature points, spatial geometric calculation methods are used to obtain body size parameters, including core body size parameters such as body height, body length, and chest circumference. The predicted body size parameters of the model are corrected by a multi-parameter linear regression equation. The regression equation covers multiple error sources such as feature point positioning deviation, feature matching error, and grid curvature. The least squares method is used to train the fit, so as to achieve high-precision body size measurement.

2. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, Image preprocessing includes: The original image is convolved pixel by pixel using Gaussian weights that are higher at the center than at the periphery to eliminate noise in the image. Using the focal length parameter and principal point coordinates in the camera intrinsic parameter matrix, the image distortion introduced by the lens is compensated by radial and tangential distortion coefficients; Linearly transform the pixel values ​​of the image to a uniform brightness target range; Images are scaled using coordinate mapping to achieve a uniform resolution across all images.

3. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, The feature extraction and matching employs algorithms that include extremum detection of key points in multi-scale scale spaces, construction of feature descriptors, and cross-view feature point correspondence.

4. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, Gaussian splash modeling uses anisotropic Gaussian parameter settings, where the scaling parameter in the major axis direction is used to represent the scale in the direction of the feathers, and the scaling parameter in the minor axis direction is used to represent the scale perpendicular to the direction of the feathers. Complex structural morphology is accurately reproduced through parameter combination.

5. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, Gaussian splash modeling employs a multi-resolution optimization strategy, which involves iterative optimization in multiple stages: global structural optimization in the low-resolution stage; local geometric refinement in the medium-resolution stage; and detail enhancement in the high-resolution stage. This progressive optimization improves model accuracy.

6. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, The Gaussian splash optimization process introduces multi-view consistency constraints, including multi-view reprojection loss constraints and adjacent geometric normal smoothing constraints. By combining the constraints, a 3D model with small geometric discontinuities is generated.

7. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, The linear regression equation corrects the initially predicted body size parameters by considering feature point positioning deviation, feature matching error, and grid curvature factor. The regression coefficients are determined by least squares fitting.

8. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, The mesh wrapping process estimates the point cloud normal vector through neighborhood covariance analysis, constructs a vector field using the normal vector, solves the surface reconstruction equation to obtain the inner and outer indicator functions, and extracts the triangular mesh from the isosurface of the indicator functions.

9. The lightweight method for reconstructing a three-dimensional high-precision model of the Anhui white goose and measuring its body size according to claim 1, characterized in that, Key feature point localization employs semantic segmentation to identify dissected feature points, clusters candidate feature points using a clustering algorithm, and refines coordinates using local curvature calculation.

10. A lightweight, high-precision three-dimensional model reconstruction and body size measurement system for Anhui white geese, used to perform the lightweight, high-precision three-dimensional model reconstruction and body size measurement method for Anhui white geese as described in any one of claims 1 to 9, characterized in that, include: The image acquisition unit includes a synchronous high-definition camera, a hardware triggering device, and a visual edge computing device; The image preprocessing unit performs filtering, distortion correction, brightness adjustment, and size unification processing. The feature matching unit is used to extract feature points and perform cross-view matching. The pose estimation unit is used to calculate camera extrinsic parameters and sparse point cloud reconstruction. Stereo matching units are used to generate dense point clouds; Gaussian splash modeling unit, used for high-fidelity modeling of point clouds; Mesh-wrapping cells are used to generate triangular mesh models; Feature point recognition unit, used to locate anatomical feature points; Body size calculation unit, used to calculate body size parameters; The error correction unit is used to perform multi-parameter linear regression correction.

Images