A method for on-line evaluation of cement pole paste slump

Abstract

The application provides a cement pole slurry slump online evaluation method, and belongs to the online evaluation field.The method comprises the following steps: collecting a slurry surface image sequence and camera parameters, extracting an edge gradient direction distribution, an extreme point high-frequency texture and a spatial position prior, and constructing an initial query feature map; inputting a current frame into a network to obtain a spatial domain feature map, combining a frequency domain mask to generate a frequency domain feature map, and fusing to obtain a dual-domain feature matrix; adjusting an attention receptive field according to a curvature change vector of a last frame mask to obtain a reinforced query feature vector; generating a polar coordinate sampling offset based on an included angle projection of the curvature change vector and a gradient main direction, and resampling to obtain a segmentation mask; calculating a three-dimensional motion vector according to a mask contour and the camera parameters, and combining a geometric weight matrix to output a slump evaluation value.The method is used to improve the stability and reliability of the slurry slump online evaluation.

Description

Technical Field

[0001] This application belongs to the field of online evaluation, and in particular relates to an online evaluation method for the slump of cement pole grout. Background Technology

[0002] Currently, slump testing mainly relies on manual sampling and physical measurement using standard cylinders. This offline testing method is not only labor-intensive and inefficient, but it can also only reflect the state of the slurry at the moment of a single sampling, making it difficult to achieve real-time monitoring in continuous production processes such as mixing, conveying, or spreading. Therefore, it cannot adequately meet the requirements of modern industrial automation for continuous sensing and process control.

[0003] To achieve continuous monitoring, existing technologies have begun to employ online assessment methods based on computer vision. Early vision-based solutions typically utilize basic image processing algorithms to extract slurry contours and estimate the collapse state based on the degree of contour deformation. However, under the influence of complex lighting conditions, slurry splashes, equipment obstructions, and background interference in industrial settings, the contours extracted by basic image processing algorithms are prone to breakage, adhesion, or misjudgment. Furthermore, convolutional neural networks and query vector-based instance segmentation models can identify slurry regions by extracting high-dimensional image features, thus improving the anti-interference capability of visual recognition to a certain extent.

[0004] However, cement slurry is a continuous, non-rigid body with strong rheological properties, and its surface morphology continuously changes during flow, vibration, and collapse. Existing visual models mostly focus on extracting features from a single spatial domain, making it difficult to simultaneously distinguish and utilize high-frequency texture components and low-frequency morphological components in an image. During the slurry flow process, high-frequency features such as surface bubbles and particle textures mix with the overall contour distribution. Without frequency domain filtering and a dual-domain fusion mechanism of spatial and frequency domains, segmentation models are prone to insufficient accuracy when processing blurred edges and complex texture regions.

[0005] Furthermore, existing video stream-based segmentation methods often fail to adequately utilize the physical geometry of the slurry during inter-frame information transfer. The evolution of slurry motion in three-dimensional space is frequently accompanied by changes in surface curvature and gradient direction deflection. However, conventional attention mechanisms and mask sampling strategies often employ fixed receptive field configurations and conventional orthogonal coordinate sampling methods, making it difficult to fully utilize geometric prior information such as curvature parameters from the previous frame and edge gradient directions from the current frame to adjust the receptive field or guide feature resampling. This results in potential discrepancies between the segmentation mask and the actual physical boundaries of the slurry. These discrepancies may be further amplified during subsequent three-dimensional spatial mapping, affecting the reliability of three-dimensional motion vector calculations and consequently impacting the slump assessment results derived from motion features.

[0006] Therefore, how to integrate spatial and frequency domain features and combine them with inter-frame geometric prior information to improve the stability and boundary fit of continuous slurry mask segmentation, thereby realizing online evaluation of the slump of cement pole slurry, is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the problems of existing technologies where cement pole slump testing relies on manual sampling, making continuous online monitoring difficult, and visual segmentation is susceptible to complex textures, blurred edges, and inter-frame geometric changes, leading to insufficient reliability in assessment, this invention provides an online assessment method for cement pole slump, the method comprising: Acquire the image sequence of the slurry surface and camera parameters; calculate the edge gradient direction distribution of the current frame, and initialize the extreme points based on the high frequency response of the current frame when the current frame is the first frame, and determine the inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame; extract the spatial location prior and high frequency texture components based on the edge gradient direction distribution and extreme points to obtain the initial query feature map. The current frame is input into the network to obtain a spatial feature map; when it is the first frame, the frequency domain mask is initialized; otherwise, a frequency domain mask is constructed based on the three-dimensional motion vector of the previous frame; the frequency domain feature map is obtained by filtering based on the frequency domain mask, and then fused with the spatial feature map to obtain a dual-domain feature matrix; the curvature change vector of the mask of the previous frame is obtained, and the initial query feature map and the dual-domain feature matrix are input into the attention module. The receptive field is adjusted according to the magnitude of the curvature change vector and cross-attention is calculated to obtain the enhanced query feature vector. The polar coordinate sampling offset is obtained by calculating the projection value of the angle between the curvature change vector and the main gradient direction of the current frame. Based on the offset, the enhanced query feature vector is resampled to obtain the segmentation mask of the current frame. The contour point set of the segmentation mask is extracted and mapped to three-dimensional space based on camera parameters to calculate the three-dimensional motion vector. A geometric weight matrix is ​​constructed based on the magnitude of the curvature change vector through dimensionless mapping. The three-dimensional motion vector is weighted and solved based on the geometric weight matrix to obtain the collapse evaluation value.

[0008] According to another aspect of the present invention, an online slump assessment system for cement pole grout is provided, the system comprising the following modules: The extraction module is used to acquire the image sequence of the slurry surface and camera parameters; calculate the edge gradient direction distribution of the current frame, and initialize the extreme points based on the high frequency response of the current frame when the current frame is the first frame, and determine the inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame; extract the spatial location prior and high frequency texture components based on the edge gradient direction distribution and extreme points to obtain the initial query feature map. The computation module is used to input the current frame into the network to obtain a spatial feature map; initialize a frequency domain mask when it is the first frame, otherwise construct a frequency domain mask based on the three-dimensional motion vector of the previous frame; filter the frequency domain mask to obtain a frequency domain feature map, and fuse it with the spatial feature map to obtain a dual-domain feature matrix; obtain the curvature change vector of the mask of the previous frame, input the initial query feature map and the dual-domain feature matrix into the attention module, adjust the receptive field according to the magnitude of the curvature change vector and calculate cross-attention to obtain the enhanced query feature vector; The evaluation module is used to calculate the polar coordinate sampling offset by the projection value of the angle between the curvature change vector and the main gradient direction of the current frame; to resample the enhanced query feature vector based on the offset to obtain the segmentation mask of the current frame; to extract the contour point set of the segmentation mask and to calculate the three-dimensional motion vector by mapping it to three-dimensional space based on camera parameters; to construct a geometric weight matrix based on the magnitude of the curvature change vector through dimensionless mapping; and to obtain the collapse evaluation value by weighting the three-dimensional motion vector based on the geometric weight matrix.

[0009] This disclosure has the following advantages: By acquiring image sequences of the slurry surface and camera parameters, and extracting edge gradient direction distribution, high-frequency texture of extreme points, and spatial location priors, a rich initial feature base can be provided for slurry region identification. By constructing a frequency domain mask by combining three-dimensional motion vectors and fusing it with spatial domain features, the feature expression capability under complex textures and blurred edges can be enhanced. Furthermore, by using curvature change vectors to adjust the attention receptive field and performing feature resampling based on polar coordinate sampling offset, the segmentation results can better fit the slurry deformation boundary. By mapping the mask contour to three-dimensional space and calculating weighted motion feature quantities, a more stable calculation basis can be provided for online slump assessment. Attached Figure Description

[0010] Figure 1 A flowchart for an online assessment method of slump in cement pole grout; Figure 2 This is a histogram of edge gradient directions. Figure 3 This is a schematic diagram of the sampling points for the segmentation mask contour. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0012] In this application, an online evaluation method for the slump of cement pole grout is provided, such as... Figure 1 As shown, the method includes: S1, acquire the image sequence of the slurry surface and camera parameters; calculate the edge gradient direction distribution of the current frame, and initialize the extreme point based on the high frequency response of the current frame when the current frame is the first frame, and determine the inter-frame extreme point based on the current frame and the previous frame when the current frame is not the first frame, and extract the spatial location prior and high frequency texture components based on the edge gradient direction distribution and extreme points to obtain the initial query feature map.

[0013] An industrial camera is installed above a centrifugal molding machine for cement poles to capture continuous image sequences of the slurry surface. The camera's internal and external parameter matrices are obtained using a checkerboard calibration method. After acquiring the original image of the current frame, the original image is cropped and scaled according to a preset region of interest in the slurry area to obtain a uniformly sized image to be processed. The cropping position and scaling ratio are recorded.

[0014] The current frame image to be processed is converted into a grayscale image. The gradient magnitude and gradient direction of each pixel are calculated using the Sobel operator. The range from 0 to 360° is divided into multiple angle intervals, and the edge gradient direction distribution of the current frame is statistically obtained.

[0015] The centroid of two-dimensional coordinates is calculated based on the pixel points corresponding to the angle intervals with the highest frequency in the edge gradient direction distribution, and the spatial position prior is generated by normalization. If the current frame is the first frame, high-frequency response is extracted from the grayscale image of the current frame, and pixels with high-frequency responses greater than a preset high-frequency threshold are used as initial extreme points; if the current frame is not the first frame, inter-frame difference is performed between the current frame and the previous frame, and pixels with local maxima whose absolute difference value is greater than a preset threshold are used as inter-frame extreme points.

[0016] A two-dimensional query grid of a preset size is established on the image to be processed, with each grid cell corresponding to a fixed query position. For any grid cell, the position code of its center coordinates is obtained, and the initial extrema or inter-frame extrema falling within the grid cell or its neighborhood are counted. If an extrema exists, the high-frequency texture components of the corresponding local image patch are extracted and averaged. If no extrema exists, the high-frequency texture components of the grid cell are set as a preset zero vector.

[0017] The spatial location prior corresponding to each grid cell is concatenated with the high-frequency texture component and then input into a multilayer perceptron with shared parameters to obtain the corresponding initial query vector. All initial query vectors are arranged according to the order of the two-dimensional query grid, forming an initial query feature map with a fixed number and fixed spatial order.

[0018] In some implementations, the calculation of the edge gradient direction distribution of the current frame, and the initialization of extreme points based on the high-frequency response of the current frame when the current frame is the first frame, and the determination of inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame, and the extraction of spatial location priors and high-frequency texture components based on the edge gradient direction distribution and extreme points to obtain an initial query feature map, including: The Sobel operator is applied to the current frame to calculate the gradient magnitude and gradient direction of each pixel. The gradient direction is divided into multiple angle intervals. The pixel frequency of each angle interval is counted to obtain the edge gradient direction distribution histogram. All pixels corresponding to the interval with the highest frequency are extracted and the centroid of the two-dimensional spatial coordinates is calculated. The centroid of the two-dimensional spatial coordinates is used as the spatial position prior. Determine whether the current frame is the first frame. If it is the first frame, extract the pixels with a frequency greater than the preset high frequency threshold based on the high frequency response map of the current frame as the initial extreme points. If it is not the first frame, then perform inter-frame difference between the current frame and the previous frame, and extract the pixels whose absolute difference value is greater than a preset threshold as inter-frame extreme points. Based on a preset two-dimensional query grid, the local high-frequency texture features corresponding to the initial extreme points or inter-frame extreme points are aggregated in a grid to obtain the high-frequency texture components corresponding to each query grid unit. The position code corresponding to the spatial location prior is concatenated with the high-frequency texture component, and mapped through a multilayer perceptron to obtain an initial query feature map arranged according to the preset two-dimensional query grid.

[0019] The original image of the slurry surface with a current frame size of 1920×1080 is acquired. Based on the preset region of interest (ROI) of the slurry area, it is cropped and scaled to a size of 1024×1024 for processing. A 3×3 Sobel operator is applied for convolution along both the horizontal and vertical directions to obtain the horizontal and vertical gradients at the corresponding pixel locations. The gradient magnitudes and gradient direction angles (greater than or equal to 0° and less than 360°) are calculated using these two gradient components.

[0020] The angle from 0 to 360° is divided into 36 angular intervals, with each interval being 10°. The frequency of pixels within each interval is counted to construct a histogram of edge gradient direction distribution. If the 5th interval (40° to 50°) has the highest frequency and contains 3450 pixels with strong gradients, the two-dimensional coordinates of all pixels within this interval are extracted. The centroid coordinates are calculated by averaging, for example, with an x-coordinate of 945 and a y-coordinate of 521. A 256-dimensional sine and cosine position code is generated using this centroid coordinate as a spatial prior. The statistical results of the edge gradient directions obtained through actual calculation are as follows: Figure 2 As shown.

[0021] If the current frame is the first frame, a two-dimensional discrete Fourier transform is performed on the grayscale image of the current frame. Low-frequency components within a preset radius of the spectrum center are set to zero. Then, an inverse two-dimensional discrete Fourier transform is performed, and the amplitude is taken to obtain the high-frequency response image of the current frame. Pixels with high-frequency responses greater than a preset high-frequency threshold are extracted as initial extreme points. The preset radius is the radius of the low-frequency center region determined with the spectrum center point as the center, preferably 5% to 20% of the short side pixel length of the spectrum image. The preset high-frequency threshold is preferably the sum of the mean pixel amplitude of the high-frequency response image and 1.5 to 3 times the standard deviation. If the current frame is not the first frame, the grayscale values ​​of the corresponding pixels in the current frame and the previous frame are subtracted bit by bit, and the absolute value is taken. The absolute value is determined to be greater than a preset threshold, preferably 15 to 40 grayscale levels.

[0022] A 64×64 query grid is established on the 1024×1024 image to be processed, with each query grid cell corresponding to a 16×16 pixel area. For any query grid cell, the initial extreme points or inter-frame extreme points falling within that grid cell and its surrounding preset neighboring grid cells are counted. The preset neighboring grid cells are preferably 3×3 neighboring grid cells centered on the current grid cell.

[0023] Centered on each extreme point, a local image patch of a preset size is extracted from the grayscale image of the current frame. The preset size is preferably 16×16 pixels. A two-dimensional discrete Fourier transform is performed on the local image patch to obtain a local frequency domain amplitude spectrum. Amplitude data located in a preset high-frequency region within the local frequency domain amplitude spectrum are extracted. The mean, variance, maximum value, and energy sum of the amplitude data are calculated to form a local high-frequency statistical vector. This local high-frequency statistical vector is then mapped to a 128-dimensional high-frequency texture feature through a fully connected mapping layer with shared parameters. When the same query grid cell corresponds to multiple extreme points, the arithmetic mean of the multiple 128-dimensional high-frequency texture features is calculated to obtain the 128-dimensional high-frequency texture component corresponding to the query grid cell. When no extreme point is corresponding, the high-frequency texture component of the query grid cell is set as a 128-dimensional preset zero vector.

[0024] In the channel dimension of the features, the 256-dimensional spatial location prior vector corresponding to each query grid cell is concatenated with the 128-dimensional high-frequency texture component vector to form a 384-dimensional concatenated feature vector. This 384-dimensional concatenated feature vector is input into a multilayer perceptron with shared parameters. The multilayer perceptron consists of three fully connected layers connected sequentially, with the dimensionality mapping relationships of 384-512, 512-256, and 256-128 respectively, and the ReLU function is used as the activation function. The multilayer perceptron is applied to each grid cell in the 64×64 query grid, outputting a 128-dimensional initial query vector corresponding to each grid cell. The outputs of all grid cells are arranged in grid order to form a 64×64×128 initial query feature map.

[0025] S2, input the current frame into the network to obtain the spatial domain feature map; initialize the frequency domain mask when it is the first frame, otherwise construct the frequency domain mask based on the three-dimensional motion vector of the previous frame; filter the frequency domain mask to obtain the frequency domain feature map, and fuse it with the spatial domain feature map to obtain the dual-domain feature matrix.

[0026] The three-channel red-green-blue image of the current frame is input into a residual convolutional neural network with a ResNet-50 backbone and a feature pyramid structure to extract multi-scale features. Specifically, the 1024×1024×3 image to be processed is normalized in pixel value and then input into the ResNet-50 backbone network. It passes through convolutional layers, batch normalization layers, activation function layers, pooling layers, and multiple residual convolutional stages to obtain residual feature maps at different spatial scales. A 1×1 convolution is applied to each scale of the residual feature map to unify the number of channels. The low-resolution feature map is then upsampled level by level and fused with adjacent high-resolution feature maps to form a feature pyramid fusion result. The fused feature map is further upsampled to 1024×1024 and then integrated and compressed through 3×3 and 1×1 convolutional layers to output a spatial feature map of size 1024×1024×64.

[0027] When the current frame is the first frame, a global pass-through matrix with the same frequency domain size as the image to be processed and all elements having a value of 1 is generated as the initial frequency domain mask. When the current frame is not the first frame, the three-dimensional motion vector of the previous frame is obtained, and a frequency domain mask is constructed based on its magnitude and projection direction in the image plane. If the magnitude is less than or equal to a preset minimum positive number, the global pass-through matrix is ​​used; if the magnitude is greater than the preset minimum positive number, a cutoff frequency parameter is determined based on the magnitude within a preset cutoff frequency range, and an anisotropic angle parameter is determined based on the projection direction, so that the frequency domain mask can suppress irrelevant directional noise while preserving the main morphological and texture responses related to the motion direction.

[0028] A two-dimensional discrete Fourier transform is performed on the grayscale image to be processed in the current frame. The resulting frequency domain data is multiplied element-wise with the frequency domain mask, and then restored to the spatial domain by a two-dimensional inverse discrete Fourier transform to obtain a frequency domain feature map. The frequency domain feature map and the spatial domain feature map are concatenated along the channel dimension, and a 1×1 convolutional layer is used for channel dimensionality reduction and feature interaction to obtain a dual-domain feature matrix.

[0029] In some implementations, the initialization of the frequency domain mask when it is the first frame, and the construction of the frequency domain mask based on the three-dimensional motion vector of the previous frame, includes: Determine if the current frame number is the first frame. If so, generate a global pass matrix with all elements being preset constants as the initial frequency domain mask. If not, the magnitude and direction of the three-dimensional motion vector of the previous frame are obtained. When the magnitude is less than or equal to a preset minimum positive number, the frequency domain mask uses the global pass-through matrix. When the magnitude is greater than the preset minimum positive number, the cutoff frequency parameter is adaptively adjusted within a preset cutoff frequency range according to the magnitude, and the anisotropic angle parameter of the frequency domain passband is determined according to the direction. An anisotropic Gaussian low-pass filter with direction selectivity is constructed by combining the Gaussian kernel function with the cutoff frequency parameter and the anisotropic angle parameter, and the frequency domain response matrix of the anisotropic Gaussian low-pass filter is used as the constructed frequency domain mask.

[0030] The timestamp and frame number information in the video stream are obtained to determine whether the current frame is the first frame. If the frame number of the current frame is 0, it is considered the first frame image, and a global pass-through matrix with the same frequency domain size as the original image, such as a length and width of 1024×1024, is generated. Each element in this matrix is ​​initialized to a fixed preset constant of 1.0, thus forming the initial frequency domain mask.

[0031] When the frame number is greater than 0, the three-dimensional motion vector calculated in the previous frame is obtained. If the three-dimensional motion vector is the zero vector obtained from the initialization of the first frame, or its magnitude is less than or equal to a preset minimum positive number, then the global pass matrix is ​​used.

[0032] If the projection components of the 3D motion vector on each axis of the 3D coordinate system are 0.15, -0.08, and 0.05 respectively, then the magnitude of the 3D motion vector is calculated to be 0.177 m / s. The 3D motion vector is transformed to the camera coordinate system according to the rotation matrix in the camera's external parameters, and its two components in the image plane are extracted to form a 2D projection vector. The projection direction angle is calculated based on this 2D projection vector, for example, a projection direction angle of 32 degrees. A preset cutoff frequency lower limit is set to 80 frequency domain pixel units, a preset cutoff frequency upper limit is set to 140 frequency domain pixel units, and a reference motion speed is set to 0.35 m / s. The ratio of the magnitude of the 3D motion vector to the reference motion speed is limited to between 0 and 1, serving as a frequency adjustment coefficient. This frequency adjustment coefficient is then multiplied by the difference between the preset cutoff frequency upper and lower limits and added to the preset cutoff frequency lower limit to obtain the current adaptive cutoff frequency parameter, for example, a cutoff frequency parameter of approximately 111 frequency domain pixel units.

[0033] An anisotropic Gaussian low-pass filter is constructed based on the cutoff frequency parameter and the projection direction angle. Specifically, a frequency domain coordinate system is established within a 1024×1024 Fourier coordinate grid, with the center point of the spectrum as the origin. For any frequency domain coordinate point, its coordinates are rotated according to the projection direction angle to obtain a first frequency domain coordinate component along the direction of motion and a second frequency domain coordinate component perpendicular to the direction of motion. The cutoff frequency parameter is used as the first directional standard deviation, and the product of the first directional standard deviation and a preset scaling factor is used as the second directional standard deviation. The preset scaling factor is preferably 0.4 to 0.8, and more preferably 0.6. The first frequency domain coordinate component, the second frequency domain coordinate component, the first directional standard deviation, and the second directional standard deviation are substituted into a two-dimensional Gaussian function to calculate the response value corresponding to the frequency domain coordinate point. The response value decreases as the distance from the frequency domain coordinate point to the center point of the spectrum increases, and it has different attenuation rates in the direction of motion and perpendicular to the direction of motion.

[0034] After traversing all frequency domain coordinate points, a 1024×1024 frequency domain response matrix is ​​obtained, and this frequency domain response matrix is ​​used as the frequency domain mask of the current frame.

[0035] In some implementations, obtaining a frequency domain feature map based on the frequency domain mask filtering and fusing it with a spatial domain feature map to obtain a dual-domain feature matrix includes: Perform a two-dimensional fast Fourier transform on the current frame to obtain the frequency domain complex matrix of the current frame; The frequency domain complex matrix is ​​multiplied by the frequency domain mask to obtain the filtered complex matrix. Perform a two-dimensional inverse fast Fourier transform on the filtered complex matrix and take the real part as the frequency domain feature map; The frequency domain feature map and the spatial domain feature map are concatenated along the channel dimension, and a 1×1 convolutional layer is used for channel dimensionality reduction and feature interaction to obtain a dual-domain feature matrix.

[0036] A single-channel grayscale image of the current frame to be processed, with a size of 1024×1024 pixels, is acquired. A two-dimensional Fast Fourier Transform (FFT) is applied to convert the discrete image to the frequency domain, and the corresponding two-dimensional complex array is output as a frequency domain complex matrix. The 1024×1024 anisotropic frequency domain mask generated in the previous step is extracted. The floating-point weight values ​​corresponding to the frequency domain mask are multiplied pixel-by-pixel at the same position with the frequency domain complex matrix, and the filtered complex matrix is ​​output.

[0037] A two-dimensional inverse fast Fourier transform is performed on the filtered complex matrix. Due to potential errors in the imaginary part of the calculation, the real part is used as the frequency domain feature map for a single channel. A 1024×1024 spatial feature map with 64 channels, output from the spatial backbone network, is extracted. The frequency domain feature map and the spatial feature map are then concatenated along the channel direction to generate a multidimensional feature tensor with a depth of 65 layers.

[0038] A 1×1 convolutional layer with 64 feature kernels, a kernel size of 1×1, and a stride of 1 is deployed to process the multidimensional feature tensor. This 1×1 convolutional layer is used for cross-channel feature interaction. After channel dimensionality reduction and feature interaction, a 64-channel dual-domain feature matrix is ​​obtained.

[0039] S3, obtain the curvature change vector of the mask of the previous frame, input the initial query feature map and the dual-domain feature matrix into the attention module, adjust the receptive field according to the magnitude of the curvature change vector and calculate cross attention to obtain the enhanced query feature vector.

[0040] The discrete curvature of each pixel on the segmentation mask contour of the previous frame is calculated using the second-order center difference algorithm. The curvature difference between adjacent boundary points is obtained, and the curvature difference is multiplied by the two-dimensional unit normal vector of the corresponding boundary point. The mean of the calculation results for each boundary point is calculated to obtain the two-dimensional curvature change vector.

[0041] The initial query feature map is used as the query tensor of the attention module, and the dual-domain feature matrix is ​​mapped by different linear projection matrices as the key tensor and value tensor, respectively.

[0042] The sum of squares of the elements of the curvature change vector is calculated and its square root is taken to obtain the Euclidean norm as the modulus. This modulus is input into the Sigmoid activation function, mapping to obtain coefficients between 0 and 1. These coefficients are multiplied by a preset base dilation rate to obtain the dilation factor. A base sampling template for deformable convolution is pre-set, comprising multiple two-dimensional sampling offset coordinates distributed around the convolution center. The dilation factor is multiplied by each sampling offset coordinate in the base sampling template to obtain the two-dimensional sampling grid offset corresponding to each sampling point. The deformable convolution is configured using this two-dimensional sampling grid offset to enlarge or shrink the receptive field window of the key tensor on the dual-domain feature matrix.

[0043] The dot product of the query tensor and the key tensor after adjusting the receptive field is calculated, divided by the square root of the feature dimension as a scaling factor, and processed by a normalized exponential function to obtain the attention weight matrix. The attention weight matrix is ​​then multiplied by the value tensor and fed into a multilayer perceptron consisting of layer normalization and a feedforward neural network for processing, outputting the enhanced query feature vector set.

[0044] In some implementations, obtaining the curvature change vector of the previous frame mask, inputting the initial query feature map and the dual-domain feature matrix into the attention module, adjusting the receptive field according to the magnitude of the curvature change vector and calculating cross-attention to obtain the enhanced query feature vector includes: Extract the discrete boundary point set of the previous frame mask, calculate the first and second derivatives of each boundary point, calculate the curvature of the boundary point based on the first and second derivatives, calculate the curvature difference between adjacent boundary points and multiply it by the boundary point normal vector, and calculate the mean of the calculation results for each boundary point to obtain the curvature change vector. If it is the first frame, assign the curvature change vector to the zero vector. The initial query feature map is used as the query feature, and the dual-domain feature matrix is ​​used as the key feature and value feature; Based on the magnitude of the curvature change vector, a sampling scale coefficient for deformable convolution is generated through an activation function. Based on the sampling scale coefficient, the receptive field is adaptively sampled and adjusted on the key features and value features. The query feature vector is obtained by aligning the dimensions of the query feature with the sampled key feature through a linear projection layer, calculating the inner product, performing a normalized exponential function operation, and then weighting and summing the product with the sampled value feature.

[0045] A tracking strategy is used to obtain the discrete boundary point set of the mask from the previous frame. For example, a set of consecutive pixels includes the horizontal coordinate. =230 and the ordinate =455 to the x-coordinate =236 and the ordinate =461 points, calculate the first and second derivatives of the boundary points at each node. Substitute them into the curvature formula to calculate the curvature at a single point, for example, the curvature of a certain control point is 0.038.

[0046] If the current frame is the first frame of the video, then all elements of the curvature change vector are assigned zero. If the current frame is not the first frame of the video, then the closed edge contour of the segmentation mask of the previous frame is read, and the curvature value of each contour point is calculated sequentially along the closed edge contour. For any contour point, the curvature difference between the contour point and its adjacent boundary points is calculated, for example, the curvature difference is 0.004; at the same time, the two adjacent contour points before and after the contour point are obtained, and the two-dimensional tangent vector at the contour point is calculated based on the coordinate difference between the two adjacent contour points. The two-dimensional tangent vector is then rotated by 90 degrees and normalized to obtain the two-dimensional unit normal vector corresponding to the contour point.

[0047] The curvature difference is used as a weighting coefficient and multiplied by the two-dimensional unit normal vector corresponding to the contour point to obtain the normal curvature change contribution vector corresponding to the contour point. All contour points in the closed edge contour are traversed, and the normal curvature change contribution vectors corresponding to each contour point are accumulated and averaged to obtain the curvature change vector used for processing the current frame. For example, the two-dimensional components of the curvature change vector are 0.003 and -0.001, respectively.

[0048] An initial query feature map of 64×64×128 is used as the query feature for the attention mechanism, and a 1024×1024×64-dimensional dual-domain feature matrix is ​​used as the key and value features. The architecture of the attention module includes deformable convolutional layers for spatial sampling, linear fully connected layers for feature mapping, and normalized exponential weighted layers for weight allocation.

[0049] Obtain the curvature change vector with components of 0.003 and -0.001, and calculate its magnitude to be 0.00316. Input the magnitude into a preset activation function and multiply it by the base dilation coefficient to obtain the sampling scale coefficient of the deformable convolution; wherein, the preset activation function is preferably the Sigmoid function, and the base dilation coefficient is preferably 3.6, thereby obtaining a sampling scale coefficient of, for example, 1.8.

[0050] A preset 3×3 basic sampling template is used, comprising nine two-dimensional offset coordinates with the convolution center as the origin: (-1,-1), (-1,0), (-1,1), (0,-1), (0,0), (0,1), (1,-1), (1,0), and (1,1). The sampling scale coefficient is multiplied by each of the nine two-dimensional offset coordinates to obtain the floating sampling offset coordinates corresponding to each sampling point, ensuring that each sampling point is distributed within the range of -1.8 to 1.8 pixels. Based on these floating sampling offset coordinates, bilinear interpolation sampling is performed on the dual-domain feature matrix.

[0051] The 128-dimensional query features are reduced to 64 dimensionality using linear projection, and matrix inner product similarity is calculated after aligning with the key feature dimensions. The results are then normalized using Softmax and weighted and aggregated with the sampled and corrected value features to obtain a 64×64×64 enhanced query feature map, where each query grid position corresponds to a 64-dimensional enhanced query feature vector.

[0052] S4. Calculate the polar coordinate sampling offset by the projection value of the angle between the curvature change vector and the main gradient direction of the current frame. Based on the offset, resample the enhanced query feature vector to obtain the segmentation mask of the current frame.

[0053] Gradient calculation is performed on the grayscale image to be processed in the current frame to obtain the horizontal gradient component, vertical gradient component, gradient magnitude and gradient direction of each pixel; pixels located outside the preset width range of the image boundary and whose gradient magnitude is greater than a preset gradient threshold are determined as valid pixels, wherein the preset gradient threshold is preferably the sum of the average gradient magnitude of all pixels in the current frame and a preset multiple standard deviation.

[0054] The gradient covariance matrix is ​​calculated based on the gradient components of all valid pixels, and the gradient covariance matrix is ​​decomposed into eigenvalues. The eigenvector corresponding to the largest eigenvalue is taken as the main gradient direction of the current frame.

[0055] When the current frame is the first frame of the video, or when the magnitude of the curvature change vector is less than or equal to a preset minimum positive number, the polar coordinate sampling offset is set to zero. When the current frame is not the first frame of the video and the magnitude of the curvature change vector is greater than the preset minimum positive number, the absolute value of the cosine similarity between the curvature change vector and the gradient principal direction vector of the current frame is calculated and used as the included angle projection value. This included angle projection value is then input into the offset prediction network, which is a multi-layer feedforward fully connected neural network. The input is a single included angle projection value, and the output is the radial offset distance and the angular offset angle parameters, which are then combined to form the polar coordinate sampling offset.

[0056] A two-dimensional polar coordinate system is constructed with the center of the feature map grid corresponding to the enhanced query feature map as the pole. The polar radius and polar angle of the original polar coordinate grid are superimposed with the radial offset distance and angular offset angle respectively to obtain the offset polar coordinate sampling points. The polar coordinate sampling points are converted into rectangular coordinate sampling points, and the corresponding feature values ​​are extracted on the enhanced query feature map by bilinear interpolation to complete the polar coordinate resampling.

[0057] The resampled feature map is input into a decoder network containing deconvolutional and classification layers, outputting a probability map of each pixel belonging to the slurry region. The Otsu method is used to determine the binarization threshold and generate the current frame segmentation mask. Based on the cropping position and scaling ratio recorded in S1, the current frame segmentation mask is restored to the original image coordinate system.

[0058] In some implementations, the calculation of the angle projection between the curvature change vector and the main gradient direction of the current frame yields a polar coordinate sampling offset. Based on this offset, the enhanced query feature vector is resampled to obtain the segmentation mask for the current frame, including: Calculate the covariance matrix of the gradients of all valid pixels in the current frame, perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the largest eigenvalue as the main gradient direction of the current frame. When the current frame is the first frame, or when the magnitude of the curvature change vector is less than or equal to a preset minimum positive number, the polar coordinate sampling offset is set to zero offset. When the current frame is not the first frame and the magnitude of the curvature change vector is greater than a preset minimum positive number, the absolute value of the cosine similarity between the curvature change vector and the main gradient direction of the current frame is calculated as the included angle projection value. The included angle projection value is then input into the offset prediction network, which outputs a unified radial offset distance and a unified angular offset angle, which are then combined into a polar coordinate sampling offset. The enhanced query feature vector set is arranged according to a preset two-dimensional query grid to form an enhanced query feature map. Based on the bilinear interpolation algorithm, the enhanced query feature map is spatially resampled according to the polar coordinate sampling offset. The spatial resolution is restored through the decoding network to obtain the segmentation mask of the current frame.

[0059] Gradient calculation is performed on the grayscale image to be processed in the current frame to obtain the gradient components of each pixel in the horizontal and vertical directions, and the corresponding gradient magnitude and gradient direction are calculated based on the gradient components. Pixels located outside the preset width range of the image boundary and whose gradient magnitude is greater than a preset gradient threshold are determined as valid pixels; wherein, the preset gradient threshold is preferably the sum of the average gradient magnitude of all pixels in the current frame and a preset multiple of the standard deviation.

[0060] Calculate the 2×2 covariance matrix of the gradients of all valid pixels in the current image, for example, with diagonal elements of 68.5 and 24.2. Perform eigenvalue decomposition on this matrix to extract the largest eigenvalue, such as 72.1, and the corresponding eigenvector, for example, with components of 0.866 and 0.500. Use this eigenvector as the principal direction of the gradient in the current frame.

[0061] A curvature change vector is obtained. If the magnitude of the curvature change vector is less than or equal to a preset minimum positive number, the subsequent polar coordinate sampling offset is set to zero. If the magnitude of the curvature change vector is greater than the preset minimum positive number, the absolute value of its cosine similarity with the main gradient direction of the current frame is calculated (e.g., a value of 0.82), and this value is used as the included angle projection value and input into the offset prediction network. The offset prediction network is a multi-layer feedforward fully connected neural network, with a single included angle projection value as input and radial offset distance and angular offset angle parameters as output.

[0062] The network outputs a radial offset distance of 3.5 pixels and an angular offset angle of 0.22 radians, which are combined into a polar coordinate sampling offset. The enhanced query feature vector set arranged in a 64×64 two-dimensional query grid is used as a two-dimensional grid feature map, such as a size of 64×64×64.

[0063] Based on the polar coordinate sampling offset, the feature map is spatially resampled using a bilinear interpolation algorithm. An upsampling restoration process is then performed through a spatial decoding network, which includes multiple transposed 2D convolutional layers, batch normalization layers, and a classification output layer. The decoding network restores the full resolution, obtaining a binary segmentation mask adapted to the size of the image to be processed. This mask is then mapped back to the original image size (1920×1080) based on the recorded cropping position and scaling ratio.

[0064] S5. Extract the contour point set of the segmentation mask, map it to three-dimensional space based on camera parameters, and calculate the three-dimensional motion vector. Construct a geometric weight matrix based on the magnitude of the curvature change vector through dimensionless mapping, and solve the three-dimensional motion vector by weighting based on the geometric weight matrix to obtain the collapse evaluation value.

[0065] Morphological edge detection is performed on the current frame segmentation mask to extract the coordinates of boundary pixels whose mask values ​​change from 0 to 1, forming a contour point set. For any contour point in the contour point set, the corresponding depth value is obtained by querying or interpolating based on the pre-established slurry reference surface model. Combined with the camera's internal and external parameter matrices, the two-dimensional pixel coordinates of the contour point are inversely projected to the three-dimensional coordinate system according to the pinhole camera imaging geometry model to obtain the corresponding three-dimensional spatial coordinates. The actual distribution of discrete sampling points of the extracted segmentation mask closed contour is as follows: Figure 3 As shown.

[0066] The chamfer distance matching algorithm is used to search for the nearest corresponding point pair between the contour points of the current frame and the contour points of the previous frame in three-dimensional space. The three-dimensional space coordinates of the current frame are subtracted from the corresponding three-dimensional space coordinates of the previous frame to obtain multiple coordinate difference vectors. The average value of the multiple coordinate difference vectors is calculated and divided by the inter-frame time interval to obtain the representative three-dimensional motion vector corresponding to the current frame.

[0067] Obtain the magnitude of the aforementioned curvature change vector and normalize it by dividing it by a preset curvature reference value to obtain a curvature representation quantity; input the curvature representation quantity into a preset monotonic mapping function to generate multiple curvature-related weight coefficients, and construct a third-order positive semi-definite geometric weight matrix using the multiple curvature-related weight coefficients as the main diagonal elements.

[0068] The representative three-dimensional motion vectors are weighted using the third-order semi-definite geometric weight matrix to calculate their weighted norm under this matrix, and this weighted norm is used as a weighted motion feature. The weighted motion feature is then input into a pre-fitted polynomial regression model obtained from experimental data to output the slump assessment value of the cement pole grout.

[0069] In some implementations, the extraction of the contour point set of the segmentation mask and the calculation of the 3D motion vector based on the camera parameters mapped to 3D space include: An edge detection algorithm is applied to extract the contour point set of the segmentation mask, and the two-dimensional pixel coordinates of each contour point are obtained. Based on the camera parameters, a perspective projection matrix from the real three-dimensional space to the two-dimensional image plane is constructed, and a preset slurry reference surface model is used as a depth constraint condition. Based on the perspective projection matrix and the slurry reference surface model, the depth value corresponding to each two-dimensional pixel coordinate is determined, and combined with the inverse matrix of the camera's internal parameter matrix, the two-dimensional pixel coordinates are inversely projected to the three-dimensional coordinate system to obtain a three-dimensional spatial point set. If the current frame is the first frame, the three-dimensional spatial point set is stored and the three-dimensional motion vector is initialized to a zero vector; if the current frame is not the first frame, a matching point pair between the three-dimensional spatial point set and the corresponding three-dimensional spatial point set of the previous frame is established through the nearest neighbor matching algorithm, the coordinate difference vector of each matching point pair is calculated and the mean is calculated, and the mean is divided by the inter-frame time interval to obtain the three-dimensional motion vector.

[0070] For each frame, the contour point set of the binary segmentation mask of the current frame is extracted, and the 3D spatial point set of the current frame is obtained based on camera parameters and a slurry reference surface model. Specifically, the Canny edge detection operator or morphological edge extraction algorithm is used to obtain the pixel coordinates of the mask boundary, forming a contour point set containing multiple positioning points, for example, the contour point set contains approximately 1000 positioning points; then, based on the 2D pixel coordinates of the contour points, the corresponding depth values, and the inverse matrix of the camera's internal parameter matrix, the contour points are inversely projected onto the 3D coordinate system to obtain the 3D spatial point set of the current frame. If the current frame is the first frame of the video, the 3D spatial point set of the current frame is stored, and the three directional components of the 3D motion vector are initialized to 0; if the current frame is not the first frame of the video, the 3D spatial point set stored in the previous frame is read, and nearest neighbor matching is performed with the 3D spatial point set of the current frame.

[0071] For any location point in the contour point set, obtain its two-dimensional pixel coordinates, for example, x=852 and y=641. Based on the lateral focal length, longitudinal focal length, optical center position, and rotation and translation parameters representing the camera's mounting attitude from the camera parameters, construct a 3×4 perspective projection matrix that maps the three-dimensional spatial coordinates to the two-dimensional image coordinates.

[0072] The slurry reference surface model is pre-established based on camera calibration results, camera installation posture, initial calibration contour of the slurry area, and geometric parameters of the cement pole mold. Specifically, before the start of each production batch, a spatial mapping relationship between the camera coordinate system and the cement pole mold coordinate system is established, the initial slurry surface contour and mold geometric parameters are obtained, the boundary range of the slurry reference surface is determined based on the initial slurry surface contour, and the height field of the slurry reference surface is fitted in the mold coordinate system in combination with the mold geometric parameters and the preset initial slurry height; the height field is used to represent the depth value corresponding to different planar positions on the reference surface.

[0073] A three-dimensional reference point mesh is established based on the height field of the slurry reference surface, and each three-dimensional reference point is projected onto a two-dimensional image plane using the perspective projection matrix to obtain the coordinates of each projected pixel and its corresponding depth value in the camera coordinate system. A depth lookup table corresponding to the image coordinates is generated based on the projected pixel coordinates and the corresponding depth value, and the depth lookup table is used as the slurry reference surface model. When a contour point falls between the meshes of the lookup table, the depth value corresponding to the contour point is obtained through bilinear interpolation, and the depth lookup table is recalibrated at the beginning of each production batch.

[0074] Based on the two-dimensional pixel coordinates of the contour points, the corresponding depth values, and the inverse matrix of the camera's internal parameter matrix, the contour points are inversely projected onto the three-dimensional coordinate system according to the pinhole camera imaging geometric model to obtain the three-dimensional spatial point set of the current frame. For example, the coordinate components of one of the three-dimensional spatial points are 0.12, 0.58, and -0.21.

[0075] The nearest neighbor matching algorithm is used to establish matching point pairs between the current frame's 3D spatial point set and the previous frame's 3D spatial point set. For each matching point pair, the corresponding 3D spatial coordinates of the previous frame are subtracted from the current frame's 3D spatial coordinates to obtain a coordinate difference vector. For example, the components of the coordinate difference vector are 0.012, 0.045, and -0.003. Multiple coordinate difference vectors are summed and averaged to obtain a representative inter-frame displacement vector. This representative inter-frame displacement vector is then divided by the video frame interval to obtain the 3D motion vector corresponding to the current frame. For example, when the video frame rate is 30 frames / second, the frame interval is 0.0333 seconds, and the corresponding components of the 3D motion vector are 0.36, 1.35, and -0.09 meters / second.

[0076] In some implementations, the step of constructing a geometric weight matrix based on the magnitude of the curvature change vector through a dimensionless mapping, and then weighting the three-dimensional motion vector based on the geometric weight matrix to obtain the slump assessment value, includes: Curvature-related weight coefficients are generated by dimensionless mapping using the magnitude of the curvature change vector, and a third-order positive semi-definite weight matrix is ​​constructed using the curvature-related weight coefficients. The third-order positive semi-definite weight matrix is ​​then used as the geometric weight matrix. Calculate the weighted norm of the three-dimensional motion vector under the geometric weight matrix to obtain the weighted motion feature quantity; The weighted motion feature is input into a pre-fitted slump conversion model, wherein the slump conversion model is a multinomial regression model constructed based on historical calibration data, and outputs the slump evaluation value of the cement pole grout corresponding to the current frame.

[0077] Obtain the magnitude of the curvature change vector and normalize it by dividing it by a preset curvature reference value to obtain the curvature representation quantity κ, for example, κ is 0.0032. Substitute the curvature representation quantity κ into a preset exponential decay mapping function and calculate the curvature-related weight coefficients by combining them with the directional decay coefficients corresponding to the three directions; wherein, the three directional decay coefficients are, for example, 60, 45 and 80 respectively. For example, according to... Calculate the weighting coefficients for each direction, where, The directional attenuation coefficients for the corresponding directions are used to obtain three curvature-related weighting coefficients, for example, 0.825, 0.866, and 0.774. These three values ​​are placed on the main diagonal of a 3×3 matrix, and the remaining positions are set to zero. Since all curvature-related weighting coefficients are non-negative, the diagonal matrix constructed using these coefficients as its main diagonal elements is a third-order positive semi-definite geometric weighting matrix. This constructed 3×3 third-order positive semi-definite matrix is ​​then used as the geometric weighting matrix.

[0078] Read the 3D motion vector, for example, components of 0.18, 0.45, and -0.06. Apply the aforementioned geometric weight matrix to calculate the weighted norm of the 3D motion vector under the geometric weight matrix. After summing the squares of each component, if the combined result is 0.2045, take the square root to obtain the weighted motion characteristic quantity, for example, 0.452.

[0079] The polynomial regression model is pre-constructed based on over 500 sets of measured standard slump calibration data within the past three months. For example, the model formula is: Substituting the obtained weighted motion characteristic quantity 0.452 into the above polynomial regression model, we obtain the corresponding slump evaluation value, such as 111.002. After numerical normalization, we obtain the cement pole slump evaluation value of 111 mm.

[0080] During the model training phase, historical calibration samples were pre-collected, including slurry surface image sequences, manually labeled slurry region masks, corresponding camera calibration parameters, and measured standard slump values. The manually labeled slurry region masks were used as segmentation supervision labels to train the residual convolutional backbone network, attention module, offset prediction network, and decoder network; the measured standard slump values ​​were used as regression supervision labels to fit the slump conversion model. During training, the segmentation part used cross-entropy loss or Dice loss, the offset prediction part used the absolute error loss between the predicted sampling offset and the labeled boundary offset, where the labeled boundary offset was calculated from the nearest point deviation between the manually labeled slurry region mask boundary and the predicted mask boundary, and the slump conversion model used the mean squared error loss between the predicted slump and the measured standard slump value. After training, the network parameters and slump conversion model parameters were saved for use in the online evaluation phase.

[0081] This application discloses an online slump assessment system for cement pole grout, the system comprising: The extraction module is used to acquire the image sequence of the slurry surface and camera parameters; calculate the edge gradient direction distribution of the current frame, and initialize the extreme points based on the high frequency response of the current frame when the current frame is the first frame, and determine the inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame; extract the spatial location prior and high frequency texture components based on the edge gradient direction distribution and extreme points to obtain the initial query feature map. The computation module is used to input the current frame into the network to obtain a spatial feature map; initialize a frequency domain mask when it is the first frame, otherwise construct a frequency domain mask based on the three-dimensional motion vector of the previous frame; filter the frequency domain mask to obtain a frequency domain feature map, and fuse it with the spatial feature map to obtain a dual-domain feature matrix; obtain the curvature change vector of the mask of the previous frame, input the initial query feature map and the dual-domain feature matrix into the attention module, adjust the receptive field according to the magnitude of the curvature change vector and calculate cross-attention to obtain the enhanced query feature vector; The evaluation module is used to calculate the polar coordinate sampling offset by the projection value of the angle between the curvature change vector and the main gradient direction of the current frame; to resample the enhanced query feature vector based on the offset to obtain the segmentation mask of the current frame; to extract the contour point set of the segmentation mask and to calculate the three-dimensional motion vector by mapping it to three-dimensional space based on camera parameters; to construct a geometric weight matrix based on the magnitude of the curvature change vector through dimensionless mapping; and to obtain the collapse evaluation value by weighting the three-dimensional motion vector based on the geometric weight matrix.

[0082] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for online evaluation of the slump of cement pole grout, characterized in that, include: Acquire the image sequence of the slurry surface and camera parameters; calculate the edge gradient direction distribution of the current frame, and initialize the extreme points based on the high frequency response of the current frame when the current frame is the first frame, and determine the inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame; extract the spatial location prior and high frequency texture components based on the edge gradient direction distribution and extreme points to obtain the initial query feature map. Input the current frame into the network to obtain the spatial domain feature map; initialize the frequency domain mask when it is the first frame, otherwise construct the frequency domain mask based on the three-dimensional motion vector of the previous frame; The frequency domain feature map is obtained based on the frequency domain masking filter, and then fused with the spatial domain feature map to obtain a dual-domain feature matrix. Obtain the curvature change vector of the mask in the previous frame, input the initial query feature map and the dual-domain feature matrix into the attention module, adjust the receptive field according to the magnitude of the curvature change vector and calculate cross attention to obtain the enhanced query feature vector; The polar coordinate sampling offset is obtained by calculating the projection value of the angle between the curvature change vector and the main gradient direction of the current frame. Based on the offset, the enhanced query feature vector is resampled to obtain the segmentation mask of the current frame. The contour point set of the segmentation mask is extracted and mapped to three-dimensional space based on camera parameters to calculate the three-dimensional motion vector. A geometric weight matrix is ​​constructed based on the magnitude of the curvature change vector through dimensionless mapping. The three-dimensional motion vector is weighted and solved based on the geometric weight matrix to obtain the collapse evaluation value.

2. The method according to claim 1, characterized in that, The calculation of the edge gradient direction distribution of the current frame, and the initialization of extreme points based on the high-frequency response of the current frame when the current frame is the first frame, and the determination of inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame, and the extraction of spatial location prior and high-frequency texture components based on the edge gradient direction distribution and extreme points to obtain an initial query feature map, including: The Sobel operator is applied to the current frame to calculate the gradient magnitude and gradient direction of each pixel. The gradient direction is divided into multiple angle intervals. The pixel frequency of each angle interval is counted to obtain the edge gradient direction distribution histogram. All pixels corresponding to the interval with the highest frequency are extracted and the centroid of the two-dimensional spatial coordinates is calculated. The centroid of the two-dimensional spatial coordinates is used as the spatial position prior. Determine whether the current frame is the first frame. If it is the first frame, extract the pixels with a frequency greater than the preset high frequency threshold based on the high frequency response map of the current frame as the initial extreme points. If it is not the first frame, then perform inter-frame difference between the current frame and the previous frame, and extract the pixels whose absolute difference value is greater than a preset threshold as inter-frame extreme points. Based on a preset two-dimensional query grid, the local high-frequency texture features corresponding to the initial extreme points or inter-frame extreme points are aggregated in a grid to obtain the high-frequency texture components corresponding to each query grid unit. The position code corresponding to the spatial location prior is concatenated with the high-frequency texture component, and mapped through a multilayer perceptron to obtain an initial query feature map arranged according to the preset two-dimensional query grid.

3. The method according to claim 1, characterized in that, The process of initializing the frequency domain mask when it is the first frame, and otherwise constructing the frequency domain mask based on the three-dimensional motion vector of the previous frame, includes: Determine if the current frame number is the first frame. If so, generate a global pass matrix with all elements being preset constants as the initial frequency domain mask. If not, the magnitude and direction of the three-dimensional motion vector of the previous frame are obtained. When the magnitude is less than or equal to a preset minimum positive number, the frequency domain mask uses the global pass-through matrix. When the magnitude is greater than the preset minimum positive number, the cutoff frequency parameter is adaptively adjusted within a preset cutoff frequency range according to the magnitude, and the anisotropic angle parameter of the frequency domain passband is determined according to the direction. An anisotropic Gaussian low-pass filter with direction selectivity is constructed by combining the Gaussian kernel function with the cutoff frequency parameter and the anisotropic angle parameter, and the frequency domain response matrix of the anisotropic Gaussian low-pass filter is used as the constructed frequency domain mask.

4. The method according to claim 3, characterized in that, The process of obtaining a frequency domain feature map based on the frequency domain mask filtering and fusing it with the spatial domain feature map to obtain a dual-domain feature matrix includes: Perform a two-dimensional fast Fourier transform on the current frame to obtain the frequency domain complex matrix of the current frame; The frequency domain complex matrix is ​​multiplied by the frequency domain mask to obtain the filtered complex matrix. Perform a two-dimensional inverse fast Fourier transform on the filtered complex matrix and take the real part as the frequency domain feature map; The frequency domain feature map and the spatial domain feature map are concatenated along the channel dimension, and a 1×1 convolutional layer is used for channel dimensionality reduction and feature interaction to obtain a dual-domain feature matrix.

5. The method according to claim 1, characterized in that, The process of obtaining the curvature change vector of the previous frame mask, inputting the initial query feature map and the dual-domain feature matrix into the attention module, adjusting the receptive field according to the magnitude of the curvature change vector and calculating cross-attention to obtain the enhanced query feature vector includes: Extract the discrete boundary point set of the previous frame mask, calculate the first and second derivatives of each boundary point, calculate the curvature of the boundary point based on the first and second derivatives, calculate the curvature difference between adjacent boundary points and multiply it by the boundary point normal vector, and calculate the mean of the calculation results for each boundary point to obtain the curvature change vector. If it is the first frame, assign the curvature change vector to the zero vector. The initial query feature map is used as the query feature, and the dual-domain feature matrix is ​​used as the key feature and value feature; Based on the magnitude of the curvature change vector, a sampling scale coefficient for deformable convolution is generated through an activation function. Based on the sampling scale coefficient, the receptive field is adaptively sampled and adjusted on the key features and value features. The query feature vector is obtained by aligning the dimensions of the query feature with the sampled key feature through a linear projection layer, calculating the inner product, performing a normalized exponential function operation, and then weighting and summing the product with the sampled value feature.

6. The method according to claim 1, characterized in that, The calculation of the angle projection between the curvature change vector and the main gradient direction of the current frame yields a polar coordinate sampling offset. Based on this offset, the enhanced query feature vector is resampled to obtain the segmentation mask for the current frame, including: Calculate the covariance matrix of the gradients of all valid pixels in the current frame, perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the largest eigenvalue as the main gradient direction of the current frame. When the current frame is the first frame, or when the magnitude of the curvature change vector is less than or equal to a preset minimum positive number, the polar coordinate sampling offset is set to zero offset. When the current frame is not the first frame and the magnitude of the curvature change vector is greater than a preset minimum positive number, the absolute value of the cosine similarity between the curvature change vector and the main gradient direction of the current frame is calculated as the included angle projection value. The included angle projection value is then input into the offset prediction network, which outputs a unified radial offset distance and a unified angular offset angle, which are then combined into a polar coordinate sampling offset. The enhanced query feature vector set is arranged according to a preset two-dimensional query grid to form an enhanced query feature map. Based on the bilinear interpolation algorithm, the enhanced query feature map is spatially resampled according to the polar coordinate sampling offset. The spatial resolution is restored through the decoding network to obtain the segmentation mask of the current frame.

7. The method according to claim 1, characterized in that, The extraction of the contour point set of the segmentation mask, and the calculation of the 3D motion vector based on the camera parameters mapped to 3D space, includes: An edge detection algorithm is applied to extract the contour point set of the segmentation mask, and the two-dimensional pixel coordinates of each contour point are obtained. Based on the camera parameters, a perspective projection matrix from the real three-dimensional space to the two-dimensional image plane is constructed, and a preset slurry reference surface model is used as a depth constraint condition. Based on the perspective projection matrix and the slurry reference surface model, the depth value corresponding to each two-dimensional pixel coordinate is determined, and combined with the inverse matrix of the camera's internal parameter matrix, the two-dimensional pixel coordinates are inversely projected to the three-dimensional coordinate system to obtain a three-dimensional spatial point set. If the current frame is the first frame, the three-dimensional spatial point set is stored and the three-dimensional motion vector is initialized to a zero vector; if the current frame is not the first frame, a matching point pair between the three-dimensional spatial point set and the corresponding three-dimensional spatial point set of the previous frame is established through the nearest neighbor matching algorithm, the coordinate difference vector of each matching point pair is calculated and the mean is calculated, and the mean is divided by the inter-frame time interval to obtain the three-dimensional motion vector.

8. The method according to claim 1, characterized in that, The process involves constructing a geometric weight matrix based on the magnitude of the curvature change vector using a dimensionless mapping, and then weighting the three-dimensional motion vectors using this geometric weight matrix to obtain the collapse assessment value, including: Curvature-related weight coefficients are generated by dimensionless mapping using the magnitude of the curvature change vector, and a third-order positive semi-definite weight matrix is ​​constructed using the curvature-related weight coefficients. The third-order positive semi-definite weight matrix is ​​then used as the geometric weight matrix. Calculate the weighted norm of the three-dimensional motion vector under the geometric weight matrix to obtain the weighted motion feature quantity; The weighted motion feature is input into a pre-fitted slump conversion model, wherein the slump conversion model is a multinomial regression model constructed based on historical calibration data, and outputs the slump evaluation value of the cement pole grout corresponding to the current frame.

9. An online evaluation system for the slump of cement pole grout, characterized in that, include: The extraction module is used to acquire the image sequence of the slurry surface and camera parameters; calculate the edge gradient direction distribution of the current frame, and initialize the extreme points based on the high frequency response of the current frame when the current frame is the first frame, and determine the inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame; extract the spatial location prior and high frequency texture components based on the edge gradient direction distribution and extreme points to obtain the initial query feature map. The computation module is used to input the current frame into the network to obtain the spatial domain feature map; when it is the first frame, the frequency domain mask is initialized; otherwise, the frequency domain mask is constructed based on the three-dimensional motion vector of the previous frame. The frequency domain feature map is obtained based on the frequency domain masking filter, and then fused with the spatial domain feature map to obtain a dual-domain feature matrix. Obtain the curvature change vector of the mask in the previous frame, input the initial query feature map and the dual-domain feature matrix into the attention module, adjust the receptive field according to the magnitude of the curvature change vector and calculate cross attention to obtain the enhanced query feature vector; The evaluation module is used to calculate the polar coordinate sampling offset by the projection value of the angle between the curvature change vector and the main gradient direction of the current frame; to resample the enhanced query feature vector based on the offset to obtain the segmentation mask of the current frame; to extract the contour point set of the segmentation mask and to calculate the three-dimensional motion vector by mapping it to three-dimensional space based on camera parameters; to construct a geometric weight matrix based on the magnitude of the curvature change vector through dimensionless mapping; and to obtain the collapse evaluation value by weighting the three-dimensional motion vector based on the geometric weight matrix.

10. The system according to claim 9, characterized in that, The calculation of the edge gradient direction distribution of the current frame, and the initialization of extreme points based on the high-frequency response of the current frame when the current frame is the first frame, and the determination of inter-frame extreme points based on the current frame and the previous frame when the current frame is not the first frame, and the extraction of spatial location prior and high-frequency texture components based on the edge gradient direction distribution and extreme points to obtain an initial query feature map, including: The Sobel operator is applied to the current frame to calculate the gradient magnitude and gradient direction of each pixel. The gradient direction is divided into multiple angle intervals. The pixel frequency of each angle interval is counted to obtain the edge gradient direction distribution histogram. All pixels corresponding to the interval with the highest frequency are extracted and the centroid of the two-dimensional spatial coordinates is calculated. The centroid of the two-dimensional spatial coordinates is used as the spatial position prior. Determine whether the current frame is the first frame. If it is the first frame, extract the pixels with a frequency greater than the preset high frequency threshold based on the high frequency response map of the current frame as the initial extreme points. If it is not the first frame, then perform inter-frame difference between the current frame and the previous frame, and extract the pixels whose absolute difference value is greater than a preset threshold as inter-frame extreme points. Based on a preset two-dimensional query grid, the local high-frequency texture features corresponding to the initial extreme points or inter-frame extreme points are aggregated in a grid to obtain the high-frequency texture components corresponding to each query grid unit. The position code corresponding to the spatial location prior is concatenated with the high-frequency texture component, and mapped through a multilayer perceptron to obtain an initial query feature map arranged according to the preset two-dimensional query grid.

Images