58 results about "Laplace operator" patented technology

The invention relates to the technical field of image processing, in particular to a lens edge detection method and system based on image generation, and the method comprises the following steps: collecting a lens image through an industrial camera, carrying out the weighted graying, extracting an edge coordinate through a Laplace operator, constructing an edge pixel coordinate set, and carrying out the image processing; screening an edge validity marking result by combining a gray difference threshold, calculating a phase difference between a gradient direction and light source incidence based on a Sobel operator, extracting pixels in a consistent direction, carrying out spatial clustering, generating a lens edge pixel cluster, and fitting a continuous curve to construct a complete lens edge contour. Noise and reflection interference are restrained through pixel neighborhood gray difference, effective pixels are screened in combination with a gradient direction and light source incident phase relation, edge direction consistency is enhanced, pixels are clustered according to spatial continuity and gradient intensity, contours are continuously fitted based on pixel cluster distribution, and lens edge integrity and geometric consistency under complex illumination are improved.

The invention relates to the technical field of machine vision detection, in particular to an automobile part surface defect online detection method and system. The method comprises the following steps: acquiring a grayscale image, calculating the size of a structural element for each pixel based on a local Gaussian Laplacian operator response variance, filtering to obtain a substrate image according to the size of the structural element, and differentiating to obtain a texture image. Determining a Gabor scale and a gray-level co-occurrence matrix statistical direction by using the size, and extracting a cooperative direction gray-level co-occurrence matrix feature; and a weight is set based on the size and is subjected to weighted fusion with a multi-scale rotation invariant local binary pattern feature to generate a texture saliency map, and texture defects are judged. On the substrate image, taking the gray value as the height, and determining a neighborhood calculation curvature feature based on the size to detect the substrate defect. According to the scheme, the image scale can be adaptively analyzed, the background texture is effectively inhibited, and therefore different types of tiny defects such as scratches and pits can be reliably detected.

The invention discloses a feature point detection method based on edge saliency and scale sensitivity. The method comprises the following steps of performing Gaussian smooth denoising on an input image; calculating the edge saliency of the pixel points through a Laplace operator; calculating a texture change degree based on the local gradient magnitude; generating a scale sensitivity weighted value through multi-scale analysis; candidate points are screened in combination with edge saliency, texture weighting and scale weighted values, and mismatching points are removed through non-maximum suppression; and finally, key points are enhanced and marked. According to the method, the edge structure, the texture information and the multi-scale features are fused, the robustness and accuracy of feature point detection are improved, and the method is particularly suitable for image matching and target recognition tasks in complex scenes.

The invention relates to the technical field of intelligent printing quality detection, in particular to an intelligent printing quality detection method based on image recognition. The system comprises an image preprocessing module, a physical model construction module, a deformation iterative calculation module, an energy decoupling analysis module and a closed-loop feedback control module. The method comprises the steps of generating a feature point matrix and a to-be-detected feature field by acquiring standard and real-time image data; the core of the method is to construct a virtual tensor spring network model, map image differences into an external acting force field, and solve deformation potential energy distribution by using finite element iteration; a Laplace operator is adopted to execute energy decoupling, low-frequency isotropic stress and high-frequency topological potential energy singular points are separated, and a residual thermodynamic diagram is generated according to the singular points to position defects and adjust transmission tension; according to the method, the transformation from pure image comparison to physical potential energy evolution is realized, the mutual exclusion limitation of deformation tolerance and detection sensitivity is broken through, and the nonlinear deformation interference in the high-speed transmission of the flexible material can be effectively filtered.

The present application relates to the technical field of simulation system, specifically to a virtual battlefield scene simulation system and method thereof, the system comprises: terrain node identification module, combat unit control radius extraction module, terrain advantage area preliminary demarcation module, control force gradient evolution analysis module, advantage area boundary dynamic adjustment module.In the present application, through the joint screening based on elevation, visible penetration rate and surface roughness, the key nodes with cover and accessibility are accurately identified, the terrain selection accuracy is improved, the spatial distance between the advantage nodes and the combat units is calculated to realize the accurate association between the control range and the deployment, the isolated points are excluded by node clustering division, the complete control domain contour is constructed, the Laplace operator is introduced to calculate the control force change rate, the boundary is dynamically adjusted in response to the situation evolution, the spatial continuity and change logic of the control force expression are strengthened, and the scene adaptability and deduction accuracy in the simulation process are enhanced.

The invention belongs to the technical field of building boundary segmentation, and provides a remote sensing image building boundary segmentation method based on a U-net network model, and the method comprises the steps: extracting a gradient magnitude image through a Sobel operator, generating a binary boundary mask, dynamically adjusting the weight of a convolution kernel in combination with a gradient direction, and enhancing the feature extraction capability of a boundary region; boundary artifacts output by cavity convolution are suppressed through a high-frequency residual boundary mask, an artifact area is restored in combination with gradient direction constraint, and original detail features are reserved; calculating complexity scores based on boundary density and tortuosity, adaptively selecting voidage, and balancing global context and local detail capture capability; extracting a high-frequency component through a Laplace operator, marking an artifact point by combining a dynamic threshold value, and repairing an abnormal region by utilizing a neighborhood gradient direction; the method solves the performance bottleneck of a traditional method in a boundary fuzzy, artifact interference and receptive field fixed scene, and is suitable for remote sensing image analysis tasks such as urban planning and disaster assessment.

The invention discloses a variable grid method grid automatic division method and device based on vector field guidance. The method comprises the following steps: extracting boundary feature points according to geometric features of an input model boundary so as to construct a feature polygon of a model, and dispersing the feature polygon into a triangular mesh model; constructing a vector field based on a Laplacian operator by using the triangular mesh model; starting from the geometric center of the model, setting an initial quadrilateral grid as an initial paving layer, paving the quadrilateral grid outwards layer by layer, and if the paving layer exceeds the boundary of the model, stopping paving; when each layer of quadrilateral grid deforms, uniform transition is achieved through a size transition template; and finally, generating a boundary quadrilateral mesh of the model by adopting a mapping method, mapping discrete points on the boundary of the feature polygon back to the boundary of the input model, and outputting a generated quadrilateral mesh model. According to the method, full-automatic generation of the quadrilateral mesh can be realized, and the method is particularly suitable for mesh generation in finite element simulation pretreatment.

The invention relates to a farmland intelligent identification method of an interactive double-branch network based on deep learning, and belongs to the technical field of remote sensing image intelligent analysis and farmland information extraction. The method is composed of two mutually interactive branches which are focused on farmland feature extraction and boundary feature extraction respectively, Gaussian filtering and a Laplace operator are introduced in the boundary feature extraction branch, a double attention module is designed, fine-grained farmland features and boundary features are fused, and a fine-grained farmland feature and boundary feature fusion model is obtained. And the perceptual ability of the model to the farmland boundary is enhanced. In a farmland feature extraction branch, a hierarchical visual Transform structure is adopted, and a channel-space grouping is used for enhancing an attention module so as to improve the expression ability of farmland semantic features; meanwhile, a boundary semantic fusion module is designed to introduce boundary detail information into high-level semantic features, so that the recognition effect is further optimized, finally, comprehensive utilization of different scale features is achieved through a pyramid fusion module, and the recognition capacity of the model for diversified farmlands is effectively improved. According to the method, the description capability and the overall recognition precision of the model on farmland details are effectively improved.

This invention discloses a method for locating abnormally overheating components on circuit boards based on thermal images. Starting from the characteristics of the thermal image itself, it obtains difference images under normal and abnormal conditions, and performs image registration using an accelerated robust feature registration algorithm optimized based on the Laplacian operator. This ensures that the difference images under abnormal and normal conditions are well aligned after registration, resulting in a clear differential thermal image. Then, outlier detection is performed on the differential thermal image using the Laplacian distribution, obtaining a differential thermal image after outlier detection, thus initially identifying abnormally overheating regions and locating the abnormally overheating components. Finally, kernel density estimation is used to determine the number of K-means clusters, and the K-means algorithm is used to segment the differential thermal image after outlier detection, obtaining a K-means segmented image. Morphological opening-closing operations are then performed on the image to remove minute lines. The thermal image segmentation results of this invention can accurately provide information on abnormally overheating components, enabling accurate location and identification of these components, helping to narrow down the defect range and providing defect judgment information.

The invention relates to a mark point hole filling method, device and equipment. The method comprises the steps of obtaining boundary point data of a target hole in the surface of a three-dimensional model, wherein the target hole is a hole formed by scanning mark points; filling the target hole according to the boundary point data of the target hole, and determining a triangular mesh and / or a triangular filling vertex used for filling; according to the triangular mesh and / or the triangular filling vertex, updating the initial topological structure of the three-dimensional model to obtain the topological structure of the updated mesh data corresponding to the three-dimensional model; and if the target hole is filled by using the triangular filling vertex, optimizing and adjusting the position of the triangular filling vertex in the topological structure of the updated grid data by using a Laplace operator linear system so as to obtain the grid data of the three-dimensional model after hole filling. By adopting the method, the hole filling effect of the marked points can be achieved.

This provides a method for denoising seismic profiles based on an improved Laplacian filter and reverse time migration. [Solution] A method for denoising earthquake profiles based on an improved Laplacian filter and reverse time migration, comprising the steps of: collecting earthquake data and obtaining imaging results after reverse time migration processing; calculating the difference of the second derivatives of the Laplace operator in the x and z directions, displaying them in 8-neighbor second derivative form, decomposing this into four Laplacian filters including two relative directions to obtain an improved Laplacian filter; and using the improved Laplacian filter relating only to the second derivative in the z-axis direction, performing noise reduction by filtering on the imaging results after reverse time migration processing.

The invention provides an online visual detection method and device for flaws of high-speed moving cloth, and relates to the technical field of cloth flaw detection.The online visual detection method comprises the steps that firstly, a response diagram is obtained through time-sharing exposure, a normalized surface gradient modulus is calculated through a normalized difference ratio and logarithm compression, and color albedo interference is eliminated; curvature features are extracted by calculating the absolute value of a Hessian matrix and the square of a Laplacian operator and combined to generate a mixed curvature tensor, and tiny deformation is accurately captured; then manifold unsupervised reconstruction is carried out based on a Gaussian weighted model of a hollowed-out center pixel to obtain an ideal reconstruction tensor, and sample-free adaptive prediction is realized; difference is calculated, a local signal-to-noise ratio abnormal response index is obtained through local statistical window standard deviation normalization, and texture roughness interference is removed; and finally, positioning flaws, carrying out double-threshold judgment by calculating a flaw structure characteristic value and a flaw oil stain characteristic value, and outputting flaw coordinates and categories.

The invention provides a rapid detection method and system for parasites in fresh food, and the method comprises the steps: collecting a multispectral image of the fresh food, selecting near-infrared and green light channels, calculating a normalized differential value, generating a parasite stress feature map, and calculating the global variance of the feature map; edge texture features are extracted through a Laplacian operator, and the original spectrogram and the two feature maps are stacked and input into a deep network; the network intermediate layer divides the features into a core group and an auxiliary group, the core group generates a channel vector, and the auxiliary group performs discrete cosine transform compression according to global variance; splicing the compression feature and the core feature to obtain a depth feature vector; meanwhile, carrying out weighted fusion on the multispectral image by utilizing a channel vector, calculating a gradient covariance matrix and flattening the gradient covariance matrix to obtain a spectral statistical vector; and splicing the depth features and the spectral statistical vectors, outputting a detection score through a classifier, and comparing the detection score with a threshold value to obtain a detection conclusion.

The application discloses a kind of semi-supervised medical image segmentation method and system, it is related to image segmentation technical field, method includes: through the foreground probability graph of student network output, calculate its three-dimensional gradient amplitude and the space average weighted sum of discrete three-dimensional Laplace operator, obtain global boundary roughness scalar;Combining voxel level entropy and soft boundary indicator constructs roughness perception consistency loss, realizes adaptive course learning;Drive scaling factor to carry out residual modulation to geometric contrast loss, enhance the structure discriminant of boundary region;Guide boundary gradient alignment, curvature smoothing and pseudo-supervised threshold, through the weighted joint optimization student network of three, teacher network is updated with exponential moving average.The application introduces global boundary roughness scalar, cooperates and controls entropy consistency, geometric boundary contrast and gradient-curvature-pseudo-supervised joint regularization, significantly improves the boundary precision and robustness of medical image segmentation under semi-supervised condition.

The invention discloses a tunnel water leakage detection method based on refrigeration type infrared thermal imaging. The method comprises the following steps: obtaining a temperature value two-dimensional matrix; traversing each part of original temperature field data, and extracting a feature threshold value; calculating a second-order gradient field by adopting an eight-direction Laplacian operator; extracting a second-order temperature gradient characteristic average value of the low-temperature domain; selecting a second-order temperature gradient upper limit and a second-order temperature gradient lower limit as leakage water quantitative screening criteria based on temperature field second-order gradient characteristics; selecting data meeting quantitative screening criteria, and preliminarily rejecting non-leakage data; for the data meeting the quantitative screening criterion, dividing dry and wet areas of the data by taking the dry and wet boundary temperature as a temperature contour line, and drawing; data which cannot be eliminated through a single temperature field identification mode and contains a regular-shape non-leakage area are eliminated; and outputting the data information containing the leakage water and a corresponding leakage water segmentation result image. The method has the advantage that the tunnel leakage water disease identification accuracy is improved.

The invention discloses a color image vision enhancement method. The method comprises the following steps: acquiring an original color image and separating three color components; a Laplace operator is used to sharpen each component image, a Sobel operator is used to extract horizontal and vertical gradients respectively, the gradients are transformed based on a Gaussian function and normalization processing is completed, natural logarithm operation is carried out on an original component image and a normalized gradient image respectively, two obtained logarithm domain images are added and then exponent recovery is carried out, and a final component image is obtained. Restoring the component image to a gray domain, and performing Sigmoid function transformation on the restored component image; and calculating the global maximum and minimum values of the Sigmoid enhanced images of the three component images, performing linear stretching, and finally combining the stretched color component images to obtain a visual enhanced color image. According to the invention, the permeability of the color image can be improved, the brightness of the color image under the condition of low illumination can be improved, and the color effect matched with the visual habit of human eyes can be kept.

The invention discloses a semi-supervised medical image segmentation method and system, and relates to the technical field of image segmentation, and the method comprises the steps: calculating the spatial average weighted sum of the three-dimensional gradient magnitude and a discrete three-dimensional Laplacian operator through a foreground probability graph outputted by a student network, and obtaining a global boundary roughness scalar; the voxel-level entropy and the soft boundary indicator are combined to construct the consistency loss of roughness perception, and adaptive course learning is realized; carrying out residual modulation on geometric contrast loss by driving a scaling factor, and enhancing the structural discrimination force of a boundary region; boundary gradient alignment, curvature smoothing and a pseudo-supervision threshold are guided, and the student network and the teacher network are optimized through weighted joint of the three so as to perform index moving average updating. According to the method, the global boundary roughness scalar is introduced, entropy consistency, geometric boundary comparison and gradient-curvature-pseudo supervision joint regularization are cooperatively regulated, and the boundary precision and robustness of medical image segmentation are remarkably improved under the semi-supervised condition.

This invention discloses an underwater target saliency detection method and system based on polarization multi-stage fusion, belonging to the field of computer vision and underwater image processing technology. The method acquires four-angle polarization images of turbid water and calculates the degree of linear polarization, then normalizes them to generate a polarization-enhanced image. Subsequently, high-frequency edge components are extracted through brightness enhancement and the Laplacian operator, and a preprocessed image is obtained by injecting learnable weights. A pre-trained deep neural network is then used to fuse the original image and the preprocessed image, outputting a saliency prediction map. This invention fully exploits polarization features, effectively suppresses scattering noise, and significantly improves the precision of underwater target edge segmentation.

This invention discloses a lidar depth completion method based on hierarchical minimum surface reconstruction, relating to the fields of computer vision and image processing technology. The method includes: performing depth value inversion and morphological dilation on a sparse depth map to obtain a scene dilated depth map and a full-resolution effective mask; then downsampling to obtain a coarse-scale scene dilated map and a coarse-scale mask; using an iterative convolution kernel of the Laplacian operator to obtain a coarse-scale filled map; upsampling the coarse-scale filled map to obtain an upsampled depth map, and then fusing it with the scene dilated depth map to obtain a fused depth map; locating all remaining holes in the fused depth map using connected component analysis, and calculating the mean depth of the effective pixels within the annular pixel band around each remaining hole; using this mean depth to fill the corresponding remaining holes to obtain a hole-free depth map; and after global Gaussian blurring, performing a depth value inversion operation to obtain a scene dense depth map, thereby achieving high-precision 3D reconstruction.

The application relates to the field of intelligent monitoring, and discloses a device defect detection method based on infrared identification, which comprises the steps of infrared thermal imaging data acquisition, noise reduction processing, nonlinear heat conduction modeling, high-order partial differential optimization and variational regularization inversion solving. An infrared thermal imager collects device surface temperature distribution data, after multi-scale wavelet noise reduction and boundary heat flow balance optimization, a nonlinear heat conduction model is established by using temperature-related thermal conductivity coefficients, the heat diffusion boundary is optimized in combination with a high-order Laplace operator, and the defect heat source distribution is inverted through a variational regularization method, so that the geometric characteristics, thermal parameters and position coordinates of the defects are accurately extracted. The application can adapt to complex working conditions, realize high-precision detection and characteristic analysis of device defects, overcome the problems of noise interference, insufficient boundary processing and weak defect inversion capability in the prior art, and is widely applicable to fault diagnosis and health monitoring of industrial equipment.

The present application relates to the technical field of computer vision, and especially relates to a multi-modal semantic segmentation method of edge enhancement and space-frequency domain feature fusion, which comprises two modal multi-scale edge enhancement and feature fusion, feature extraction, enhancement, inhibition and interaction of different frequency bands of two modal tensors, and fusion and interaction of space and frequency domain features. The present application extracts edges through Laplace operator, dynamically adjusts the contribution of different modalities and scales through multi-scale edge enhancement and different scale adaptive weight fusion and different modal edge feature fusion, and realizes edge feature emphasis. In view of the problem that the segmentation accuracy is limited due to the ignored frequency domain features, the present application adopts a two-step fusion method of space and frequency domains, adds frequency domain for fusion in the calculation of the space attention weight matrix, and finally fuses the space and frequency attention matrices, so as to realize the full fusion of space and frequency domain features and the cross-complementation of features between different modalities.

The application belongs to the field of piezoelectric driver nonlinear modeling, and particularly discloses a piezoelectric driver creep nonlinear dynamics model and a construction method and system thereof. c (s)=s ‑μ ; wherein s is a Laplace operator; the identification method of the parameter mu includes: constructing input voltage signals with different frequencies omega i , calculating input voltage signal amplitudes U i ; according to the input voltage signals, obtaining output displacement signals of the piezoelectric driver, and simultaneously calculating output displacement signal amplitudes D i ; calculating the amplitude ratio A i of D i and U i ; obtaining a discrete data point set (x i , y i ) which satisfies: according to the linear fitting of the discrete data point set (x i , y i ), determining the slope of the fitted straight line, and determining the parameter mu according to the slope. The model constructed based on the fractional order involves fewer unknown parameters, is easy to identify, and at the same time, the modeling method is not limited by special discrete points in the time / frequency domain, and has higher universality.

The invention discloses a step frequency ground penetrating radar data feature-oriented adaptive edge detection method, which comprises the following steps of: preprocessing original frequency domain echo data of a step frequency ground penetrating radar, executing inverse Fourier transform on the preprocessed frequency domain data, converting the preprocessed frequency domain data into time domain signals, and combining the time domain signals to generate a two-dimensional radar profile map; time synchronization image calculation processing is carried out to obtain a one-dimensional energy curve, target scattering features are analyzed, and broadband data feature extraction is carried out; based on energy distribution and scattering characteristics, adaptively optimizing scale parameters of a Gaussian Laplacian operator; and applying the scale parameter to edge detection, and outputting a detection result. According to the method, normal form transformation from dependence on specific shape priori to data feature self-adaption is realized, various complex scenes can be adapted without shape priori, and the universality and the anti-interference capability of edge detection of different underground targets are remarkably improved.

The invention discloses a land water reserve inversion method and device, a medium and equipment, and relates to the technical field of hydrogeodetics. The method determines the average distance of the observation stations according to the ratio of the area of the target area to the number of the global navigation satellite system observation stations, and determines the average distance of the observation stations for each grid point according to the spatial distance from each global navigation satellite system observation station to the grid point and the ratio of the average distance of the observation stations to the average distance of the observation stations in the target range around the grid. The method comprises the following steps of: determining a weighting coefficient at a grid point to weight a Laplacian operator, combining the weighted Laplacian operator and a GNSS observation equation to construct a weighted Tikhonov regularization inversion model, and inverting the land water reserve change by minimizing a target function, thereby effectively improving the precision, reliability and space balance of an inversion result, and greatly improving the accuracy, reliability and space balance of the inversion result. Especially, effective constraints can still be obtained in areas with sparse or non-uniform observation stations, the land water reserve inversion deviation is reduced, and the inversion precision is improved.

The invention relates to the technical field of display panels, and discloses an automatic focusing method, device and equipment based on image processing and a storage medium, and the method comprises the steps: setting light intensity, exposure time, an initial direction, an initial step length, an initialization highest score and a corresponding position according to a defect type and an objective lens parameter; obtaining a defect position image, adaptively extracting an ROI region, filtering and denoising, and calculating a Laplacian operator mean value, an image variance and a gradient magnitude mean value of the ROI region; distributing feature weights based on defect types, and performing weighted summation to obtain a current focusing score; according to the comparison between the current focusing score and the historical focusing score, the moving direction and the step length are adjusted, and the highest score and the corresponding position are updated; when the step length is less than or equal to 0.25 [mu] m, if the current position is the highest score position or the distance between the current position and the highest score position is less than or equal to 0.25 [mu] m, stopping focusing and locking the optimal focusing position; according to the invention, the focusing efficiency, accuracy and consistency are improved.

A quantum computer-implemented system, method, and computer program product for quantum topological domain analysis (QTDA). The QTDA method achieves an improved exponential speedup and depth complexity of O(n log(1 / (δ∈))) where n is the number of data points, ∈ is the error tolerance, δ is the smallest nonzero eigenvalue of the restricted Laplacian, and achieves quantum advantage on general classical data. The QTDA system and method efficiently realizes a combinatorial Laplacian as a sum of Pauli operators; performs a quantum rejection sampling and projection approach to build the relevant simplicial complex repeatedly and restrict the superposition to the simplices of a desired order in the complex; and estimates Betti numbers using a stochastic trace / rank estimation method that does not require Quantum Phase Estimation. The quantum circuit and QTDA method exhibits computational time and depth complexities for Betti number estimation up to an error tolerance ∈.

The application belongs to the technical field of building boundary segmentation, and provides a remote sensing image building boundary segmentation method based on a U-net network model, which comprises the following steps: a Sobel operator is used to extract a gradient amplitude graph to generate a binary boundary mask, and a convolution kernel weight is dynamically adjusted in combination with a gradient direction to enhance the boundary region feature extraction capability; a high-frequency residual boundary mask is used to suppress boundary artifacts output by a hollow convolution, and a gradient direction is used to constrain a repair artifact area to retain original detail features; a complexity score is calculated based on a boundary density and a tortuosity to adaptively select a hollow rate and balance global context and local detail capture capability; a Laplace operator is used to extract a high-frequency component, a dynamic threshold is used to mark an artifact point, and a neighborhood gradient direction is used to repair an abnormal area; the performance bottleneck of a traditional method under the scenes of a blurred boundary, artifact interference and a fixed receptive field is solved, and the method is suitable for remote sensing image analysis tasks such as city planning and disaster assessment.

The present application relates to the technical field of deep gas risk exploration, and particularly relates to a deep low exploration area volcanic reservoir prediction method, device, equipment and medium. The method comprises the following steps: according to the reservoir sensitive parameters of a research area, a post-stack virtual well constrained inversion is selected to obtain a low frequency model of a full low frequency band of a virtual well; a geological model is established based on the low frequency model of the full low frequency band and seismic interface information; based on the reflection characteristics of the volcanic rock in the research area, a Laplace operator frequency division configuration attribute is used to depict the volcanic rock body; according to the geological model and the volcanic rock depiction result, a configuration attribute of a slice position of a target layer is extracted, low frequency modeling is carried out based on the configuration attribute constraint to obtain a configuration attribute volume controlled low frequency model; and inversion is carried out based on the configuration attribute volume controlled low frequency model to obtain a volcanic reservoir prediction result of the research area. The present application realizes the spatial distribution depiction of irregular geological bodies, and the depiction of special geological bodies is clearer, which has an important supporting role for risk well deployment.

The invention discloses a mask-based hierarchical Laplacian matting method, and the method employs an end-to-end deep learning network, achieves the generation from a rough mask to a fine gray-scale map through a strategy of combining multi-scale feature extraction and multi-scale Laplacian supervision, and achieves the matting of the rough mask and the fine gray-scale map. The method comprises the main technical steps of multi-scale feature coding, construction of Laplacian pyramid supervision signals and multi-scale independent prediction and fusion. According to the method, Laplacian pyramid supervision is adopted, and a complex prediction task is decomposed into low-frequency rough main body contour prediction and high-frequency fine edge residual prediction, so that a model learning target is clearer. A real image is decomposed into a low-frequency orthogonal component and a high-frequency orthogonal component through a Laplacian operator, and a high-resolution decoding layer is specially supervised to predict a high-frequency residual error extracted by the Laplacian operator, so that the network can more directly and effectively pay attention to and recover an extremely fine edge and a semitransparent region, which is difficult to solve by a traditional method.

The application relates to the technical field of online monitoring of power supply and distribution circuit devices, and discloses a product quality detection method based on deep learning and machine vision, which comprises the following steps: acquiring discrete space distribution signals of the surface state of a controlled power supply and distribution physical entity, adopting multiple scale factors to construct a multi-scale structure tensor group based on gradient characteristics of sampling points to be analyzed, determining energy evolution rates of the multi-scale structure tensor group at different scales, determining regional attributes of the sampling points, executing local directional energy suppression mapping to peel off interference components, superimposing a virtual disturbance component matrix on the multi-scale structure tensor group, determining topological singular points according to a principal axis direction deflection vector, and filling a virtual gradient into an amplitude saturation connected domain by using a topological extrapolation algorithm constrained by a Laplace operator, so that blind source decoupling of an environmental interference field and a material defect field can be realized, a signal blind area caused by strong reflection can be compensated, and the detection precision of a power supply conducting component is improved.