Mesh image edge feature extraction method and device
Through the Mesh image edge feature extraction method, Gaussian filtering and neural network screening feature nodes are used to solve the problem of noise interference and computing complexity in object detection, and more accurate edge positioning and real-time detection are achieved, suitable for resource-constrained environments.
Patent Information
- Application Number
- CN202411500393.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2024-12-27
AI Technical Summary
Existing object detection algorithms have challenges in noise interference, inaccurate edge positioning, computational complexity and real-time performance, especially in low contrast areas or high noise conditions, which make it difficult to accurately detect image edges, and computational complexity is difficult to meet real-time processing. need.
Using the Mesh image edge feature extraction method, the two-stage Gaussian filtering, Mesh node networking and screening features can reduce noise interference and calculation complexity. The N-layer neural network is used to predict similarity and gradient thresholds, and the feature nodes are filtered out to generate feature images. , reduce edge positioning errors.
It significantly reduces computing complexity, improves edge positioning accuracy, and enables real-time object detection in a shorter time, suitable for resource-constrained environments.
Smart Images
Figure CN119206250A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image edge feature extraction, and in particular to a method and device for extracting edge features of a Mesh image. Background Art
[0002] As one of the core tasks in the field of computer vision, object detection aims to identify and locate the position and category of specific objects in images or videos. This technology plays a vital role in many fields such as autonomous driving, video surveillance, and medical image analysis. However, object detection faces some challenges, mainly including target size changes, occlusion, lighting and background interference, category diversity, real-time requirements, lack and imbalance of datasets, and changes in perspective and posture.
[0003] Object detection algorithms can be divided into two-stage methods based on candidate regions, such as the R-CNN series of algorithms, and single-stage methods based on regression, such as YOLO and SSD. Each method has its own characteristics. For example, the YOLO series of algorithms are known for their high efficiency, while Faster R-CNN performs better in terms of accuracy. At present, object edge detection has the following problems:
[0004] (1) Noise interference: Noise in the image may affect the accuracy of edge detection, especially in low-contrast areas or high-noise conditions, where edges may be masked by noise or misdetected;
[0005] (2) Inaccurate edge positioning: Edge detection requires accurate positioning of the edge, but due to factors such as noise and lighting conditions, accurate edge positioning may become a challenge;
[0006] (3) Diversity of edge types: The edges in an image may be of different types, such as step edges, roof edges, etc. Different types of edges may require different processing methods;
[0007] (4) Computational complexity: With the development of image processing technology, although deep learning models such as CNN provide powerful feature description mechanisms, these models may be computationally intensive and difficult to apply in real-time or resource-constrained environments;
[0008] (5) The need for real-time edge detection: With the development of technology, real-time edge detection has become increasingly important, which requires the algorithm to be not only accurate but also efficient to meet the needs of real-time processing. Summary of the invention
[0009] In order to solve the problems of computational complexity, inaccurate edge positioning and noise interference in image edge detection, the present invention proposes a Mesh image edge feature extraction method and device. The Mesh image edge feature extraction algorithm adopts two-stage Gaussian filtering, Mesh node networking and feature screening, which greatly reduces noise and computational complexity and highlights edge features more accurately, providing algorithm support for target detection such as YOLOX series and SSD, such as detection of motor vehicles, pedestrians and objects.
[0010] The present invention discloses a method for extracting edge features of a Mesh image, which comprises:
[0011] Preprocess the original Mesh image to obtain the preprocessed image; arrange the position of each pixel in the preprocessed image, then extract the RGB value of each pixel, and calculate the similarity and gradient between each pixel and all the pixels in its preset neighborhood; each pixel is a node, and each pixel is locally meshed to obtain a Mesh network; based on the initial threshold, preliminarily screen out feature nodes from all pixels in each Mesh network;
[0012] The similarity threshold and gradient threshold corresponding to each node are predicted through an N-layer neural network. The required feature nodes are screened out from the initially screened feature nodes according to the similarity threshold and gradient threshold corresponding to each node. The feature image is drawn according to the position and RGB value of the required feature node, and finally the boundary is removed to generate the feature image.
[0013] Furthermore, the preprocessing of the original Mesh image to obtain a preprocessed image includes:
[0014] Fill the borders of the original Mesh image and mark each pixel position and RGB value;
[0015] The image is subjected to one-dimensional Gaussian blurring in the y-axis direction using a separated Gaussian filter, and one-dimensional Gaussian blurring in the x-axis direction is performed on the one-dimensional Gaussian blurring result to obtain a preprocessed image.
[0016] Furthermore, the preprocessing of the original Mesh image to obtain a preprocessed image includes:
[0017] Fill the borders of the original Mesh image and mark each pixel position and RGB value;
[0018] The image is subjected to one-dimensional Gaussian blurring in the x-axis direction using a separated Gaussian filter, and one-dimensional Gaussian blurring in the y-axis direction is performed on the one-dimensional Gaussian blurring result to obtain a preprocessed image.
[0019] Furthermore, each pixel point Γ=(P, RGB, sim, G, LNum) in the preprocessed image includes the pixel position P, RGB value, similarity sim and gradient G of all pixels in a preset neighborhood, and effective number of threshold filtering LNum, where LNum represents the cumulative number of times after passing the threshold filtering.
[0020] Furthermore, the local mesh networking of each pixel point includes:
[0021] With each pixel as the center point, Mesh networking is performed within its preset neighborhood. A pixel is selected from the Mesh networking as the head node of the linked list, and then another pixel is selected from the Mesh networking as the tail node. Communication is performed between all nodes from the head node to the tail node.
[0022] Furthermore, the initial threshold is used to preliminarily select feature nodes from all pixels, including:
[0023] Compare the similarity and gradient corresponding to each node in the Mesh network with the initial similarity threshold and the initial gradient threshold, filter out nodes whose similarity is greater than the initial similarity threshold and whose gradient is less than the initial gradient threshold, and update the threshold filtering effective times corresponding to the remaining nodes in the Mesh network respectively.
[0024] Furthermore, predicting the similarity threshold and gradient threshold corresponding to each node through the N-layer neural network includes:
[0025] Assume that there is a trained image model M, and the features F in the image model M are known to be F = (f1, f2, f3, ..., f n ) and the eigenvector f n Corresponding similarity threshold and gradient threshold, the N-layer neural network includes neural network layers with weights and activation functions;
[0026] Set the initial similarity and gradient values;
[0027] Through N layers of neural network with initial weights, the similarity threshold and gradient threshold corresponding to each node are predicted; the loss error is obtained by evaluation results; the weights of each layer are updated by back propagation from top to bottom, and the predicted similarity threshold and gradient threshold corresponding to each node in the Mesh network are saved; the similarity threshold and gradient threshold corresponding to each node are continued to be predicted, and the loss error is obtained by evaluation results; the weights of each layer are updated by back propagation from top to bottom, and the predicted similarity threshold and gradient threshold corresponding to each node in the Mesh network are saved, until the error of predicted correlation value and marked correlation value converges.
[0028] Furthermore, according to the effective number of threshold filtering, after updating a round of weights of each layer, the forward propagation calculation is stopped. The larger the value of the effective number of threshold filtering, the more obvious the pixel features are.
[0029] Furthermore, the method of selecting required feature nodes from the initially selected feature nodes according to the similarity threshold and the gradient threshold corresponding to each node includes:
[0030] Nodes whose gradients are greater than a gradient threshold and whose similarities are less than a similarity threshold are selected from the initially selected feature nodes as required feature nodes.
[0031] The present invention also discloses a Mesh image edge feature extraction device, which implements any of the above-mentioned Mesh image edge feature extraction methods, and comprises:
[0032] The image preprocessing module is used to preprocess the original Mesh image to obtain a preprocessed image;
[0033] The image feature extraction module is used to arrange the position of each pixel in the preprocessed image, then extract the RGB value of each pixel, and calculate the similarity and gradient between each pixel and all the pixels in its preset neighborhood; each pixel is a node, and each pixel is locally meshed to obtain a Mesh network; based on the initial threshold, the feature nodes are preliminarily screened out from all the pixels in each Mesh network; the similarity threshold and gradient threshold corresponding to each node are predicted through an N-layer neural network, and the required feature nodes are screened out from the preliminarily screened feature nodes according to the similarity threshold and gradient threshold corresponding to each node; the feature image is drawn according to the position and RGB value of the required feature node, and finally the boundary is removed to generate the feature image.
[0034] Due to the adoption of the above technical solution, the present invention has the following advantages: the present invention greatly reduces the computational complexity and can perform edge positioning more accurately, so as to facilitate the realization of real-time target detection in a shorter time. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.
[0036] Figure 1 A schematic diagram of the main process of the Mesh network image edge feature extraction algorithm according to an embodiment of the present invention;
[0037] Figure 2A schematic diagram of adding a layer of boundary pixels at the edge of an image according to an embodiment of the present invention;
[0038] Figure 3 A schematic diagram of an image after Gaussian filtering according to an embodiment of the present invention;
[0039] Figure 4 A schematic diagram of the "networking" communication of Mesh nodes including data transmission in an embodiment of the present invention;
[0040] Figure 5 A schematic diagram of filtering valid pixels by updating the threshold according to an embodiment of the present invention;
[0041] Figure 6 A schematic diagram of screening effective feature nodes by predicting a similarity gradient threshold according to an embodiment of the present invention;
[0042] Figure 7 This is a schematic diagram of the final image features according to an embodiment of the present invention. DETAILED DESCRIPTION
[0043] The present invention is further described in conjunction with the accompanying drawings and embodiments, and the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present invention.
[0044] See also Figure 1 The present invention provides an embodiment of a method for extracting edge features of a Mesh image, which comprises:
[0045] The original Mesh (magic palace grid) image is preprocessed to obtain a preprocessed image; each pixel in the preprocessed image is arranged in position, and then the RGB value of each pixel is extracted, and the similarity and gradient between each pixel and all the pixels in its preset neighborhood are calculated; each pixel is a node, and each pixel is locally meshed to obtain a Mesh network; based on the initial threshold, the feature nodes are preliminarily screened from all the pixels in each Mesh network;
[0046] The similarity threshold and gradient threshold corresponding to each node are predicted through an N-layer neural network. The required feature nodes are screened out from the initially screened feature nodes according to the similarity threshold and gradient threshold corresponding to each node. The feature image is drawn according to the position and RGB value of the required feature node, and finally the boundary is removed to generate the feature image.
[0047] In one embodiment of the present application, preprocessing the original Mesh image to obtain a preprocessed image includes:
[0048] See also Figure 2, fill the border of the original Mesh image and mark each pixel position and RGB value;
[0049] The image is subjected to one-dimensional Gaussian blurring in the y-axis direction using a separated Gaussian filter, and one-dimensional Gaussian blurring in the x-axis direction is performed on the one-dimensional Gaussian blurring result to obtain a preprocessed image.
[0050] In one embodiment of the present application, preprocessing the original Mesh image to obtain a preprocessed image includes:
[0051] Fill the borders of the original Mesh image and mark each pixel position and RGB value;
[0052] The image is subjected to one-dimensional Gaussian blurring in the x-axis direction using a separated Gaussian filter, and one-dimensional Gaussian blurring in the y-axis direction is performed on the one-dimensional Gaussian blurring result to obtain a preprocessed image.
[0053] Specifically,
[0054] f(x,y)=g(x)*h(y)
[0055] Among them, g(x) and h(y) are one-dimensional Gaussian functions. Since the product of g(x) and h(y) still satisfies the Gaussian distribution, it means that each direction can be filtered separately, and finally the image after noise removal can be obtained. The advantage is that the computational complexity of the image can be reduced. The following are the specific steps:
[0056] Generate 1D Gaussian kernel: First, generate two 1D Gaussian kernels according to the required standard deviation σ and kernel size. These kernels will be applied in the horizontal and vertical directions respectively;
[0057] First filtering: Apply the first one-dimensional Gaussian kernel to each channel of the image to filter horizontally. This can be done by using a convolution operation;
[0058] Second filtering: Based on the first step, a second one-dimensional Gaussian kernel is applied to each channel of the image, but this time along the vertical direction;
[0059] Merge result: After two filterings, the image with noise removed is obtained, as shown in Figure 3 shown.
[0060] Then, by adjusting the standard deviation σ of the Gaussian filter and the size of the convolution kernel, using a smaller convolution kernel can reduce the amount of calculation.
[0061]
[0062] Among them, G(x,y) is a Gaussian function.
[0063] In one embodiment of the present application, each pixel point Γ=(P, RGB, sim, G, LNum) in the preprocessed image includes the pixel position P, RGB value, the similarity sim and gradient G of all pixels within a preset neighborhood, and the effective number of threshold filtering LNum, where LNum represents the cumulative number of times after passing the threshold filtering.
[0064] Optionally, each 1*1 pixel point is used as a node in the mesh network. Each node in the mesh network can "communicate" with each other to transfer data using a linked list data structure. Compared with the CNN convolutional neural network, although it increases a certain communication overhead, it reduces a large number of complex calculations while reducing time and space complexity, and can lock the feature boundaries more accurately.
[0065] Taking each node adjacent to 8 pixels as an example, calculate the similarity and gradient of each node:
[0066] G x (x,y)=H(x+1,y)-H(x-1,y)
[0067] G y (x,y)=H(x,y+1)-H(x,y-1)
[0068]
[0069] Among them, H(x+1,y) and H(x-1,y) are one-dimensional Gaussian functions, which are smoothed on the image in the x+1 and x-1 directions respectively to obtain a one-stage Gaussian function G x (x,y); similarly for G y (x,y); G(x,y) is the gradient; α(x,y) is G y (x,y) vs. G x (x, y); d(x, y) is the distance between x and y in space; sim(x, y) is the similarity. The following is an explanation: the range is [0, 1]. The larger the value, the smaller d is, that is, the closer the distance is, the greater the similarity is.
[0070] In one embodiment of the present application, the local mesh networking of each pixel point includes:
[0071] See also Figure 4 , taking each pixel point as the center point, Mesh networking is performed within its preset neighborhood range, a pixel point is selected from the Mesh networking as the head node of the linked list, and then another pixel point is selected from the Mesh networking as the tail node, and communication is performed between all nodes from the head node to the tail node.
[0072] After calculating the similarity and gradient, Figure 4Each node in the green matrix part can be regarded as the head of the linked list, and the end of the communication is the tail of the linked list. Since the use of linked lists can greatly reduce the computational complexity while reducing the time complexity and space complexity, it increases the communication overhead between nodes to a certain extent. Each mesh node can communicate with each other directly to transmit data. The similarity and gradient of each node are compared with the initial set thresholds TH and G to filter out the values with small gradients and small similarities, and the filtered nodes update their LNUM values.
[0073] In one embodiment of the present application, the preliminary screening of feature nodes from all pixel points based on the initial threshold includes:
[0074] See also Figure 5 , compare the similarity and gradient corresponding to each node in the Mesh network with the initial similarity threshold and the initial gradient threshold, screen out the nodes whose similarity is less than the initial similarity threshold and whose gradient is greater than the initial gradient threshold, and add 1 to the LNUM of the corresponding adjacent screened mesh nodes; the following is an explanation: the adjacent mesh nodes are 8 mesh nodes: upper left, upper, upper right, left, right, lower left, lower, and lower right.
[0075] Figure 5 In the figure, assuming that the initial similarity and gradient threshold of the mark are Lsim and LG, the similarity of the mesh node relative to the surrounding 8 mesh nodes (i.e., 8 pixels) is calculated: [sim(leftTop),sim(top),sim(rightTop),sim(left),sim(right),sim(leftDown),sim(down),sim(rightDown)] and gradient [G(leftTop),G(top),G(rightTop),G(left),G(right),G(leftDown),G(down),G(rightDown)], and then the similarity and gradient thresholds are predicted each time as PreSim and PreG respectively; compare PreSim, PreG and the similarity and gradient of the above mesh node relative to the surrounding 8 mesh nodes (the first time is to use Lsim and LG with the similarity and gradient of the 8 mesh nodes), filter out the nodes whose similarity is less than the similarity threshold and whose gradient is greater than the gradient threshold, and the LNUM of the filtered mesh nodes is accumulated by 1.
[0076] In one embodiment of the present application, predicting the similarity threshold and gradient threshold corresponding to each node through an N-layer neural network includes:
[0077] Assume that there is a trained image model M, and the features F in the image model M are known to be F = (f1, f2, f3, ..., f n) and the eigenvector f n The corresponding similarity threshold and gradient threshold, the N-layer neural network contains a neural network layer with weights and an activation function; f1, f n are the first eigenvector and the nth eigenvector respectively, then n is the number of eigenvectors and also the length of the eigenvector.
[0078] Set the initial similarity and gradient values;
[0079] Through N layers of neural network with initial weights, the similarity threshold and gradient threshold corresponding to each node are predicted; the loss error is obtained by evaluation; the weights of each layer are updated by backpropagation from top to bottom, and the predicted similarity threshold and gradient threshold corresponding to each node in the Mesh network are saved; the similarity threshold and gradient threshold corresponding to each node are continued to be predicted, and the loss error is obtained by evaluation; the weights of each layer are updated by backpropagation from top to bottom, and the predicted similarity threshold and gradient threshold corresponding to each node in the Mesh network are saved until the error of the predicted correlation value and the marked correlation value converges; the predicted correlation value includes the predicted gradient and similarity; the marked correlation value includes the marked gradient and similarity.
[0080] In one embodiment of the present application, according to the effective number of threshold filtering, after a round of updating of weights of each layer is completed, the forward propagation calculation is stopped. The larger the value of the effective number of threshold filtering, the more obvious the pixel features are.
[0081] In one embodiment of the present application, the step of selecting required feature nodes from the initially selected feature nodes according to the similarity threshold and the gradient threshold corresponding to each node includes:
[0082] See also Figure 6 , select nodes with gradients greater than the gradient threshold and similarities less than the similarity threshold from the initially selected feature nodes as the required feature nodes. Draw the feature image according to the position and RGB value of the required feature node, and finally remove the boundary to generate the feature image, such as Figure 7 shown.
[0083] The present invention also provides an embodiment of a Mesh image edge feature extraction device, which is used to implement the Mesh image edge feature extraction method described in the above embodiment, and comprises:
[0084] The image preprocessing module is used to preprocess the original Mesh image to obtain a preprocessed image;
[0085] The image feature extraction module is used to arrange the position of each pixel in the preprocessed image, then extract the RGB value of each pixel, and calculate the similarity and gradient between each pixel and all the pixels in its preset neighborhood; each pixel is a node, and each pixel is locally meshed to obtain a Mesh network; based on the initial threshold, the feature nodes are preliminarily screened out from all the pixels in each Mesh network; the similarity threshold and gradient threshold corresponding to each node are predicted through an N-layer neural network, and the required feature nodes are screened out from the preliminarily screened feature nodes according to the similarity threshold and gradient threshold corresponding to each node; the feature image is drawn according to the position and RGB value of the required feature node, and finally the boundary is removed to generate the feature image.
[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for extracting edge features of a Mesh image, characterized in that: include: Preprocess the original Mesh image to obtain a preprocessed image; Arrange the position of each pixel in the preprocessed image, extract the RGB value of each pixel, and calculate the similarity and gradient between each pixel and all pixels in its preset neighborhood. Perform local Mesh networking on each pixel to obtain a Mesh network; Based on the initial threshold, feature nodes are preliminarily screened from all pixels in each Mesh network; The similarity threshold and gradient threshold corresponding to each node are predicted through an N-layer neural network. The required feature nodes are screened out from the initially screened feature nodes according to the similarity threshold and gradient threshold corresponding to each node. The feature image is drawn according to the position and RGB value of the required feature node, and finally the boundary is removed to generate the feature image.
2. The Mesh image edge feature extraction method according to claim 1, characterized in that: The preprocessing of the original Mesh image to obtain the preprocessed image includes: Fill the borders of the original Mesh image and mark each pixel position and RGB value; The image is subjected to one-dimensional Gaussian blurring in the y-axis direction using a separated Gaussian filter, and one-dimensional Gaussian blurring in the x-axis direction is performed on the one-dimensional Gaussian blurring result to obtain a preprocessed image.
3. The Mesh image edge feature extraction method according to claim 1, characterized in that: The preprocessing of the original Mesh image to obtain the preprocessed image includes: Fill the borders of the original Mesh image and mark each pixel position and RGB value; The image is subjected to one-dimensional Gaussian blurring in the x-axis direction using a separated Gaussian filter, and one-dimensional Gaussian blurring in the y-axis direction is performed on the one-dimensional Gaussian blurring result to obtain a preprocessed image.
4. The method for extracting edge features of a Mesh image according to any one of claims 1 to 3, characterized in that: Each pixel point Γ = (P, RGB, sim, G, LNum) in the preprocessed image includes the pixel position P, RGB value, the similarity sim and gradient G of all pixels in the preset neighborhood, and the effective number of threshold filtering LNum, where LNum represents the cumulative number of times after passing the threshold filtering.
5. The Mesh image edge feature extraction method according to claim 1, characterized in that: The local mesh networking of each pixel point includes: With each pixel as the center point, Mesh networking is performed within its preset neighborhood. A pixel is selected from the Mesh networking as the head node of the linked list, and then another pixel is selected from the Mesh networking as the tail node. Communication is performed between all nodes from the head node to the tail node.
6. The method for extracting edge features of a Mesh image according to claim 1, characterized in that: According to the initial threshold, feature nodes are initially screened out from all pixels, including: Compare the similarity and gradient corresponding to each node in the Mesh network with the initial similarity threshold and the initial gradient threshold, filter out nodes whose similarity is greater than the initial similarity threshold and whose gradient is less than the initial gradient threshold, and update the threshold filtering effective times corresponding to the remaining nodes in the Mesh network respectively.
7. The method for extracting edge features of a Mesh image according to claim 1, characterized in that: The method of predicting the similarity threshold and gradient threshold corresponding to each node through an N-layer neural network includes: Assume that there is a trained image model M, and the features F in the image model M are known to be F = (f1, f2, f3, ..., f n ) and the eigenvector f n Corresponding similarity threshold and gradient threshold, the N-layer neural network includes neural network layers with weights and activation functions; Set the initial similarity and gradient values; Through N layers of neural network with initial weights, the similarity threshold and gradient threshold corresponding to each node are predicted; the loss error is obtained by evaluation results; the weights of each layer are updated by back propagation from top to bottom, and the predicted similarity threshold and gradient threshold corresponding to each node in the Mesh network are saved; the similarity threshold and gradient threshold corresponding to each node are continued to be predicted, and the loss error is obtained by evaluation results; the weights of each layer are updated by back propagation from top to bottom, and the predicted similarity threshold and gradient threshold corresponding to each node in the Mesh network are saved, until the error of predicted correlation value and marked correlation value converges.
8. The method for extracting edge features of a Mesh image according to claim 7, characterized in that: According to the effective number of threshold filtering, after updating a round of weights of each layer, the forward propagation calculation is stopped. The larger the effective number of threshold filtering, the more obvious the pixel features.
9. The method for extracting edge features of a Mesh image according to claim 1, characterized in that: The method of selecting required feature nodes from the initially selected feature nodes according to the similarity threshold and the gradient threshold corresponding to each node includes: Nodes whose gradients are greater than a gradient threshold and whose similarities are less than a similarity threshold are selected from the initially selected feature nodes as required feature nodes.
10. A device for extracting edge features of a Mesh image, implementing the method for extracting edge features of a Mesh image according to any one of claims 1 to 9, characterized in that: include: The image preprocessing module is used to preprocess the original Mesh image to obtain a preprocessed image; The image feature extraction module is used to arrange the position of each pixel in the preprocessed image, then extract the RGB value of each pixel, and calculate the similarity and gradient between each pixel and all the pixels in its preset neighborhood; each pixel is a node, and each pixel is locally meshed to obtain a Mesh network; based on the initial threshold, the feature nodes are preliminarily screened out from all the pixels in each Mesh network; the similarity threshold and gradient threshold corresponding to each node are predicted through an N-layer neural network, and the required feature nodes are screened out from the preliminarily screened feature nodes according to the similarity threshold and gradient threshold corresponding to each node; the feature image is drawn according to the position and RGB value of the required feature node, and finally the boundary is removed to generate the feature image.
Citation Information
Cited By
Panoramic video stream processing method and system for guaranteeing distribution continuity
CN122205086A