High-precision image measurement system based on multi-feature fusion and measurement method thereof

Abstract

The invention discloses a high-precision image measurement system and method based on multi-feature fusion, and the method comprises the steps: collecting the image data of a to-be-measured object through a plurality of sensors, extracting the color, texture, shape and depth features, and carrying out the processing through an independent algorithm. According to the environment and requirements, an intelligent algorithm selects related features and distributes weights; fusing selected features by using a deep learning model, performing time synchronization and space registration of data, and eliminating space-time errors between sensors; noise suppression, illumination adaptation and background processing are carried out on the image data, finally fusion data are optimized, and a high-precision measurement result is output; according to the high-precision image measurement system based on multi-feature fusion and the measurement method thereof, through intelligent fusion and dynamic optimization of multiple features, the measurement precision and robustness of the system in a complex environment are improved.

Description

Technical Field

[0001] The present invention belongs to the field of high-precision image measurement based on multi-feature fusion, and particularly relates to a high-precision image measurement system and a measurement method based on multi-feature fusion. Background Art

[0002] A high-precision image measurement system based on multi-feature fusion is a system that improves measurement accuracy and robustness by fusing multiple image features or information. This system is often used in fields requiring high precision, such as 3D reconstruction, precision manufacturing, quality inspection, and topographic surveying. It uses color information in images to identify different objects or surfaces; uses texture patterns in images to identify or measure objects; uses geometric information such as edges, corners, and contours to assist in object positioning and measurement; combines depth images or stereo vision to obtain 3D information; and uses changes in illumination to infer surface details or the spatial position of objects. Feature fusion is typically achieved using image processing algorithms, deep learning, and sensor fusion technologies. By fusing multiple features, the system can maintain high-precision measurement performance in various complex environments and reduce the errors and uncertainties caused by single features. Therefore, image measurement systems based on multi-feature fusion are generally more robust in practical applications and can adapt to complex scenarios such as illumination changes, occlusion, and noise interference.

[0003] However, although the existing high-precision image measurement systems based on multi-feature fusion have significant advantages in practical applications, they also have some defects and challenges. Although the existing image measurement systems can achieve relatively accurate measurements in some high-precision applications, they are often affected by problems such as ambient lighting changes, noise interference, complex scene occlusions, and sensor inconsistency, resulting in insufficient measurement accuracy and system robustness. Summary of the Invention

[0004] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a high-precision image measurement system and a measurement method based on multi-feature fusion, which improves the measurement accuracy and robustness of the system in complex environments through intelligent fusion and dynamic optimization of multiple features.

[0005] The technical solution adopted by the present invention to solve its technical problem is:

[0006] High-precision image measurement system based on multi-feature fusion, including:

[0007] Multi-feature extraction module, used to extract multiple feature information from the input image, including but not limited to color features, texture features, shape features, and depth features. Each feature is extracted using an independent algorithm;

[0008] Intelligent feature selection module, which uses adaptive algorithms to automatically select the most relevant features based on different scenarios and measurement requirements, and dynamically adjusts the importance weights of features;

[0009] Feature fusion module, which is used to intelligently fuse features from multiple sources using a deep learning model;

[0010] Multi-sensor data synchronization and registration module, used for accurate time synchronization and spatial registration of data from different types of sensors;

[0011] The noise suppression and environment adaptation module combines deep learning with traditional image processing methods to suppress noise in images in real time.

[0012] The measurement result optimization module is used to optimize the fused features and output high-precision measurement results. Through the optimization algorithm, the stability and accuracy of the measurement are further improved.

[0013] As a preference, the feature fusion module performs feature fusion by weighted synthesis of information from multiple channels; the multi-sensor data synchronization and registration module adopts a multimodal sensor data fusion algorithm; the noise suppression and environmental adaptation module is configured with the ability to automatically adapt to different lighting and environmental changes; the measurement result optimization module dynamically adjusts the optimization strategy according to the actual measurement situation.

[0014] The high-precision image measurement method based on multi-feature fusion includes the following steps:

[0015] Multiple sensors are used to collect image data of the object to be measured, including color images, depth images, and sensor data of lidar data;

[0016] Extract color, texture, shape and depth feature information from the collected image data, and each feature is processed and extracted by an independent algorithm;

[0017] According to the current environment and measurement requirements, an intelligent algorithm is used to select relevant features and assign appropriate weights to each feature;

[0018] By integrating the selected features through a deep learning model, the information of multiple features is integrated to synchronize the time and space of data from different sensors, eliminating the temporal and spatial errors between sensors;

[0019] The image data is subjected to noise suppression, adapted to illumination changes and complex background factors, and the fused feature data is finally optimized to output high-precision measurement results.

[0020] Preferably, a method for using multiple sensors to collect image data of an object to be measured, wherein the data includes color images, depth images, and sensor data of lidar data is as follows:

[0021] For the data collected by each sensor, denoising and alignment processing are performed, and after noise removal and color correction processing, I color After denoising and smoothing, the depth image D is obtained depth , by filtering the point cloud data and removing outliers, we can obtain the point cloud data P lidar ;

[0022] LiDAR point cloud data P lidar ={p1, p2, ..., p n The projection matrix is ​​aligned with the depth data of the color image to obtain the corresponding image pixel value. This process is achieved through the following projection formula:

[0023] p′ i =Proj(R·p i +t)

[0024] Among them, R is the rotation matrix, t is the translation vector, and p i is the i-th point in the lidar point cloud, p′ i is the projection position of the point in the image coordinate system;

[0025] The depth image D depth With the point cloud data P of the laser radar lidar Alignment is performed to obtain the corresponding relationship between the depth information and the lidar point cloud. The depth value d(x, y) is aligned with the point cloud data using the following formula:

[0026] P lidar =K·[x,y,d(x,y)] T

[0027] Among them, K is the camera intrinsic parameter matrix, (x, y) is the pixel coordinate in the image, d(x, y) is the depth value corresponding to the pixel, P lidar is the corresponding 3D coordinate in the point cloud data;

[0028] The data of different sensors are integrated through weighted average or deep learning model, and the data of each sensor is set up separately using I color 、D depth 、P lidar It indicates that its fusion process is carried out through the following steps;

[0029] From the color image I color , depth image D depth , LiDAR point cloud data P lidar Extract features; Different sensors provide different types of information, and feature fusion is performed in a weighted manner, setting the weight w of each feature color , wdepth , w lidar , the weighted formula for feature fusion is:

[0030] F fusion (x, y) = w color ·f color (x, y) + w depth ·f depth (x, y) + w lidar ·f lidar (x,y)

[0031] Among them, f color (x, y) is the feature extracted from the color image, f depth (x, y) is the feature extracted from the depth image, f lidar (x, y) is the feature extracted from the lidar point cloud;

[0032] Based on the fused data, the final accurate measurement is performed to measure the three-dimensional coordinates P of the object. object , is calculated based on the fused image data as follows:

[0033] P object =f(I color , D depth , P lidar )

[0034] Where f is a function of the three-dimensional object position or other geometric parameters calculated according to the multi-feature fusion model.

[0035] Preferably, the color, texture, shape and depth feature information is extracted from the collected image data, and each feature is processed and extracted by an independent algorithm as follows:

[0036] The color histogram describes the frequency of pixel distribution in each color channel of the image. The color distribution of the image is composed of three channels C r , C g , C b describe:

[0037] ParseError:KaTeX parse error:Can′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N} \( r-r_i)

[0038] ParseError:KaTeX parse error:Can′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N}\( g-g_i)

[0039] ParseError: KaTeX parse error: Can′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N} \( b-b_i)

[0040] Among them, H r , H g , H b is the histogram of the red, green and blue channels, is the Dirac function, r i , g i , b i For the image The red, green and blue channel values ​​of each pixel, where N is the number of pixels in the image;

[0041] Texture features are extracted through the gray-level co-occurrence matrix or local binary pattern method. The gray-level co-occurrence matrix describes the spatial relationship between the gray values ​​of pixel pairs and is used for texture analysis. The gray value of the image is set as I(x, y). The GLCM matrix is ​​defined as:

[0042] ParseError:KaTeX parseerror:Can′t use function′\(′in math mode atposition 31:...=\sum_{x, y} \( I(x,y)=i,I(...

[0043] Where p(i, j, θ) represents the frequency of the pixel pair (i, j) in the specified direction θ, and _x, _y are the distances between the pixel pairs.

[0044] LBP is used to describe local texture structure. Let I(i, j) be the pixel value of the image at position (x, y). The LBP operator converts the neighborhood of a pixel into a binary sequence:

[0045]

[0046] Where s(x) is a sign function. When x≥0, s(x)=1, otherwise s(x)=0. I(x n ,y n ) is the neighborhood pixel value, and N is the size of the neighborhood;

[0047] Shape feature extraction is used to analyze the geometric shape of an image. The Hu moment is a set of invariant moments derived from the central moment of the image, which describes the geometric features of the shape. The central moment μ pq Defined as:

[0048]

[0049] Among them, (c x , c y ) is the centroid of the image, p and q are the orders of the moments, and the general form of the Hu moment is:

[0050] φ1=μ 20 +μ 02

[0051]

[0052] φ3=(μ 30 -3μ 12 ) 2 +(3μ 21 -μ 03 ) 2

[0053] By extracting the contour of the image, the methods include edge detection and contour simplification. Depth feature extraction is used to extract the three-dimensional geometric information of the object from the depth image. The algorithms include normal estimation and curvature analysis.

[0054] The normal of each point is calculated through the depth image to obtain the direction information of the object surface. For a point P(x, y) in the depth image, its normal is estimated by the neighborhood points. The depth of the image is set to D(x, y). The normal n is calculated as follows:

[0055]

[0056] in, is the gradient of the depth image at point (x, y);

[0057] Curvature is a feature that describes the degree of curvature of the surface, including Gaussian curvature and mean curvature. For each pixel in the depth image, the curvature is calculated by the second-order derivative;

[0058]

[0059] Among them, K g is the Gaussian curvature;

[0060] For the extraction of multiple features, multiple algorithms are performed in parallel to combine different features, and the color, texture, shape and depth features are spliced ​​into a high-dimensional feature vector for subsequent classification and detection tasks. The formula of the fusion process is:

[0061] F total =[F color , F texture , F shape , Fdepth ]

[0062] Among them, F color , F texture , F shape , F depth They are color, texture, shape and depth feature vectors respectively, and finally a comprehensive feature vector F is obtained total .

[0063] As a preferred method, according to the current environment and measurement requirements, an intelligent algorithm is used to select relevant features and assign appropriate weights to each feature:

[0064] The feature is selected based on its correlation with the target variable, and the Pearson correlation coefficient (ρ) is used to calculate the correlation between each feature and the target:

[0065]

[0066] Among them, Cov(X, Y) is the covariance of feature X and target variable Y, σ X and σ Y are the standard deviations of the feature and target variables, respectively;

[0067] Lasso regression selects features and automatically adjusts their weights through penalty terms. The objective function of Lasso regression is:

[0068]

[0069] Among them, y i is the target value, x i is the feature vector, w is the feature weight, and λ is the regularization parameter;

[0070] Weighting is to assign weights to each feature according to its importance. The weights are adaptively obtained through a learning algorithm using the following formula:

[0071] w=argmin w {L(w)+λ||w||2}

[0072] Where w is the feature weight vector, L(w) is the loss function, λ is the regularization coefficient, and ||w||2 is the L2 regularization term;

[0073] When using decision tree, random forest or gradient boosted tree models, extract feature importance directly from the model. Feature importance in random forest is j Calculate the impact of each feature on the decision tree:

[0074]

[0075] Among them, S is the feature subset, Si For the specific features, accuracy(S i ) is based on the subset S i Model accuracy;

[0076] PSO is used to find the optimal feature selection scheme and the corresponding weight distribution. The position of the particle represents the combination of feature selection, the speed of the particle represents the weight adjustment, and the optimization goal is to minimize the loss function, error or loss:

[0077]

[0078] Combining feature selection and weighting, we get the following optimization formula:

[0079]

[0080] Among them, F selected is the weighted feature set, F i is the i-th feature, w i is the corresponding weight, and M is the number of features.

[0081] As a preferred method, the selected features are fused through a deep learning model, the information of multiple features is integrated, the data of different sensors are synchronized in time and space, and the method of eliminating the spatiotemporal errors between sensors is as follows:

[0082] The purpose of feature fusion is to merge features from different sources. A neural network is used to perform feature fusion. The input of the model is the features of multiple sensor data, and the output is the fused features. Its basic form is expressed as:

[0083] z=Fusion(x1,x2,...,x n )

[0084] Among them, x1, x2, ..., x n are the data features from different sensors;

[0085] z is the fused feature representation;

[0086] Set up two sensors S1 and S2, whose data sampling times are t1 and t2 respectively. Align them through interpolation and set up a deep learning-based model to learn the time synchronization transformation:

[0087] t2=f(t1,θ)

[0088] Where t1 is the timestamp of sensor 1;

[0089] t2 is the timestamp of sensor 2;

[0090] f(t1, θ) is the time synchronization function learned, where θ is the model parameter;

[0091] To ensure that multiple sensor data are spatially aligned and eliminate errors caused by differences in sensor positions, two sensors S1 and S2 are set up, whose spatial coordinates are p1 and p2 respectively. The purpose of spatial registration is to transform p1 to p2 through a transformation function g(p1, θ):

[0092] p2=g(p1,θ)

[0093] Where p1 is the spatial position of sensor 1;

[0094] p2 is the spatial position of sensor 2;

[0095] g(p1,θ) is the spatial registration function;

[0096] The elimination of spatiotemporal errors is achieved by fusing the data after time synchronization and spatial registration. The deep learning model for eliminating errors is a loss function that optimizes the data after time synchronization and spatial registration to minimize the error. The model output is set as The target is y, and the loss function is expressed as:

[0097]

[0098] in, The data output by the deep learning model has undergone spatiotemporal synchronization and spatial registration;

[0099] y i is the target data;

[0100] N is the total number of data.

[0101] As a preferred method, the image data is subjected to noise suppression, adapted to illumination changes and complex background factors, and the fused feature data is subjected to final optimization processing to output high-precision measurement results:

[0102] Noise suppression and illumination change adaptation are performed through filtering and normalization methods. The input image is set as I noisy (x, y), where x and y are pixel coordinates. Noise suppression is achieved through image denoising algorithms, and illumination adaptation is achieved through brightness normalization and histogram equalization. The calculation formula is as follows:

[0103]

[0104]

[0105] in, represents the noise suppression operation;

[0106] Represents light adaptation operation;

[0107] Complex background factors are processed through background modeling or background subtraction, which is completed through deep learning models or classic background modeling methods. The calculation formula is:

[0108]

[0109] in, represents the background suppression algorithm;

[0110] In image processing, feature fusion is used to combine feature data from multiple different sources, and to establish multiple feature images {F1, F2, ..., F n}, then the feature fusion operation formula is expressed as:

[0111]

[0112] in, It is the feature fusion operation;

[0113] The fused feature data is optimized and processed through the deep learning model, and the fused feature data is set as F fusion , and finally optimize the processing to output high-precision measurement results R final , the calculation formula is:

[0114]

[0115] in, represents the final optimization operation, performed by the deep neural network, and θ is the network parameter.

[0116] Another technical problem to be solved by the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, a high-precision image measurement system and a measurement method based on multi-feature fusion as described above are implemented.

[0117] Another technical problem to be solved by the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a high-precision image measurement system and a measurement method thereof based on multi-feature fusion.

[0118] The beneficial effects of the present invention are:

[0119] By using color images, depth images, and lidar data sensors, more comprehensive object information can be obtained. Independent algorithms extract features such as color, texture, shape, and depth to better understand the various attributes of an object. Data collected by different sensors often exhibit spatiotemporal errors. Data fusion and spatiotemporal registration between sensor data using deep learning models can eliminate these errors, ensuring that the multidimensional data provided by each sensor can be analyzed uniformly within the same reference frame, improving data reliability. Image data may be subject to noise during actual acquisition, especially in low light, complex backgrounds, or in poor sensor performance. Fusion of selected features using deep learning models can uncover higher-level patterns in the data, thereby optimizing results. By integrating multiple data sources and using powerful optimization processing, the final output measurement results will have higher accuracy and greater adaptability. The advantages of this solution make it suitable for a variety of fields, including autonomous driving, robotic navigation, industrial inspection, and 3D modeling. It can efficiently and accurately process multi-sensor data and adapt to various dynamic changes, providing strong technical support for these fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] Figure 1 It is a flow chart of the high-precision image measurement system based on multi-feature fusion of the present invention. DETAILED DESCRIPTION

[0121] The principles and features of the present invention are described below. The examples provided are intended to illustrate the present invention only and are not intended to limit the scope of the present invention. The following paragraphs describe the present invention in more detail by way of example. The advantages and features of the present invention will become more apparent from the following description and claims.

[0122] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0123] Example

[0124] The technical solution adopted by the present invention to solve its technical problem is:

[0125] High-precision image measurement system based on multi-feature fusion, including:

[0126] Multi-feature extraction module, used to extract multiple feature information from the input image, including but not limited to color features, texture features, shape features, and depth features. Each feature is extracted using an independent algorithm;

[0127] Intelligent feature selection module, which uses adaptive algorithms to automatically select the most relevant features based on different scenarios and measurement requirements, and dynamically adjusts the importance weights of features;

[0128] Feature fusion module, which is used to intelligently fuse features from multiple sources using a deep learning model;

[0129] Multi-sensor data synchronization and registration module, used for accurate time synchronization and spatial registration of data from different types of sensors;

[0130] The noise suppression and environment adaptation module combines deep learning with traditional image processing methods to suppress noise in images in real time.

[0131] The measurement result optimization module is used to optimize the fused features and output high-precision measurement results. Through the optimization algorithm, the stability and accuracy of the measurement are further improved.

[0132] The system combines multi-sensor data and adopts advanced feature extraction and deep learning models to output high-precision measurement results; intelligent feature selection and noise suppression modules enable the system to automatically adapt to different environments and measurement requirements, enhancing the applicability of the system; through a variety of technical means (such as feature fusion, optimization processing, spatiotemporal alignment, etc.), the system can still maintain stability and accuracy in complex environments; the system can flexibly adjust parameters according to different sensors and data sources, adapt to various measurement tasks, and has high flexibility and scalability; through multi-sensor fusion, it can make up for the shortcomings of a single sensor and improve the integrity and reliability of the measurement.

[0133] The feature fusion module performs feature fusion by weighted synthesis of information from multiple channels; the multi-sensor data synchronization and registration module adopts a multimodal sensor data fusion algorithm; the noise suppression and environmental adaptation module is configured with the ability to automatically adapt to different lighting and environmental changes; the measurement result optimization module dynamically adjusts the optimization strategy according to the actual measurement situation.

[0134] Through weighted synthesis of multiple features, fusion of multiple sensor data and real-time noise suppression, the system can generate high-quality and high-precision measurement results, especially in complex environments; the noise suppression and environmental adaptation modules ensure that the system can cope with various lighting changes and environmental interference, so that the system can still maintain efficient operation in dynamic environments; the system can dynamically adjust the feature fusion weights, optimization strategies and noise suppression methods according to actual needs, making the measurement process more flexible and adaptable; multimodal sensor data fusion can integrate information from different sensors, provide more comprehensive measurement data, and make up for the possible shortcomings of a single sensor; through real-time optimization and adaptive processing, the system can ensure the stability and robustness of measurement results under various complex conditions.

[0135] The high-precision image measurement method based on multi-feature fusion includes the following steps:

[0136] Multiple sensors are used to collect image data of the object to be measured, including color images, depth images, and sensor data of lidar data;

[0137] Extract color, texture, shape and depth feature information from the collected image data, and each feature is processed and extracted by an independent algorithm;

[0138] According to the current environment and measurement requirements, an intelligent algorithm is used to select relevant features and assign appropriate weights to each feature;

[0139] By integrating the selected features through a deep learning model, the information of multiple features is integrated to synchronize the time and space of data from different sensors, eliminating the temporal and spatial errors between sensors;

[0140] The image data is subjected to noise suppression, adapted to illumination changes and complex background factors, and the fused feature data is finally optimized to output high-precision measurement results.

[0141] Through the fusion of multi-sensor data and deep learning models, it can provide higher-precision measurement results and is suitable for scenarios requiring higher precision; it has the ability to adapt to environments such as lighting changes and complex backgrounds, and can work stably in changing environments; it can effectively deal with problems such as spatiotemporal errors and noise interference between sensors, and enhance the robustness of the system in complex environments; through the automatic selection and weighted processing of intelligent algorithms, the system can efficiently process data, avoid unnecessary computing overhead, and optimize measurement results; through multi-feature fusion, the system can comprehensively analyze the object to be measured and improve the reliability and accuracy of the measurement results.

[0142] The method for using multiple sensors to collect image data of the object to be measured, including color images, depth images and lidar data, is as follows:

[0143] For the data collected by each sensor, denoising and alignment processing are performed, and after noise removal and color correction processing, I color After denoising and smoothing, the depth image D is obtained depth , by filtering the point cloud data and removing outliers, we can obtain the point cloud data P lidar ;

[0144] LiDAR point cloud data P lidar ={p1, p2, ..., p n The projection matrix is ​​aligned with the depth data of the color image to obtain the corresponding image pixel value. This process is achieved through the following projection formula:

[0145] p′ i =Proj(R·p i +t)

[0146] Among them, R is the rotation matrix, t is the translation vector, and p i The first point, p′ i is the projection position of the point in the image coordinate system;

[0147] The depth image D depth With the point cloud data P of the laser radar lidar Alignment is performed to obtain the corresponding relationship between the depth information and the lidar point cloud. The depth value d(x, y) is aligned with the point cloud data using the following formula:

[0148] P lidar =K·[x, y, d(r, y)] T

[0149] Among them, K is the camera intrinsic parameter matrix, (x, y) is the pixel coordinate in the image, d(x, y) is the depth value corresponding to the pixel, P lidar is the corresponding 3D coordinate in the point cloud data;

[0150] The data of different sensors are integrated through weighted average or deep learning model, and the data of each sensor is set up separately using I color 、D depth 、P lidar It indicates that its fusion process is carried out through the following steps;

[0151] From the color image I color , depth image D depth , LiDAR point cloud data P lidar Extract features from

[0152] Different sensors provide different types of information, and feature fusion is performed in a weighted manner, setting the weight w of each feature. color , w depth , w lidar , the weighted formula for feature fusion is:

[0153] F fusion (x, y) = w color ·f color (x, y) + w depth ·f depth (x, y) + w lidar ·f lidar (r, y)

[0154] Among them, f color (x, y) is the feature extracted from the color image, f depth (x, y) is the feature extracted from the depth image, f lidar (x, y) is the feature extracted from the lidar point cloud;

[0155] Based on the fused data, the final accurate measurement is performed to measure the three-dimensional coordinates P of the object. object , is calculated based on the fused image data as follows:

[0156] P object =f(I color , D depth , P lidar )

[0157] Where f is a function of the three-dimensional object position or other geometric parameters calculated according to the multi-feature fusion model.

[0158] By fusing information from different sensors, a more comprehensive and accurate description of objects can be obtained. The strengths of each sensor can offset the shortcomings of the others. Color images provide detailed surface color and texture information, facilitating object recognition. Depth images provide accurate surface shape and distance information. LiDAR point clouds provide precise three-dimensional spatial information. Through precise alignment and feature fusion, errors from individual sensors can be eliminated, significantly improving the accuracy of the final measurement results. During data processing, denoising and filtering steps effectively eliminate interference from sensor noise and outliers, ensuring data reliability and stability. This solution is adaptable to diverse environmental conditions, such as changing lighting and complex backgrounds. Depth images and LiDAR provide stable measurements under various environmental conditions, while the texture and color information provided by color images further enhances object recognition capabilities. Through intelligent algorithms such as deep learning and weighted averaging, the system automatically selects the most appropriate features for fusion, reducing the need for manual intervention and improving processing efficiency and accuracy. This method is highly scalable and can flexibly incorporate other sensor types to enhance the system's capabilities and adapt to a wider range of application scenarios.

[0159] Extract color, texture, shape, and depth feature information from the collected image data. Each feature is processed and extracted using an independent algorithm as follows:

[0160] The color histogram describes the frequency of pixel distribution in each color channel of the image. The color distribution of the image is composed of three channels C r , C g , C b describe:

[0161] ParseError:KaTeX parse error:Can′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N} \( r–r_i)

[0162] ParseError:KaTeX parse error:Can′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N} \( g-g_i)

[0163] ParseError:KaTeXparse error:Can′t use function′\(′in math mode atposition 25:\sum_{i=1}^{N} \( b-b_i)

[0164] Among them, H r , H g , H b is the histogram of the red, green and blue channels, is the Dirac function, r i , g i , b i For the image The red, green and blue channel values ​​of each pixel, where N is the number of pixels in the image;

[0165] Texture features are extracted through the gray-level co-occurrence matrix or local binary pattern method. The gray-level co-occurrence matrix describes the spatial relationship between the gray values ​​of pixel pairs and is used for texture analysis. The gray value of the image is set as I(x, y). The GLCM matrix is ​​defined as:

[0166] ParseError:KaTeX parse error:Can′t use function′\(′in math mode atposition 31:...=\sum_{x, y} \( I(x, y) = i, I(...

[0167] Where p(i, j, θ) represents the frequency of the pixel pair (i, j) in the specified direction θ, and -x, -y are the distances between the pixel pairs.

[0168] LBP is used to describe local texture structure. Let I(x, y) be the pixel value at position (x, y) of the image. The LBP operator converts the neighborhood of a pixel into a binary sequence:

[0169]

[0170] Where s(x) is a sign function. When x≥0, s(x)=1, otherwise s(x)=0. I(x n ,y n )

[0171] is the neighborhood pixel value, N is the size of the neighborhood;

[0172] Shape feature extraction is used to analyze the geometric shape of an image. The Hu moment is a set of invariant moments derived from the central moment of the image, which describes the geometric features of the shape. The central moment μ pq Defined as:

[0173]

[0174] Among them, (c x , c y ) is the centroid of the image, p and q are the orders of the moments, and the general form of the Hu moment is:

[0175] φ1=μ 20 +μ 02

[0176]

[0177] φ3=(μ 30 -3μ 12 ) 2 +(3μ 21 -μ 03 ) 2

[0178] By extracting the contour of the image, the methods include edge detection and contour simplification. Depth feature extraction is used to extract the three-dimensional geometric information of the object from the depth image. The algorithms include normal estimation and curvature analysis.

[0179] The normal of each point is calculated through the depth image to obtain the direction information of the object surface. For a point P(x, y) in the depth image, its normal is estimated by the neighborhood points. The depth of the image is set to D(x, y). The normal n is calculated as follows:

[0180]

[0181] in, is the gradient of the depth image at point (x, y);

[0182] Curvature is a feature that describes the degree of curvature of the surface, including Gaussian curvature and mean curvature. For each pixel in the depth image, the curvature is calculated by the second-order derivative;

[0183]

[0184] Among them, K g is the Gaussian curvature;

[0185] For the extraction of multiple features, multiple algorithms are performed in parallel to combine different features, and the color, texture, shape and depth features are spliced ​​into a high-dimensional feature vector for subsequent classification and detection tasks. The formula of the fusion process is:

[0186] F total =[F color , F texture , F shape , F depth ]

[0187] Among them, F color , F texture , F shape , F depth They are color, texture, shape and depth feature vectors respectively, and finally a comprehensive feature vector F is obtainedtotal .

[0188] By extracting multiple features such as color, texture, shape, and depth, the system can analyze objects from different dimensions, avoiding the limitations of a single feature and enhancing its ability to understand complex images. Fusion of multiple features can improve classification and recognition accuracy, especially when dealing with target detection tasks in complex environments, as fused features provide richer input information. By extracting different features in parallel, hardware resources can be fully utilized, improving processing efficiency. This approach can accelerate the computational process when processing large-scale data sets. The solution's feature extraction and fusion methods can cope with diverse environmental conditions such as changing lighting and complex backgrounds, improving the robustness of the system. Fusion of multiple sensor data allows the solution to adapt to multiple application areas, such as autonomous driving, robotic vision, and medical imaging. It can process color, depth, and geometric information, adapting to a wider range of application scenarios.

[0189] Based on the current environment and measurement requirements, the method of using intelligent algorithms to select relevant features and assigning appropriate weights to each feature is as follows:

[0190] The feature is selected based on its correlation with the target variable, and the Pearson correlation coefficient (ρ) is used to calculate the correlation between each feature and the target:

[0191]

[0192] Among them, Cov(X,Y) is the covariance of feature X and target variable Y, σ X and σ Y are the standard deviations of the feature and target variables, respectively;

[0193] Lasso regression selects features and automatically adjusts their weights through penalty terms. The objective function of Lasso regression is:

[0194]

[0195] Among them, y i is the target value, x i is the feature vector, w is the feature weight, and λ is the regularization parameter;

[0196] Weighting is to assign weights to each feature according to its importance. The weights are adaptively obtained through a learning algorithm using the following formula:

[0197] w=argmin w {L(w)+λ||w||2}

[0198] Where w is the feature weight vector, L(w) is the loss function, λ is the regularization coefficient, and ||w||2 is the L2 regularization term;

[0199] When using decision tree, random forest or gradient boosted tree models, extract feature importance directly from the model. Feature importance in random forest is j Calculate the impact of each feature on the decision tree:

[0200]

[0201] Among them, S is the feature subset, S i For the specific features, accuracy(S i ) is based on the subset S i Model accuracy;

[0202] PSO is used to find the optimal feature selection scheme and the corresponding weight distribution. The position of the particle represents the combination of feature selection, the speed of the particle represents the weight adjustment, and the optimization goal is to minimize the loss function, error or loss:

[0203]

[0204] Combining feature selection and weighting, we get the following optimization formula:

[0205]

[0206] Among them, F selected is the weighted feature set, F i For the Features, w i is the corresponding weight, and M is the number of features.

[0207] Using the Pearson correlation coefficient to evaluate the correlation between features and target variables can effectively remove low-correlation features, retain the most informative features, and improve model performance; methods such as Lasso regression, decision trees, and random forests can automatically select features related to the target variable and assign appropriate weights to each feature; by integrating multiple feature selection algorithms and optimization algorithms, this solution can adapt to different data characteristics and environmental changes and achieve good performance in various tasks; through the combination of particle swarm optimization and other intelligent algorithms, the optimal feature combination and weight distribution can be found in a shorter time, avoiding the tedious process of traditional manual adjustment and improving computational efficiency; by combining different feature selection methods and weighting strategies, this solution can enhance the robustness of the model to different data sets and environments, and improve the stability of classification and prediction tasks.

[0208] The method of fusing the selected features through the deep learning model, integrating the information of multiple features, and performing time synchronization and spatial registration on the data of different sensors to eliminate the spatiotemporal errors between sensors is as follows:

[0209] The purpose of feature fusion is to merge features from different sources. A neural network is used to perform feature fusion. The input of the model is the features of multiple sensor data, and the output is the fused features. Its basic form is expressed as:

[0210] z=Fusion(x1,x2,...,x n )

[0211] Among them, x1, x2, ..., x n are the data features from different sensors;

[0212] z is the fused feature representation;

[0213] Set up two sensors S1 and S2, whose data sampling times are t1 and t2 respectively. Align them through interpolation and set up a deep learning-based model to learn the time synchronization transformation:

[0214] t2=f(t1,θ)

[0215] Where t1 is the timestamp of sensor 1;

[0216] t2 is the timestamp of sensor 2;

[0217] f(t1,θ) is the time synchronization function learned, where θ is the model parameter;

[0218] To ensure that multiple sensor data are spatially aligned and eliminate errors caused by differences in sensor positions, two sensors S1 and S2 are set up, whose spatial coordinates are p1 and p2 respectively. The purpose of spatial registration is to transform p1 to p2 through a transformation function g(p1, θ):

[0219] p2=g(p1,θ)

[0220] Where p1 is the spatial position of sensor 1;

[0221] p2 is the spatial position of sensor 2;

[0222] g(p1,θ) is the spatial registration function;

[0223] The elimination of spatiotemporal errors is achieved by fusing the data after time synchronization and spatial registration. The deep learning model for eliminating errors is a loss function that optimizes the data after time synchronization and spatial registration to minimize the error. The model output is set as The target is y, and the loss function is expressed as:

[0224]

[0225] in, The data output by the deep learning model has undergone spatiotemporal synchronization and spatial registration;

[0226] y i is the target data;

[0227] N is the total number of data.

[0228] While traditional methods rely on manual design and rules, this solution uses deep learning to automatically learn the relationships between features, spatiotemporal synchronization, and registration methods, significantly reducing manual intervention and improving efficiency. This method is adaptable enough to handle a variety of sensor data and is suitable for different data types and sensor configurations. Deep learning models can extract more complex data features and spatiotemporal synchronization patterns, improving system accuracy, especially when processing heterogeneous data. Joint training for temporal synchronization, spatial registration, and feature fusion not only reduces errors but also ensures spatiotemporal consistency across data sources, improving data quality.

[0229] The method for suppressing noise in image data, adapting to illumination changes and complex background factors, and finally optimizing the fused feature data to output high-precision measurement results is as follows:

[0230] Noise suppression and illumination change adaptation are performed through filtering and normalization methods. The input image is set as L noisy (x, y), where x and y are pixel coordinates. Noise suppression is achieved through image denoising algorithms, and illumination adaptation is achieved through brightness normalization and histogram equalization. The calculation formula is as follows:

[0231]

[0232]

[0233] in, represents the noise suppression operation;

[0234] Represents light adaptation operation;

[0235] Complex background factors are processed through background modeling or background subtraction, which is completed through deep learning models or classic background modeling methods. The calculation formula is:

[0236]

[0237] in, represents the background suppression algorithm;

[0238] In image processing, feature fusion is used to combine feature data from multiple different sources, and to establish multiple feature images {F1, F2, ..., F n}, then the feature fusion operation formula is expressed as:

[0239]

[0240] in, It is the feature fusion operation;

[0241] The fused feature data is optimized and processed through the deep learning model, and the fused feature data is set as F fusion , and finally optimize the processing to output high-precision measurement results R final , the calculation formula is:

[0242]

[0243] in, represents the final optimization operation, performed by the deep neural network, and θ is the network parameter.

[0244] Removing noise from images can effectively improve image quality and make subsequent processing more accurate, which is especially important in low-quality or noisy images; changes in lighting conditions are a common problem in image processing. Through illumination adaptation methods, it is possible to ensure that images remain consistent under different lighting conditions, thereby improving the robustness of image processing; through background modeling or background subtraction, complex background factors in the image can be effectively processed to ensure the accuracy of target object detection, which is especially important in dynamic environments; feature fusion can integrate information from multiple processing steps, utilize the complementarity of different features, enhance the overall performance of the image, and make subsequent analysis more comprehensive and accurate; the final optimization processing, through the refined training of deep neural networks, makes the output measurement results have higher accuracy, can cope with complex image data, and provide more reliable analysis.

[0245] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the high-precision image measurement system and measurement method based on multi-feature fusion as described above are implemented.

[0246] This embodiment also provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the high-precision image measurement system and measurement method based on multi-feature fusion as described above are implemented.

[0247] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0248] Those skilled in the art will clearly understand that for the sake of convenience and brevity in description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0249] The above embodiments of the present invention are not intended to limit the scope of protection of the present invention, and the implementation methods of the present invention are not limited thereto. All other modifications, replacements or changes made to the above structures of the present invention based on the above contents of the present invention, in accordance with common technical knowledge and customary means in this field, without departing from the above basic technical ideas of the present invention, should fall within the scope of protection of the present invention.

Claims

1. High-precision image measurement system based on multi-feature fusion, characterized by: Includes: Multi-feature extraction module, used to extract multiple feature information from the input image, including but not limited to color features, texture features, shape features, and depth features. Each feature is extracted using an independent algorithm; Intelligent feature selection module, which uses adaptive algorithms to automatically select the most relevant features based on different scenarios and measurement requirements, and dynamically adjusts the importance weights of features; Feature fusion module, which is used to intelligently fuse features from multiple sources using a deep learning model; Multi-sensor data synchronization and registration module, used for accurate time synchronization and spatial registration of data from different types of sensors; The noise suppression and environment adaptation module combines deep learning with traditional image processing methods to suppress noise in images in real time. The measurement result optimization module is used to optimize the fused features and output high-precision measurement results. Through the optimization algorithm, the stability and accuracy of the measurement are further improved.

2. The high-precision image measurement system based on multi-feature fusion according to claim 2 is characterized in that: The feature fusion module performs feature fusion by weighted synthesis of information from multiple channels; the multi-sensor data synchronization and registration module adopts a multimodal sensor data fusion algorithm; the noise suppression and environmental adaptation module is configured with the ability to automatically adapt to different lighting and environmental changes; the measurement result optimization module dynamically adjusts the optimization strategy according to the actual measurement situation.

3. A high-precision image measurement method based on multi-feature fusion, characterized in that: The following steps are involved: Multiple sensors are used to collect image data of the object to be measured, including color images, depth images, and sensor data of lidar data; Extract color, texture, shape and depth feature information from the collected image data, and each feature is processed and extracted by an independent algorithm; According to the current environment and measurement requirements, an intelligent algorithm is used to select relevant features and assign appropriate weights to each feature; By integrating the selected features through a deep learning model, the information of multiple features is integrated to synchronize the time and space of data from different sensors, eliminating the temporal and spatial errors between sensors; The image data is subjected to noise suppression, adapted to illumination changes and complex background factors, and the fused feature data is finally optimized to output high-precision measurement results.

4. The high-precision image measurement method based on multi-feature fusion according to claim 3 is characterized in that: The method for using multiple sensors to collect image data of the object to be measured, including color images, depth images and lidar data, is as follows: For the data collected by each sensor, denoising and alignment processing are performed, and after noise removal and color correction processing, I color After denoising and smoothing, the depth image D is obtained depth , by filtering the point cloud data and removing outliers, we can obtain the point cloud data P lidar ; LiDAR point cloud data P lidar ={p1, p2, ..., p n The projection matrix is ​​aligned with the depth data of the color image to obtain the corresponding image pixel value. This process is achieved through the following projection formula: p′ i =Proj(R·p i +t) Among them, R is the rotation matrix, t is the translation vector, and p i is the i-th point in the lidar point cloud, p′ i is the projection position of the point in the image coordinate system; The depth image D depth With the point cloud data P of the laser radar lidar Alignment is performed to obtain the corresponding relationship between the depth information and the lidar point cloud. The depth value d(x, y) is aligned with the point cloud data using the following formula: P lidar =K·[x,y,d(x,y)] T Among them, K is the camera intrinsic parameter matrix, (x, y) is the pixel coordinate in the image, d(x, y) is the depth value corresponding to the pixel, P lidar is the corresponding 3D coordinate in the point cloud data; The data of different sensors are integrated through weighted average or deep learning model, and the data of each sensor is set up separately using I color 、D depth 、P lidar It indicates that its fusion process is carried out through the following steps; From the color image I color , depth image D depth , LiDAR point cloud data P lidar Extract features from Different sensors provide different types of information, and feature fusion is performed in a weighted manner, setting the weight w of each feature. color , w depth , w lidar , the weighted formula for feature fusion is: F fusion (x,y)=w color ·f color (x,y)+w depth ·f depth (x,y)+w lidar ·f lidar (x,y) Among them, f color (x, y) is the feature extracted from the color image, f depth (x, y) is the feature extracted from the depth image, f lidar (x, y) is the feature extracted from the lidar point cloud; Based on the fused data, the final accurate measurement is performed to measure the three-dimensional coordinates P of the object. object , is calculated based on the fused image data as follows: P object =f(I color ,D depth ,P lidar ) Where f is a function of the three-dimensional object position or other geometric parameters calculated according to the multi-feature fusion model.

5. The high-precision image measurement method based on multi-feature fusion according to claim 4 is characterized in that: Extract color, texture, shape, and depth feature information from the collected image data. Each feature is processed and extracted using an independent algorithm as follows: The color histogram describes the frequency of pixel distribution in each color channel of the image. The color distribution of the image is composed of three channels C r , C g , C b describe: ParseError:KaTeX parse error:Cat′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N} \( r-r_i) ParseError:KaTeX parse error:Cat′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N} \( g-g_i) ParseError:KaTeX parse error:Can′t use function′\(′in math mode atposition 25:...\sum_{i=1}^{N} \( b-b_i) Among them, H r , H g , H b is the histogram of the red, green and blue channels, is the Dirac function, r i , g i , b i is the red, green and blue channel value of the i-th pixel in the image, and N is the number of pixels in the image; Texture features are extracted through the gray-level co-occurrence matrix or local binary pattern method. The gray-level co-occurrence matrix describes the spatial relationship between the gray values ​​of pixel pairs and is used for texture analysis. The gray value of the image is set as I(x, y). The GLCM matrix is ​​defined as: ParseError:KaTeX parse error:Can′t use function′\(′in math mode atposition 31:...=\sum_{x,y} \( I(x,y)=i,I(... Where p(i, j, θ) represents the frequency of the pixel pair (i, j) in the specified direction θ, and _x, _y are the distances between the pixel pairs. LBP is used to describe local texture structure. Let I(x, y) be the pixel value at position (x, y) of the image. The LBP operator converts the neighborhood of a pixel into a binary sequence: Where s(x) is a sign function. When x≥0, s(x)=1, otherwise s(x)=0. I(x n ,y n ) is the neighborhood pixel value, and N is the size of the neighborhood; Shape feature extraction is used to analyze the geometric shape of an image. The Hu moment is a set of invariant moments derived from the central moment of the image, which describes the geometric features of the shape. The central moment μ pq Defined as: Among them, (c x , c y ) is the centroid of the image, p and q are the orders of the moments, and the general form of the Hu moment is: φ1=μ 20 +m 02 φ3=(μ 30 -3m 12 ) 2 +(3m 21 -m 03 ) 2 By extracting the contour of the image, the methods include edge detection and contour simplification. Depth feature extraction is used to extract the three-dimensional geometric information of the object from the depth image. The algorithms include normal estimation and curvature analysis. The normal of each point is calculated through the depth image to obtain the direction information of the object surface. For a point P(x, y) in the depth image, its normal is estimated by the neighborhood points. The depth of the image is set to D(x, y). The normal n is calculated as follows: in, is the gradient of the depth image at point (x, y); Curvature is a feature that describes the degree of curvature of the surface, including Gaussian curvature and mean curvature. For each pixel in the depth image, the curvature is calculated by the second-order derivative; Among them, K g is the Gaussian curvature; For the extraction of multiple features, multiple algorithms are performed in parallel to combine different features, and the color, texture, shape and depth features are spliced ​​into a high-dimensional feature vector for subsequent classification and detection tasks. The formula of the fusion process is: F total =[F color ,F texture ,F shape ,F depth ] Among them, F color , F texture , F shape , F depth They are color, texture, shape and depth feature vectors respectively, and finally a comprehensive feature vector F is obtained total .

6. The high-precision image measurement method based on multi-feature fusion according to claim 5 is characterized in that: Based on the current environment and measurement requirements, the method of using intelligent algorithms to select relevant features and assigning appropriate weights to each feature is as follows: The feature is selected based on its correlation with the target variable, and the Pearson correlation coefficient (ρ) is used to calculate the correlation between each feature and the target: Among them, Cov(X, Y) is the covariance of feature X and target variable Y, σ X and σ Y are the standard deviations of the feature and target variables, respectively; Lasso regression selects features and automatically adjusts their weights through penalty terms. The objective function of Lasso regression is: Among them, y i is the target value, x i is the feature vector, w is the feature weight, and λ is the regularization parameter; Weighting is to assign weights to each feature according to its importance. The weights are adaptively obtained through a learning algorithm using the following formula: w=argmin w {L(w)+λ||w||2} Where w is the feature weight vector, L(w) is the loss function, λ is the regularization coefficient, and ||w||2 is the L2 regularization term; When using decision tree, random forest or gradient boosted tree models, extract feature importance directly from the model. Feature importance in random forest is j Calculate the impact of each feature on the decision tree: Among them, S is the feature subset, S i For the specific features, accuracy(S i ) is based on the subset S i Model accuracy; PSO is used to find the optimal feature selection scheme and the corresponding weight distribution. The position of the particle represents the combination of feature selection, the speed of the particle represents the weight adjustment, and the optimization goal is to minimize the loss function, error or loss: Combining feature selection and weighting, we get the following optimization formula: Among them, F selected is the weighted feature set, F i is the i-th feature, w i is the corresponding weight, and M is the number of features.

7. The high-precision image measurement method based on multi-feature fusion according to claim 6 is characterized in that: The method of fusing the selected features through the deep learning model, integrating the information of multiple features, and performing time synchronization and spatial registration on the data of different sensors to eliminate the spatiotemporal errors between sensors is as follows: The purpose of feature fusion is to merge features from different sources. A neural network is used to perform feature fusion. The input of the model is the features of multiple sensor data, and the output is the fused features. Its basic form is expressed as: z=Fusion(x1,x2,...,x n ) Among them, x1, x2, ..., x n are the data features from different sensors; z is the fused feature representation; Set up two sensors S1 and S2, whose data sampling times are t1 and t2 respectively. Align them through interpolation and set up a deep learning-based model to learn the time synchronization transformation: t2=f(t1,θ) Where t1 is the timestamp of sensor 1; t2 is the timestamp of sensor 2; f(t1, θ) is the time synchronization function learned, where θ is the model parameter; To ensure that multiple sensor data are spatially aligned and eliminate errors caused by differences in sensor positions, two sensors S1 and S2 are set up, whose spatial coordinates are p1 and p2 respectively. The purpose of spatial registration is to transform p1 to p2 through a transformation function g(p1, θ): p2=g(p1,θ) Where p1 is the spatial position of sensor 1; p2 is the spatial position of sensor 2; g(p1,θ) is the spatial registration function; The elimination of spatiotemporal errors is achieved by fusing the data after time synchronization and spatial registration. The deep learning model for eliminating errors is a loss function that optimizes the data after time synchronization and spatial registration to minimize the error. The model output is set as The target is y, and the loss function is expressed as: in, The data output by the deep learning model has undergone spatiotemporal synchronization and spatial registration; y i is the target data; N is the total number of data.

8. The high-precision image measurement method based on multi-feature fusion according to claim 7 is characterized in that: The method for suppressing noise in image data, adapting to illumination changes and complex background factors, and finally optimizing the fused feature data to output high-precision measurement results is as follows: Noise suppression and illumination change adaptation are performed through filtering and normalization methods. The input image is set as I noisy (x, y), where x and y are pixel coordinates. Noise suppression is achieved through image denoising algorithms, and illumination adaptation is achieved through brightness normalization and histogram equalization. The calculation formula is as follows: in, represents the noise suppression operation; Represents light adaptation operation; Complex background factors are processed through background modeling or background subtraction, which is completed through deep learning models or classic background modeling methods. The calculation formula is: in, represents the background suppression algorithm; In image processing, feature fusion is used to combine feature data from multiple different sources, and to establish multiple feature images {F1, F2, ..., F n }, then the feature fusion operation formula is expressed as: in, It is the feature fusion operation; The fused feature data is optimized and refined through the deep learning model, and the fused feature data is set as E fusion , and finally optimize the processing to output high-precision measurement results R final , the calculation formula is: in, represents the final optimization operation, performed by the deep neural network, and θ is the network parameter.

9. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method realizes the high-precision image measurement method based on multi-feature fusion as claimed in any one of claims 3 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the high-precision image measurement method based on multi-feature fusion as described in any one of claims 3 to 8 is implemented.

Images