Multi-spectral imaging fused pattern recognition image acquisition method and device

Abstract

The invention relates to the technical field of multispectral imaging, in particular to a multispectral imaging fused pattern recognition image acquisition method and device, and the method comprises the following steps: obtaining a multiband substrate response value through a multispectral sensor, generating response difference distribution through moving average filtering, inputting cross-correlation analysis, extracting a space offset vector, and compensating displacement; after the signal intensity of the mark points is captured, Gaussian fitting is carried out to generate registration coordinates, triangular mesh division is carried out, offset is calculated, interpolation mapping is carried out, a registration error distribution diagram is generated, bilinear interpolation remapping is carried out to collect time sequence parameters, and a multispectral imaging data set is output. According to the method, noise is eliminated through moving average filtering, offset vectors are extracted in a cross-correlation mode, array displacement is compensated, mark points are positioned through two-dimensional Gaussian fitting, a registration error graph is constructed through triangular meshes, parameters are collected through bilinear interpolation remapping, geometric distortion is corrected, feature consistency and collection synchronism are enhanced, and the imaging resolution ratio and the signal-to-noise ratio are improved.

Description

Technical Field

[0001] This invention relates to the field of multispectral imaging technology, and in particular to a method and device for pattern recognition image acquisition that integrates multispectral imaging. Background Technology

[0002] The field of multispectral imaging technology mainly involves acquiring image information of the same scene or target using light in multiple differentiated bands to achieve more comprehensive object recognition and analysis. This technology includes core aspects such as multi-band light source control, imaging sensor design, optical system integration, and synchronous acquisition and processing of multi-band images. It is widely used in fields such as remote sensing detection, medical diagnosis, food safety detection, and agricultural monitoring.

[0003] Among them, the traditional pattern recognition image acquisition method of fusion multispectral imaging refers to the process of combining and recombining image data acquired from different bands under specific rules during multispectral imaging to form an acquired image for subsequent pattern recognition analysis. After acquiring multi-band images sequentially using time-division exposure, the multi-band images are superimposed at the pixel level or statistical features are extracted in a preset order in the acquisition terminal to obtain an image containing multi-band feature information, so as to provide a data foundation for subsequent sample recognition and target detection.

[0004] Traditional multispectral imaging technology relies on time-sharing exposure mechanisms to acquire multi-band images. In dynamic scenes, images of different bands are out of time, resulting in motion artifacts after fusion. Pixel-level stacking methods ignore the spatial offset characteristics of the sensor array, leading to the accumulation of geometric registration errors across band images. Pre-ordered statistical feature extraction lacks the ability to self-match local response differences. Multispectral image feature fusion is distorted in complex lighting or sudden changes in target reflection, resulting in decreased spatial resolution and signal-to-noise ratio, which affects the accuracy and stability of target detection and recognition. Summary of the Invention

[0005] To address the technical problems of traditional multispectral imaging techniques that rely on time-sharing exposure mechanisms to acquire multi-band images, resulting in temporal misalignment between images of different bands in dynamic scenes and motion artifacts after fusion, pixel-level overlay methods neglecting the spatial offset characteristics of the sensor array leading to accumulated geometric registration errors across bands, the lack of self-matching capability for local response differences in preset-order statistical feature extraction, and distortion of multispectral image feature fusion in complex lighting or sudden changes in target reflection, leading to decreased spatial resolution and signal-to-noise ratio, and affecting target detection accuracy and recognition stability, this invention provides a pattern recognition image acquisition method and device that fuses multispectral imaging. The technical solution is as follows:

[0006] On the one hand, a pattern recognition image acquisition method integrating multispectral imaging is provided, the method comprising:

[0007] S1: Obtain the base response values ​​of the target area under multiple bands through a multispectral sensor, and use a moving average filtering algorithm to perform convolution operation on the base response values ​​of adjacent bands to generate the response difference distribution of adjacent bands;

[0008] S2: Input the response difference distribution of adjacent bands into the cross-correlation analysis algorithm, extract the spatial offset vector between adjacent bands, perform displacement compensation on the sensor array based on the offset vector amplitude, and generate a set of displacement compensation parameters.

[0009] S3: Call the compensated three-dimensional coordinates in the displacement compensation parameter set, capture the characteristic signal intensity of the preset spectral marker points through a multispectral sensor, perform two-dimensional Gaussian fitting using the least squares method for the regional signal, and generate a registered marker point coordinate dataset.

[0010] S4: Divide the registered marker point coordinate dataset into spatial triangular meshes, calculate the offset between the side length of the triangular unit and the preset geometric layout, map it to a two-dimensional plane coordinate system using interpolation, and generate a spatial registration error distribution map.

[0011] S5: Input the spatial registration error distribution map into the bilinear interpolation algorithm to perform spatial domain remapping calculation on the acquisition time sequence parameters of the multispectral sensor array, and output the multispectral imaging dataset.

[0012] As a further aspect of the present invention, the adjacent band response difference distribution includes distribution characteristics, statistical amplitude, and spectral difference index; the displacement compensation parameter set includes compensation amplitude, compensation direction, and compensation accuracy; the registered marker point coordinate dataset includes marker point spatial coordinates, attribute labels, and signal strength information; the spatial registration error distribution map includes error value distribution, spatial position mapping, and error extreme value distribution; and the multispectral imaging dataset includes remapping acquisition parameters, acquisition timing information, and registered multispectral data.

[0013] As a further aspect of the present invention, the specific steps of S1 include:

[0014] S101: Acquire reflection signals of the target area under multiple bands through a multispectral sensor, collect multi-band floor response data, statistically analyze the numerical distribution of the floor response based on the response signals of each band, establish a complete band response database by collecting the sampled values ​​under different bands, and generate band floor response values.

[0015] S102: Based on the base response value of the band, a moving average filtering algorithm is used to set a sliding window for the base response values ​​of adjacent bands, and a weighted average of adjacent response data is performed using a convolution processing method. The convolution result is matched with the original band order to generate a filtered average response value.

[0016] S103: Based on the average response value of the filter, the difference calculation method is used to perform numerical difference on the average response values ​​of each pair of adjacent bands, summarize the difference results between the bands, and statistically analyze the range of difference variation to obtain the distribution of response differences between adjacent bands.

[0017] As a further embodiment of the present invention, the sliding window width of the moving average filtering algorithm is set by self-matching according to the signal-to-noise ratio parameters of the multispectral sensor in the differentiated bands, and the convolution weights are corrected by parameters from the original acquired data.

[0018] The band substrate response value is a statistical characteristic value of the target area reflection signal collected by the multispectral sensor in multiple bands, and is a statistical characteristic of the target area reflection signal.

[0019] The filtered average response value is the result obtained by weighting and smoothing the base response values ​​of adjacent bands using a moving average filtering algorithm, and represents the statistical characteristics of the reflected signal in the target area.

[0020] As a further aspect of the present invention, the specific steps of S2 include:

[0021] S201: Obtain the response difference distribution data of the adjacent bands, set the normalization parameters for normalization processing, select the cross-correlation analysis method to set the operation window and rules, input the normalized pixels point by point into the analysis algorithm, calculate the correlation coefficient according to the pixel sliding window, record the result of each point, and form a cross-correlation coefficient sequence.

[0022] The normalization process refers to linearly scaling the response difference distribution data according to a preset normalization parameter to stabilize its value within a specified range. The initial value of the parameter is set based on the original sample statistical mean.

[0023] S202: Based on the cross-correlation coefficient sequence, identify the trend of pixel correlation change, record the original coordinates and the coordinates corresponding to the peak, calculate the coordinate offset, organize the offset values ​​between pixels, and generate a spatial offset vector set;

[0024] S203: Call the spatial offset vector set, retrieve the corresponding offset vector amplitude for each sensor array channel one by one, adjust the position parameters of the channel pixels in combination with the offset vector amplitude, summarize the adjusted channel pixel compensation data, and establish a displacement compensation parameter set.

[0025] As a further aspect of the present invention, the specific steps of S3 include:

[0026] S301: Call the compensated three-dimensional coordinates in the displacement compensation parameter set, set the sampling path of the multispectral sensor, and collect the characteristic signal intensity at the preset spectral marker points in sequence according to the spatial coordinates. For each spatial coordinate point, the measured raw signal is spatially grouped, the signal distribution data of the marker points is statistically analyzed, and a spatial signal distribution value is generated.

[0027] S302: Call the spatial signal distribution value, integrate the two-dimensional matrix signal intensity according to the spatial neighborhood for each group of signal distribution areas, calculate the spatial neighborhood weighted mean, fit the two-dimensional signal intensity distribution by the least squares method, and use the two-dimensional Gaussian function to extract the distribution feature parameters of the region to obtain the Gaussian parameter coefficients.

[0028] S303: Based on the Gaussian parameter coefficients, determine the signal center position of the marker point in the two-dimensional space, match the center coordinates obtained from the positioning with the compensated three-dimensional coordinates, calculate their corresponding Euclidean distance, filter out the three-dimensional points with small Euclidean distances, and obtain the registered marker point coordinate dataset.

[0029] As a further aspect of the present invention, the spatial signal distribution value is the signal intensity data collected by the multispectral sensor at the spatial coordinate point, reflecting the spectral signal characteristics of the detected object at that coordinate point.

[0030] The Gaussian parameter coefficients are characteristic parameters of the center position, amplitude, and diffusion width obtained by fitting the signal distribution based on a two-dimensional Gaussian distribution function.

[0031] The least squares method is a mathematical method that optimizes parameters by minimizing the sum of squared errors between the fitted model and the actual observed data. The initial values ​​of the parameters are set based on the statistical characteristics of the sampled data.

[0032] The two-dimensional Gaussian function characterizes the spatial distribution of the signal, and its parameters include amplitude, center coordinates, and standard deviation. The initial value is determined based on the distribution center, extrema, and diffusion range of the sampled signal.

[0033] As a further aspect of the present invention, the specific steps of S4 include:

[0034] S401: Based on the registered marker point coordinate dataset, all marker points are filtered in sequence, and the marker points are grouped and corresponding triangular units are constructed using a spatial triangular mesh partitioning method. Three sides are extracted for each group of triangular units, and the actual side length is calculated for the coordinate data of the marker points at both ends of each side to generate the side length value of the triangular unit.

[0035] S402: Call the side length value of the triangular unit, extract the corresponding preset geometric layout standard side length for each triangular unit side, calculate the numerical offset one by one according to the actual side length and the standard side length, and classify and statistically analyze the offset data of the triangular unit to obtain the spatial offset of the triangular unit.

[0036] S403: Based on the spatial offset of the triangular unit, the spatial offset data is mapped to a two-dimensional plane coordinate system using an interpolation method. The error distribution information of multiple points in the two-dimensional plane is allocated through the interpolation scheme to obtain a spatial registration error distribution map.

[0037] As a further aspect of the present invention, the specific steps of S5 include:

[0038] S501: Obtain the spatial registration error distribution map data, and according to the spatial coordinate points of each group of pixels in the distribution map, use the spatial mapping rules of the bilinear interpolation algorithm to perform a remapping calculation on the corresponding relationship of the coordinate points to generate spatial coordinate remapping values.

[0039] The spatial coordinate remapping value refers to the remapping spatial coordinates of the pixel after spatial registration error compensation.

[0040] S502: Based on the spatial coordinate remapping value, combined with the original acquisition time series parameters of the multispectral sensor array, for each set of time series parameters and pixel coordinates, the error distribution and the pixels under the time series acquisition parameters are mapped to new spatial positions through the bilinear interpolation algorithm to obtain the multispectral time series spatial mapping value.

[0041] The multispectral temporal spatial mapping value refers to the new spatial location to which a pixel is mapped after combining temporal parameters and channel information;

[0042] S503: Based on the multispectral temporal spatial mapping value, the channels of the original data acquired by the multispectral sensor array are reordered in spatial order, the mapping sequence of the channel data is integrated, and a multispectral imaging dataset is established.

[0043] On the other hand, a pattern recognition image acquisition device integrating multispectral imaging is provided. This device is used to execute the aforementioned pattern recognition image acquisition method integrating multispectral imaging. The device includes:

[0044] The floor response filtering module is used to acquire multi-band floor response values ​​of the target area through a multispectral sensor, and to perform convolution operation on the floor response values ​​of adjacent bands using a moving average filtering algorithm to generate the differential distribution of responses of adjacent bands, which is then transmitted to the spatial offset compensation module.

[0045] The spatial offset compensation module is used to call the response difference distribution of adjacent bands, input the difference distribution into the cross-correlation analysis algorithm to extract the spatial offset vector between adjacent bands, perform displacement compensation on the three-dimensional coordinates of the sensor array based on the offset vector amplitude, generate a set of displacement compensation parameters, and pass them to the marker registration module.

[0046] The marker registration module is used to call the compensated three-dimensional coordinates in the displacement compensation parameter set, capture the characteristic signal intensity of the preset spectral markers in a directional manner, perform two-dimensional Gaussian fitting on the regional signal using the least squares method, generate a registered marker coordinate dataset, and pass it to the error mapping module.

[0047] The error mapping module is used to divide the registered marker point coordinate dataset into spatial triangular meshes, calculate the offset between the side length of the triangular unit and the preset geometric layout, map it to a two-dimensional plane coordinate system by interpolation, generate a spatial registration error distribution map, and pass it to the temporal remapping module.

[0048] The temporal remapping module is used to call the spatial registration error distribution map, input the acquisition timing parameters of the multispectral sensor array into the bilinear interpolation algorithm, perform spatial domain pixel coordinate transformation on the timing parameters according to the error distribution, and output the multispectral imaging dataset.

[0049] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0050] By introducing a moving average filtering algorithm to convolve the base response values ​​of adjacent bands, sensor noise interference is eliminated. A cross-correlation analysis algorithm extracts the spatial offset vector and dynamically compensates for the displacement deviation of the sensor array. The least squares two-dimensional Gaussian fitting accurately locates the coordinates of spectral marker points. Triangular meshing combined with geometric layout offset calculation constructs a spatial registration error distribution map. A bilinear interpolation algorithm performs spatial domain remapping of the acquired time-series parameters, corrects geometric distortion in the multispectral image fusion process, and improves the spatial consistency of multi-band features. The synergistic effect of time-series optimization and error compensation enhances the synchronization of data acquisition. The output imaging dataset has higher spatial resolution and signal-to-noise ratio, providing a low-distortion and high-reliability underlying data foundation for pattern recognition. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0052] Figure 2 This is a flowchart of the equipment of the present invention. Detailed Implementation

[0053] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0054] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0055] In the embodiments of the present invention, the terms "image" and "picture" may be used interchangeably. It should be noted that, when the distinction between them is not emphasized, their intended meanings are the same. The terms "of," "corresponding," and "corresponding" may be used interchangeably. It should be noted that, when the distinction between them is not emphasized, their intended meanings are the same.

[0056] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0057] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0058] Please see Figure 1 This invention provides a method for pattern recognition image acquisition that integrates multispectral imaging. The processing flow of this method may include the following steps:

[0059] S1: Obtain the base response values ​​of the target area under multiple bands through a multispectral sensor, and use a moving average filtering algorithm to perform convolution operation on the base response values ​​of adjacent bands to generate the response difference distribution of adjacent bands;

[0060] S2: Input the response difference distribution of adjacent bands into the cross-correlation analysis algorithm, extract the spatial offset vector between adjacent bands, perform displacement compensation on the sensor array based on the offset vector amplitude, and generate a set of displacement compensation parameters;

[0061] S3: Call the compensated three-dimensional coordinates in the displacement compensation parameter set, capture the characteristic signal intensity of the preset spectral marker points through a multispectral sensor, perform two-dimensional Gaussian fitting using the least squares method for the regional signal, and generate a registered marker point coordinate dataset.

[0062] S4: Divide the registered marker point coordinate dataset into spatial triangular meshes, calculate the offset between the side length of the triangular unit and the preset geometric layout, map it to the two-dimensional plane coordinate system by interpolation, and generate a spatial registration error distribution map.

[0063] S5: Input the spatial registration error distribution map into the bilinear interpolation algorithm to perform spatial domain remapping calculation on the acquisition time-series parameters of the multispectral sensor array, and output the multispectral imaging dataset;

[0064] The distribution of response differences between adjacent bands includes distribution characteristics, statistical amplitude, and spectral difference indices. The set of displacement compensation parameters includes compensation amplitude, compensation direction, and compensation accuracy. The registered marker point coordinate dataset includes marker point spatial coordinates, attribute labels, and signal strength information. The spatial registration error distribution map includes error numerical distribution, spatial location mapping, and error extreme value distribution. The multispectral imaging dataset includes remapping acquisition parameters, acquisition timing information, and registered multispectral data.

[0065] Specifically, the steps of S1 are as follows:

[0066] S101: Acquire reflection signals of the target area under multiple bands through a multispectral sensor, collect multi-band floor response data, statistically analyze the numerical distribution of the floor response based on the response signals of each band, establish a complete band response database by collecting the sampled values ​​under different bands, and generate band floor response values.

[0067] To acquire the reflection signals of a target area from a multispectral sensor across multiple bands, the target area needs to be spatially divided into grids, for example, each pixel is set to a resolution of 10×10cm. Data acquisition is performed grid by grid. The multispectral sensor used for acquisition is set on a carrier platform (such as a ground rail or UAV platform), and its configuration includes 16 discrete bands in the range of 400nm to 1000nm. The center wavelength and bandwidth of each band are preset, for example, the center wavelengths are 420, 450, 480, 510... to 960nm, and the bandwidth is 10nm. When acquiring the reflection signal, the sensor integration time needs to be adjusted to keep the incident irradiance value within the sensor's linear response range. The reflectance is converted using a calibration plate or grayscale plate reflectance reference value to eliminate the interference of ambient light and changes in the angle of sunlight on the data. The reflectance in each band is recorded as an array and stored in the database. For example, for a single pixel, the reflectance values ​​recorded in the 16 bands are as follows:

[0068] [0.32, 0.35, 0.41, 0.47, 0.53, 0.57, 0.60, 0.61, 0.59, 0.56, 0.52, 0.48, 0.44, 0.40, 0.36, 0.33], where reflectance is a unit normalized value (between 0 and 1). The set of reflectance values ​​collected for each band is statistically analyzed. For example, if 2000 pixel response values ​​are collected for each band, a frequency distribution histogram is constructed for each band, and its statistical characteristics, such as mean, standard deviation, and skewness, are extracted to build a complete response database. When aggregating the values, it is necessary to distinguish between different response bands, such as by calculating λ for any band. i With λ j Rate of difference between lower reflectance values:

[0069]

[0070] If the difference rate exceeds a set threshold (e.g., 0.15), the band is considered to have significant differences and is included in the difference band set. For example, for a certain vegetation sample, the reflectances at 450nm and 680nm are 0.35 and 0.60, respectively. The difference rate is calculated as follows:

[0071]

[0072] Since this value is greater than the threshold of 0.15, both are grouped into the difference band set. A traversal operation is performed on the pixels to statistically analyze the distribution intervals under the difference bands. The sampled value of each pixel under the difference band set is then constructed as a multi-dimensional vector sample for storage. For example, a dataset can be constructed as follows:

[0073] [R 450 R 680 R 720 = [0.35, 0.60, 0.52];

[0074] Repeating this operation yields multiple sets of sample vectors, which are then used to build a band response database, including fields such as sample ID, location coordinates, multi-band reflectance values, and difference markers, as shown in Table 1 below:

[0075] Table 1: Band Response Database Structure Representation

[0076]

[0077] As shown in Table 1, each sample records multi-band reflectance data under spatial positioning. The difference threshold is used to determine whether the band combination has classification value, and a complete band response database is constructed. This result can be used as the input basis for subsequent filtering and differential analysis.

[0078] S102: Based on the band floor response value, a moving average filtering algorithm is used to set a sliding window for the floor response values ​​of adjacent bands, and a weighted average of adjacent response data is performed using convolution processing. The convolution result is matched with the original band order to generate the filtered average response value.

[0079] Based on the band-based response values, the size of the sliding window must first be set. If the current sensor includes 16 bands, the window width can be set to 3 bands. Then, the first moving average calculation is performed on bands 1 to 3, the second moving average on bands 2 to 4, and so on, until the last window covers bands 14 to 16. The band response values ​​within the window are calculated using a weighted convolution method. The weights are symmetrically set according to the distance of each band from the center band within the window. For example, the center weight is 0.5, and the weights of the left and right neighboring bands are 0.25. For band positions λ2, λ3, and λ4 with corresponding response values ​​R2 = 0.35, R3 = 0.41, and R4 = 0.47, the weighting operation R can be performed. avg =0.25×0.35+0.5×0.41+0.25×0.47=0.0875+0.205+0.1175=0.41. From this weighted result, the average filtered response value for band 3 is 0.41. When processing bands using a sliding window as described above, attention should be paid to boundary handling methods. For example, the first and last bands can be filled using edge replication or weight normalization strategies. After the sliding window completes the weighting operation, the filtered response value sequence corresponding to each band position can be obtained. For example, the original sequence is:

[0080] [0.32, 0.35, 0.41, 0.47, 0.53, 0.57, 0.60, 0.61, 0.59, 0.56, 0.52, 0.48, 0.44, 0.40, 0.36, 0.33];

[0081] After performing sliding window weighting, a filtered response sequence can be generated, such as:

[0082] [0.335, 0.3725, 0.4225, 0.495, 0.555, 0.595, 0.605, 0.603, 0.58625, 0.5525, 0.503, 0.48, 0.44, 0.405, 0.3675, 0.345];

[0083] The weight settings required for the weighting operation should be set according to the symmetric distribution principle. When the window width is odd, the highest weight is set at the center position, and the weights decrease symmetrically on both sides. The sum of the weights needs to be normalized to 1. For different window widths, the corresponding coefficient values ​​need to be adjusted. For example, when the window width is 5, the corresponding weights are [0.1, 0.2, 0.4, 0.2, 0.1]. If normalization is not performed, the weighted values ​​need to be manually divided by the sum of the weights. For each convolution weighting operation, the product of the three response values ​​and the weight array is actually calculated and then summed. This operation is equivalent to the finite response form of a linear filter, ensuring that the filtering result does not introduce non-physical fluctuations. For example, a point where a response value changes abruptly can be smoothed by the local surrounding values ​​through the moving average. The final generated filtered response sequence will be used as the data input for subsequent difference analysis.

[0084] S103: Based on the average response value of the filter, the difference calculation method is used to perform numerical difference on the average response value of each pair of adjacent bands, summarize the difference results between the bands, statistically analyze the range of difference variation, and obtain the distribution of response difference between adjacent bands.

[0085] Based on the average filtered response value, the response values ​​between adjacent bands need to be differentially analyzed item by item. The calculation steps include iterating through the response difference D between the i-th and (i+1)-th items in the sequence. i =R i+1 -R iFor example, if the current response sequence is [0.335, 0.3725, 0.4225, 0.495, 0.555], then the first pair of differences is 0.3725 - 0.335 = 0.0375, the second pair of differences is 0.4225 - 0.3725 = 0.05, and so on, generating a difference sequence for adjacent bands as [0.0375, 0.05, 0.0725, 0.06]. This difference sequence represents... This refers to the monotonic variation of the response value with the band. For difference sequences, further statistical analysis of their distribution characteristics is needed, including the maximum difference, minimum difference, mean difference, and standard deviation. For example, in a sample, the maximum value of the difference sequence is 0.0725, the mean is 0.055, and the standard deviation is 0.0157. When statistically analyzing the range of difference variation, a significant variation interval needs to be set. If the absolute value of the difference is greater than a certain threshold, for example, setting 0.06 as the threshold, then it is marked as significant. The band pairs are considered to be significantly different. For each sample, the above judgment is performed, and it is counted whether there is a change exceeding the threshold at the difference position. If the difference values ​​of band pairs (λ3, λ4) and (λ4, λ5) in a sample are 0.0725 and 0.06 respectively, then it is marked as containing two significantly different pairs. Further, the frequency of significantly different band pairs in the sample can be counted, and high-frequency difference positions can be analyzed in detail. For example, if the frequency of significant difference in band pair (λ3, λ4) in the sample exceeds 70%, it can be regarded as one of the basic bands for differentiation. In the process of summarizing the difference distribution, a difference matrix needs to be constructed. Each sample is constructed as a 15-dimensional vector representation (if the number of bands is 16, then the adjacent difference terms are 15 terms) for the feature input of the subsequent classification model. This difference distribution reflects the response change trend of the sample in the band space, providing a basis for subsequent spatial structure extraction and classification rule construction.

[0086] Specifically, the steps of S2 are as follows:

[0087] S201: Obtain the distribution data of response differences between adjacent bands, set normalization parameters for normalization processing, select the cross-correlation analysis method to set the operation window and rules, input the normalized pixels point by point into the analysis algorithm, calculate the correlation coefficient according to the pixel sliding window, record the result of each point, and form a cross-correlation coefficient sequence.

[0088] Normalization refers to linearly scaling response difference distribution data according to preset normalization parameters to stabilize its values ​​within a specified range. The initial values ​​of the parameters are set based on the original sample statistical mean.

[0089] When acquiring the distribution data of response differences between adjacent bands, it is necessary to first load the band difference sequence of each pixel in the sample. This difference sequence is an array of response value changes after sliding weighting and point-by-point differencing in the previous steps. The data structure is in one-dimensional vector form. For example, the difference value of a certain pixel is:

[0090] [0.015, 0.032, 0.047, 0.061, 0.050, 0.036, 0.024, 0.011, -0.004, -0.019, -0.031, -0.044, -0.053, -0.061, -0.070];

[0091] Next, normalization parameters are set for linear scaling. The parameters used for normalization must be obtained from the population statistics of the sample, specifically the mean and range of the current difference sequence. If the normalization range is set to [0, 1], then the normalization formula is:

[0092]

[0093] Where R i R represents the original difference value. min With R max These are the minimum and maximum pixel difference values ​​for the sample or the entire sample set, respectively. For example, if the maximum value of a current pixel difference sequence is 0.070 and the minimum value is -0.070, then the normalization parameter is R. max =0.070, R min = -0.070, substituting the value of the fourth term R4 = 0.061, the normalized result is:

[0094]

[0095] Normalized data as follows:

[0096] [0.607, 0.728, 0.839, 0.9357, 0.857, 0.757, 0.671, 0.589, 0.471, 0.364, 0.278, 0.186, 0.121, 0.064, 0.0]. After normalization, the values ​​are input point by point into the analysis algorithm for cross-correlation analysis. The analysis requires setting a calculation window, the size of which corresponds to the comparison range of adjacent sequences. Assuming a window width of 5, each analysis selects 5 consecutive difference values. For example, the first window is [0.607, 0.728, 0.839, 0.9357, 0.857]. Within this window, a sliding correlation calculation is performed with the standard reference vector. The standard reference vector can be set as the ideal gradient ascent vector [0.2, 0.4, 0.6, 0.8, 1.0]. The correlation coefficient is calculated based on the Pearson correlation coefficient expression:

[0097]

[0098] Where x i For normalized difference window data, y i As the reference vector, and These are the means of both, substituted into the first window data and the reference vector value mentioned above:

[0099]

[0100]

[0101] Calculate the numerator:

[0102]

[0103] Calculate the denominator:

[0104] The correlation coefficient was obtained:

[0105] This value is recorded as the cross-correlation coefficient of the pixel at the starting point of the current window. The window is slid sequentially, and the above calculation process is repeated. Finally, a sequence of cross-correlation coefficient values ​​corresponding to the starting point of each window is obtained, forming a complete cross-correlation coefficient sequence [0.3672, 0.4421, 0.5089, 0.5312, 0.4784], which provides a numerical basis for subsequent coordinate peak location and offset extraction.

[0106] S202: Based on the cross-correlation coefficient sequence, identify the trend of pixel correlation change, record the original coordinates and the coordinates corresponding to the peak, calculate the coordinate offset, organize the offset values ​​between pixels, and generate a spatial offset vector set;

[0107] Based on the cross-correlation coefficient sequence formed in the previous stage, it is necessary to sequentially extract the correlation value sequence generated by each pixel in its neighborhood as the sliding window changes. For a single pixel, the center position of the sliding window is fixed, and the correlation values ​​are obtained by sliding horizontally and vertically, respectively, as shown in r. x = [0.921, 0.934, 1.000, 0.975] and r γ For a sequence with correlations of [0.948, 0.982, 0.971], the following steps are taken: First, an extreme value detection operation is performed, and then the maximum value index is extracted from the sequence, for example, r. x The maximum value is 1.000, corresponding to the 3rd element, with a horizontal offset index of +1 step (i.e., corresponding to a center offset of 1 column). γThe maximum value is 0.982, corresponding to the second element. The vertical offset index is 0 (i.e., keeping the current row). The original coordinates of the cell (x=2, y=2) and the coordinates of its peak (x=3, y=2) are recorded respectively. Then the offset is calculated as (Δx=+1, Δy=0). The same logic is performed on all cells to form a two-dimensional vector difference dataset. Each vector represents the spatial offset direction and magnitude from the current cell to its maximum correlation position. In order to further avoid the interference of outliers, outlier removal operation needs to be performed on the generated offset set. That is, a full-map mean and standard deviation model is established for each offset dimension. If the Δx or Δy of a certain vector deviates from the full-map mean by more than 2 times the standard deviation, it is determined to be outlier data and removed. The main trend data is retained as valid input. For example, in the entire image, the mean offset in the Δx direction is 0.3 pixels, and the standard deviation is 0.6 pixels. If a vector Δx = 2.1, its Z value is (2.1-0.3) / 0.6 ≈ 3.0, which exceeds 2 and is considered an outlier. Then, the remaining valid offset data are organized into a spatial offset vector set V = {(x1, y1, Δx1, Δy1), (x2, y2, Δx2, Δy2), ...}. Each record in the dataset represents the corresponding offset between the original pixel position and its most correlated point, serving as the input basis for the subsequent displacement compensation module. To illustrate this data structure, a list of partial sample data is constructed below:

[0108] Table 2: Example Table of Spatial Offset Vector Set Representation

[0109] Original x-coordinate Original ordinate lateral offset Vertical offset 1 1 +1 0 1 2 +1 -1 1 3 0 -1 2 1 0 0 2 2 +1 0 2 3 +1 +1 3 1 0 +1 3 2 0 0 3 3 -1 +1

[0110] Table 2 lists some data points in the spatial offset vector set, with the offset step size represented by integer pixels. For cases where multiple values ​​are equal in the "peak corresponding coordinates" detection, the direction with the smallest corresponding center distance is prioritized; that is, when equal values ​​exist, the nearest neighbor is selected. The offset is calculated directly by taking the difference between the target point and the center point coordinates, avoiding the need to introduce a direction judgment model. This stage does not require an algorithm model; the process can be completed simply by using position indexing and difference calculation. Therefore, each step can be completed using array indexing and position subtraction, greatly simplifying the overall processing.

[0111] S203: Call the spatial offset vector set, retrieve the corresponding offset vector amplitude for each sensor array channel one by one, adjust the position parameters of the channel pixels in combination with the offset vector amplitude, summarize the adjusted channel pixel compensation data, and establish a displacement compensation parameter set.

[0112] After calling the spatial offset vector set V, the spatial position of each cell in the sensor array needs to be processed channel by channel. Assuming the current sensor array has three channels C1, C2, and C3, and each channel contains m×n cells, the processing flow is as follows: Taking channel C1 as an example, read the original coordinates (x, y) of each cell, calculate the target coordinates (x+Δx, y+Δy) based on the offset vector (Δx, Δy), and use these as the adjusted position parameters. If the coordinates of a point are (2, 2) and its offset vector is (+1, 0), then the updated position is (3, 2), denoted as C1(2, 2) → C1′(3, 2). For points without offset records, the original coordinates remain unchanged. After processing the cells, establish the mapping relationship between the updated position and the original cell value for each channel, forming a set of cells with updated positions, constituting the compensated datasets C1′, C2′, and C3′, which have the same spatial structure but different coordinate positions.

[0113] To ensure spatial continuity after pixel compensation, if multiple pixels converge at a target coordinate location (i.e., multiple source points are offset to the same target location), the average superposition of values ​​is performed. For example, if multiple pixels point to coordinates (3, 2) in channel C1, the final pixel value at that location is the average value of the pixels. Conversely, if no pixels are mapped to a certain location, a gap compensation operation is performed, for example, the average value of its four neighboring pixels is used as the fill value.

[0114] After processing, a channel compensation parameter set P = {(x′, y′, v′)} can be formed, where v′ is the compensated pixel value and (x′, y′) are the updated coordinates. For ease of understanding, some channel-level compensation results are listed below in text format:

[0115] In channel C1, the compensation value for pixel (2,2) after being shifted to (3,2) is 0.485, and the value for pixel (3,2) after being shifted to (4,2) is 0.472.

[0116] In channel C2, the value of cell (1,1) after being shifted to (2,1) is 0.498, and the value of cell (2,2) after being shifted to (3,2) is 0.501;

[0117] In channel C3, the value of cell (2,2) after being shifted to (3,3) is 0.476, and the value of cell (3,2) after being shifted to (4,3) is 0.482.

[0118] The compensation values ​​mentioned above are all obtained through simulation calculations. In the final spatial compensation output dataset, the position and value of each channel after pixel offset are recorded individually, serving as the input basis for subsequent calibration images or multi-channel data fusion. This compensation process does not introduce complex models and is completed entirely through vector offset and mapping relocation, exhibiting good stability and repeatability, and is suitable for large-scale pixel compensation processing tasks.

[0119] Specifically, the steps of S3 are as follows:

[0120] S301: Call the compensated three-dimensional coordinates in the displacement compensation parameter set, set the sampling path of the multispectral sensor, and collect the characteristic signal intensity at the preset spectral marker points in sequence according to the spatial coordinates. For each spatial coordinate point, the measured raw signal is spatially grouped, the signal distribution data of the marker points is statistically analyzed, and the spatial signal distribution value is generated.

[0121] When calling the compensated 3D coordinates from the displacement compensation parameter set, the compensated coordinates (x′, y′, z′) of each 3D point should be read pixel by pixel and loaded into the path planning module in the form of a coordinate sequence. For example, if a certain compensated coordinate set is [(12.3, 8.5, 3.0), (13.3, 8.5, 3.0), (14.3, 9.0, 3.1)], each point is extracted sequentially as the sampling starting point. Then, the sampling path of the multispectral sensor is set. The path setting process needs to be sorted according to the coordinate order and adjacent points are connected. For example, first sort them in ascending order according to the x′ coordinates, then calculate the distance between adjacent points and generate path segments. The path segment is defined as the straight line trajectory vector between points. If the angle between path segments is greater than 45 degrees, it is recorded as a path change point. For example, the angle between (12.3, 8.5, 3.0) and (14.3, 9.0, 3.1) is about 53 degrees, and its turning attribute is recorded. After the path is constructed, the sensors are sequentially moved to the spatial coordinate points. The positioning module of the control platform completes the positioning verification. After the sensors stabilize, the spectral signal acquisition operation is performed. During the sampling process, the sampling time is set to 0.5 seconds and the sampling interval is 1 second. N sets of data are collected for each spatial coordinate. Let N=3, then 3 spectral signal values ​​are collected at each point and recorded as [I1, I2, I3]. The average value of these three sets of data is calculated as the final signal value I at that point. If the difference between any sampling value exceeds ±10% of the average value, it is judged as abnormal, discarded, and resampled to ensure numerical stability. For example, at (13.3, 8.5, 3). 0) The location data is [0.496, 0.493, 0.450], with a mean of 0.4797 and a maximum deviation of 0.0297. The deviation percentage is 6.2%, which is less than 10% and is considered a valid value. Next, each measured spatial signal value is spatially grouped. The spatial grouping is divided into regional grids according to the (x′, y′) coordinate axis. Each regional grid is set to a size of 5 pixels × 5 pixels. Points in the same region are grouped together. For example, (13.3, 8.5) and (14.1, 9.2) belong to the same grid region and are uniformly identified as the nth group. The signal values ​​in each group are numbered to form a sequence [I]. 1，n , I 2，n , ..., I m，nLet m be the number of points in the group, and then calculate the mean I of the signal within the group. n With standard deviation σ n The validity is judged based on its value. If the difference between the signal strength at a certain point and the group mean exceeds 2σ... n Then the signal at that point is removed, such as I. n =0.480, σ n =0.015, and the signal at a certain point is 0.510, then its Z value is (0.510-0.480) / 0.015=2.0, and the critical point is retained. If the value at a certain point is 0.545, then the Z value is 4.3, and it is determined to be an outlier and removed. After obtaining the stable point set, the spatial signal distribution value of each group is statistically analyzed, and the point position and its effective signal value are recorded to form the final distribution mapping dataset. Some specific data are shown in Table 3.

[0122] Table 3: Sample Table of Spatial Signal Distribution Values

[0123]

[0124] Table 3 lists the coordinates and signal strength data of some sampling points, as well as their spatial group numbers, for subsequent distribution calculation and fitting.

[0125] The location information of each measurement point and its signal strength value are recorded by using spatial coordinates (x′, y′, z′). The group number n is divided according to the spatial location. During the execution, the side length of each group grid can be flexibly set according to the actual region of interest. For example, 2 pixels × 2 pixels can be used for high-precision scenes, and 10 pixels × 10 pixels can be used for fast filtering scenes. In the table, n = 1 represents the signal area of ​​the first group.

[0126] In summary, each step in the execution process uses three-dimensional compensation coordinates as the basic data source, controls the sensor to collect data coordinate by coordinate, and forms a stable and reliable spatial signal distribution value dataset based on the aggregation and anomaly removal actions of signal data within the grouped execution area.

[0127] S302: Call the spatial signal distribution value, integrate the signal intensity of the two-dimensional matrix according to the spatial neighborhood for each group of signal distribution areas, calculate the weighted mean of the spatial neighborhood, fit the two-dimensional signal intensity distribution by the least squares method, and use the two-dimensional Gaussian function to extract the distribution feature parameters of the region to obtain the Gaussian parameter coefficients.

[0128] After calling the spatial signal distribution values, an initial signal strength matrix needs to be formed for the two-dimensional coordinate points within each signal distribution area. During matrix construction, the group n must first be determined. For example, suppose there is a spatial signal area in group 1 containing M = 5 pixels located at coordinates (2, 3), (3, 3), (3, 4), (2, 4), and (4, 3), with corresponding signal strengths of 0.521, 0.544, 0.533, 0.527, and 0.548, respectively. Calculate the mean signal strength I within the group. n It can be obtained by summing and averaging, i.e., I n = (0.521+0.544+0.533+0.527+0.548) / 5 = 0.5346. Based on this, a weighted mean is calculated for each pixel within its spatial neighborhood. The specific process is as follows: the absolute difference between the signal strength of the current pixel and the group mean |I| is used. i,n -I n | is the basis for signal offset weighting, multiplied by its coordinates and (x) i,n +y i,n This process yields the weighted signal value for each point. Taking point (2, 3) as an example, its offset weight is |0.521 - 0.5346| = 0.0136, and the sum of its coordinates is 5, therefore the weighted signal is 0.068. The same logic applies to the other points. After completion, the sum of the weighted signal values ​​is calculated as the numerator.

[0129] Simultaneously, it is necessary to calculate the coordinate variances of the pixels within this group in the x-axis and y-axis directions, denoted as follows: and For example, taking the x-axis coordinates [2, 3, 3, 2, 4], the mean is 2.8 and the variance is... Similarly, the y-axis coordinates are [3, 3, 4, 4, 3], with a mean of 3.4 and a variance of... The weighted mean of spatial neighborhoods is calculated using the following formula:

[0130]

[0131] Where, μ n I represents the weighted average signal intensity of the nth spatial signal distribution region. i,n I represents the signal intensity of the i-th pixel within the n-th spatial signal distribution area. n x represents the average signal intensity of pixels within the nth spatial signal distribution region. i,n With y i,n These represent the x-coordinates and y-coordinates of the i-th pixel within the n-th spatial signal distribution area, respectively. represents the x-axis variance of the pixel points within the distribution region of the nth group of spatial signals. represents the y-axis variance of the pixel points within the nth group of spatial signal distribution areas, M represents the total number of pixel points within the nth group of spatial signal distribution areas, and n represents the group of the signal distribution areas.

[0132]

[0133] The weighted mean of the neighborhood was 0.032. Subsequently, based on the original two-dimensional signal intensity distribution and the signal attenuation characteristics between adjacent points, a least squares method was used to construct a fitting objective within this region. This objective was to minimize the sum of squared errors between the actual signal intensity and the fitted function. The two-dimensional coordinates and their signal values ​​within the region were then fitted to a two-dimensional Gaussian distribution function.

[0134]

[0135] where μ x μ y With the center coordinates, σ x σ y Let A be the standard deviation and A be the peak intensity. The fitting process involves substituting each pixel into a (x, y, I) triplet, calculating the sum of squared residuals, and iteratively approximating the minimum value. Finally, the Gaussian distribution parameter coefficients of the signal intensity in this region are extracted as (A = 0.546, μ). x =2.95, μ y =3.4, σ x =0.63, σ y =0.49).

[0136] The above coefficient extraction results are based on the actual pixel coordinate distribution and sampling signal intensity in the region. They are completed through a clear calculation process to form a set of regional feature parameters, which provides a basis for subsequent Gaussian fitting center position determination and spatial matching.

[0137] S303: Based on the Gaussian parameter coefficients, determine the signal center position of the marker point in two-dimensional space, match the center coordinates obtained from the positioning with the compensated three-dimensional coordinates, calculate their corresponding Euclidean distance, filter out the three-dimensional points with small Euclidean distances, and obtain the registered marker point coordinate dataset.

[0138] Based on the Gaussian parameter coefficients obtained in the previous step, firstly, extract the coordinates (x0, y0) of the center point of the two-dimensional Gaussian function from multiple sets of data. Let the center point in the first set be (13.4, 8.7). Then, call the stored compensated three-dimensional coordinate set to extract the corresponding (x′, y′, z′) data, and calculate the Euclidean distance between the two-dimensional center point and the three-dimensional point projected onto the xy plane one by one. The formula is:

[0139]

[0140] Assuming the candidate 3D points are (13.3, 8.5, 3.0), (14.1, 9.2, 3.1), and (13.1, 8.9, 3.0), the calculation is as follows:

[0141] Point 1:

[0142] Point 2:

[0143] Point 3:

[0144] Sort the points by distance, filter the N points with the smallest distance, set the threshold to 0.4mm, and keep the points with an Euclidean distance of less than 0.4mm, namely point 1 and point 3; the resulting registered marker point coordinate dataset is [(13.3, 8.5, 3.0), (13.1, 8.9, 3.0)].

[0145] The results show that an effective spatial registration model can be constructed based on the geometric relationship between the Gaussian center and the compensated 3D coordinates, and the obtained 3D points can be further used as signal anchor points or reference markers in a spatial positioning system.

[0146] Specifically, the steps of S4 are as follows:

[0147] S401: Based on the registered marker point coordinate dataset, all marker points are filtered in sequence. The marker points are grouped and corresponding triangular units are constructed using a spatial triangular mesh partitioning method. Three sides are extracted for each group of triangular units. The actual side length is calculated based on the coordinate data of the marker points at both ends of each side, and the side length value of the triangular unit is generated.

[0148] Based on the registered marker point coordinate dataset [(13.3, 8.5, 3.0), (13.1, 8.9, 3.0)], the marker points in the entire dataset are first sequentially selected, with two different points chosen each time as boundary vertices. The 3D coordinates (x, y, z) of each point are decomposed, and the x, y, and z components of each point are extracted. Then, according to the spatial triangular mesh division method, every three adjacent points are combined into a triangular unit. If the number of points is less than three, no triangular unit is formed. In this example, since there are only two points, it is impossible to complete the construction of triangular units. It is assumed that an additional point P3 (13.5, 8.7, 3.2) can be added to form a triangular mesh unit. During the execution, any three points P1 (13.3, 8.5, 3.0), P2 (13.1, 8.9, 3.0), and P3 (13.5, 8.7, 3.2) are combined into a triangular unit, and three edges are extracted respectively. The Euclidean distance between the two endpoints of each edge is calculated. The calculation formula is as follows:

[0149]

[0150] Where (x1, y1, z1) and (x2, y2, z2) are the coordinates of the two ends of each edge, respectively. Substitute them one by one:

[0151] Calculate the side lengths of P2, P3, P1, and P3 sequentially to obtain the actual lengths of the three sides of each triangular unit. Then, record the three side length values ​​for each triangular unit to generate a set of side length values ​​for that unit. During execution, the coordinates of point pairs must be obtained from actual data collection within the dataset; virtual point sets cannot be used. If the number of actual collected points is N, then triangular unit construction needs to be performed C(N, 3) times. Each side length calculation must be recorded in the side length value table. For example, the side lengths of the triangular units are as follows:

[0152] l 12 ≈0.447;

[0153]

[0154] Assign the three side lengths to the data table respectively, and establish the following set of side length values. After the final set of side length values ​​is generated, it will be used as the input parameter for subsequent steps. The calculation results of the side lengths must all reach three decimal places with precision. The side length value table is shown in the table below:

[0155] Table 4: Side Length Values ​​of Triangular Units

[0156] Triangular unit numbering Side 1 (unit: m) Side 2 (unit: m) Side 3 (unit: m) 1 0.447 0.490 0.346

[0157] As shown in Table 4, the lengths of the three sides of each triangular unit have been calculated step by step, and the obtained side length data serve as the basic input parameters for subsequent spatial offset calculations.

[0158] S402: Call the side length value of the triangular unit, extract the corresponding preset geometric layout standard side length for each triangular unit side, calculate the numerical offset one by one according to the actual side length and the standard side length, and classify and statistically analyze the offset data of the triangular unit to obtain the spatial offset of the triangular unit.

[0159] The algorithm retrieves the side length values ​​of the triangular units. For each triangular unit side, it extracts the standard side length value under the preset geometric layout. The actual side length of the triangular unit is then compared to the standard side length, and the difference is calculated one-to-one to obtain the numerical offset of multiple sides. In the specific execution process, the standard length of each side must first be determined. The standard length can be measured based on design drawings or experimental benchmarks. Assuming the standard lengths of the three sides are 0.450m, 0.500m, and 0.350m respectively, for side 1, comparing its actual side length of 0.447m with the standard 0.450m yields a difference of -0.003m. For side 2, the difference is the difference between 0.490m and 0.500m, resulting in -0.010m. For side 3, the difference is the difference between 0.346m and 0.350m, resulting in -0.004m. All numerical offsets must be recorded. The spatial offset value set requires the offset data of the edges to be organized into groups and statistically assigned to the corresponding triangular units. The statistical method is to directly accumulate the absolute values ​​of the offsets of the three edges of the same triangular unit. This can be used for subsequent classification and statistical analysis of the overall spatial offset. If the absolute value of the offset of a certain edge is less than 0.005m, it is classified as a low offset interval; if it is between 0.005m and 0.015m, it is classified as a medium offset interval; and if it is greater than 0.015m, it is classified as a high offset interval. Further, for the three-edge offset results in this example, the overall spatial offset of the triangular unit can be defined as the sum of the absolute values ​​of the three-edge offsets, i.e., |-0.003|+|-0.010|+|-0.004|=0.017m. This value is used as the spatial offset of the unit and recorded in the spatial offset table as the basic data for subsequent interpolation mapping.

[0160] S403: Based on the spatial offset of the triangular unit, the spatial offset data is mapped to a two-dimensional plane coordinate system using an interpolation method. The error distribution information of multiple points in the two-dimensional plane is allocated through the interpolation scheme to obtain a spatial registration error distribution map.

[0161] Based on the spatial offset of the triangular unit obtained in the previous paragraph, an interpolation method is used to map the spatial offset data to a two-dimensional plane coordinate system. During interpolation, the coordinates (x, y) of the projection points of the triangular unit vertices in three-dimensional space must first be extracted. Corresponding to the two-dimensional plane, for each triangular unit's spatial offset, such as 0.017m, it is allocated to the plane region where that unit is located. Using linear interpolation or spatially weighted interpolation, the error distribution values ​​of the mesh region are calculated sequentially in the plane using the triangular mesh vertex coordinates and offsets. During interpolation, the weights used can be inversely proportional to the distance from the vertex to the interpolation point. The weights can be set as w = 1 / d, where d is the Euclidean distance from the vertex to the interpolation point. In the actual case, assuming the distances from the interpolation point to P1, P2, and P3 are 0.2, 0.3, and 0.25m respectively, then the weights are 5.00, 3.33, and 4.00 respectively, with a total weight of 12.33. The weighted average of the three offsets is:

[0162]

[0163] The weighted distribution values ​​were calculated sequentially to be 0.00689, 0.00459, and 0.00552, and the sum of the three was 0.017m, representing the offset distribution of the interpolation points. After the interpolation points were assigned, the spatial registration error distribution data of multiple points in the two-dimensional plane were organized and plotted into a spatial registration error distribution map. The error values ​​of the regions on the error distribution map were all calibrated with an accuracy of 0.001m. The final generated error distribution information reflects the detailed distribution pattern of the entire spatial registration error in the two-dimensional plane.

[0164] Specifically, the steps of S5 are as follows:

[0165] S501: Obtain spatial registration error distribution map data. Based on the spatial coordinates of each group of pixels in the distribution map, use the spatial mapping rules of the bilinear interpolation algorithm to perform remapping calculations on the corresponding relationships of the coordinates and generate spatial coordinate remapping values.

[0166] Spatial coordinate remapping value refers to the remapped spatial coordinates of a pixel after spatial registration error compensation;

[0167] To obtain the spatial registration error distribution map data, firstly, extract the spatial coordinate points of each group of pixels in the error distribution map according to the pixel number. Each group of pixel coordinate points includes (x... i ,y i Two-dimensional numerical values ​​are generated using actual collected data. For each coordinate point, its corresponding pixel index in the error distribution map is found. For example, the spatial coordinates of pixel 1 are (20.5, 45.2), and the spatial coordinates of pixel 2 are (21.0, 45.5). These are recorded one by one. For each coordinate point, according to the spatial mapping rules of bilinear interpolation, the four neighboring grid points of that point in the error distribution map need to be found, denoted as the top left, top right, bottom left, and bottom right points. Assuming the coordinates of the four neighboring points are (20, 45), (21, 45), (20, 46), and (21, 46), they are denoted as A, B, C, and D respectively. Retrieve the error values ​​corresponding to A, B, C, and D from the registration error distribution map, assuming they are 0.015m, 0.017m, 0.013m, and 0.016m respectively. For the pixel coordinates (20.5, 45.2), calculate its corresponding weights among points A, B, C, and D. Using the bilinear interpolation weight allocation method, the horizontal difference is (20.5-20) = 0.5, and the vertical difference is (45.2-45) = 0.2. The horizontal weights are 0.5 and 0.5 respectively, and the vertical weights are 0.8 and 0.2 respectively. The weight allocation for the error values ​​A, B, C, and D can be written as follows:

[0168] E=A×(1-α)(1-β)+B×α(1-β)+C×(1-α)β+D×αβ;

[0169] Where α = 0.5, β = 0.2, substitute the specific values:

[0170] E=0.015×(1-0.5)(1-0.2)+0.017×0.5×(1-0.2)+0.013×(1-0.5)×0.2+0.016×0.5×0.2=0.006+0.0068+0.0013+0.0016=0.0157;

[0171] The final error compensation amount for the remapped coordinates is 0.0157m. Subsequently, the corresponding compensation value is obtained for the spatial coordinates of each pixel through the above process. The original coordinate points are then mapped and adjusted accordingly. The mapping adjustment method is to compensate the original coordinate points along the spatial error distribution direction by a corresponding distance. After completion, the spatial coordinate remapped value is generated. For pixel 1, the remapped coordinates after compensation of the original spatial coordinates (20.5, 45.2) are (20.5+Δx, 45.2+Δy), where Δx and Δy can be decomposed into compensation amounts according to the principal error direction. After pixel compensation, a new set of spatial coordinates is generated as the spatial coordinate remapped value. Finally, the remapped spatial coordinates corresponding to the pixel after spatial registration error compensation are obtained.

[0172] S502: Based on the spatial coordinate remapping value, combined with the original acquisition time series parameters of the multispectral sensor array, for each set of time series parameters and pixel coordinates, the error distribution and the pixels under the time series acquisition parameters are mapped to new spatial positions through the bilinear interpolation algorithm to obtain the multispectral time series spatial mapping value.

[0173] Multispectral temporal spatial mapping value refers to the new spatial location to which a pixel is mapped after combining temporal parameters and channel information;

[0174] Based on the spatial coordinate remapping values, the original acquisition timing parameters of each channel of the multispectral sensor array are first organized. These timing parameters typically include exposure time, frame number, and spectral channel number. Each set of timing parameters is matched with the pixel coordinates using a correspondence table. For example, the original acquisition coordinates of the k-th pixel in the n-th frame of a certain channel are (20.5, 45.2), with timing parameters of frame number 12 and channel 3. Under each timing parameter, for its pixel coordinates, the corresponding spatial coordinate remapping value is found. The remapping coordinates are compared with the timing parameters, and bilinear interpolation is used to calculate the new mapped coordinates for each pixel's position under the timing acquisition parameters, based on its surrounding neighborhood spatial coordinates and error distribution weights. Specifically, the four neighboring points of the remapping coordinates are first found, and the horizontal and vertical weights are calculated. Then, according to the weight allocation method, the interpolation of the remapping coordinates in the two-dimensional plane is calculated. In the example, for frame number 12, channel 3, and pixel (20.5, 45.2), the remapping error is 0.0157m. Therefore, the new coordinates are (20.5 + Δx, 45.2 + Δy). Δx and Δy can be set as the error distribution value multiplied by a specified direction coefficient, which takes values ​​between 0 and 1, allocated according to the actual angle. Assuming Δx = 0.008m and Δy = 0.0077m, the new coordinates are (20.508, 45.2077). The new coordinates after pixel interpolation compensation must be recorded completely. Combining the channel number and acquisition sequence, the new spatial positions of pixels in each channel and time sequence are mapped sequentially. The results are then compiled and summarized to generate multispectral temporal spatial mapping values. The data structure can use a triple (channel number, time sequence frame number, new spatial coordinates). Arranging the triples sequentially yields a complete multispectral temporal spatial mapping value dataset. Example data is shown in the table below:

[0175] Table 5: Multispectral Temporal Spatial Mapping Values

[0176] Channel number Frame number New spatial x-coordinate (unit: m) New spatial y-coordinate (unit: m) 3 12 20.508 45.2077 4 12 20.512 45.2135 3 13 20.510 45.2100

[0177] As shown in Table 5, the table records the correspondence between different channels, different frame numbers and the corresponding new spatial coordinates of pixels, and lists some data examples of multispectral temporal spatial mapping values.

[0178] S503: Based on the multispectral temporal spatial mapping values, the channels of the original data acquired by the multispectral sensor array are reordered in spatial order, the mapping sequence of the channel data is integrated, and a multispectral imaging dataset is established.

[0179] Based on the multispectral temporal spatial mapping data, the channel order of the original data acquired by the multispectral sensor array is spatially reordered. Specifically, each channel's pixels are traversed, and their corresponding new spatial coordinates are extracted. These new spatial coordinates are then sorted in ascending order, first by the x-coordinate, and then by the y-coordinate if the x-coordinates are the same. If the new spatial coordinates of two pixels are less than a spatial threshold of 0.005m, they are determined to be in the same spatial region, and the pixels are grouped into the same spatial group. Then, the channels are numbered sequentially according to the new spatial coordinate sequence. The channel numbering results are organized into a new channel sorting table. For multispectral imaging data, the new spatial mapping position of the pixels in each channel and each temporal frame is extracted, and the corresponding pixel data is rearranged according to the new sorting table. The channel mapping sequences are integrated according to the new sorting table, ultimately forming a new multispectral imaging dataset. The order of this dataset strictly follows the spatial position sorting and channel combination results. The original data of each channel is rearranged according to the new spatial mapping relationship and then stored, thus completing the multispectral imaging dataset.

[0180] like Figure 2 As shown, a pattern recognition image acquisition device integrating multispectral imaging is disclosed. The device includes:

[0181] The floor response filtering module is used to acquire multi-band floor response values ​​of the target area through a multispectral sensor, and to perform convolution operation on the floor response values ​​of adjacent bands using a moving average filtering algorithm to generate the differential distribution of responses of adjacent bands, which is then transmitted to the spatial offset compensation module.

[0182] The spatial offset compensation module is used to call the response difference distribution of adjacent bands, input the difference distribution into the cross-correlation analysis algorithm to extract the spatial offset vector between adjacent bands, perform displacement compensation on the three-dimensional coordinates of the sensor array based on the offset vector amplitude, generate a set of displacement compensation parameters, and pass them to the marker registration module.

[0183] The marker registration module is used to call the compensated three-dimensional coordinates in the displacement compensation parameter set, capture the characteristic signal intensity of the preset spectral markers in a directional manner, perform two-dimensional Gaussian fitting on the regional signal using the least squares method, generate a registered marker coordinate dataset, and pass it to the error mapping module.

[0184] The error mapping module is used to divide the registered marker point coordinate dataset into spatial triangular meshes, calculate the offset between the side length of the triangular unit and the preset geometric layout, map it to the two-dimensional plane coordinate system by interpolation, generate a spatial registration error distribution map, and pass it to the temporal remapping module.

[0185] The temporal remapping module is used to call the spatial registration error distribution map, input the acquisition time-series parameters of the multispectral sensor array into the bilinear interpolation algorithm, perform spatial domain pixel coordinate transformation on the time-series parameters according to the error distribution, and output the multispectral imaging dataset.

[0186] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A pattern recognition image acquisition method integrating multispectral imaging, characterized in that, The following steps are involved: S1: Obtain the base response values ​​of the target area under multiple bands through a multispectral sensor, and use a moving average filtering algorithm to perform convolution operation on the base response values ​​of adjacent bands to generate the response difference distribution of adjacent bands; S2: Input the response difference distribution of adjacent bands into the cross-correlation analysis algorithm, extract the spatial offset vector between adjacent bands, perform displacement compensation on the sensor array based on the offset vector amplitude, and generate a set of displacement compensation parameters. S3: Call the compensated three-dimensional coordinates in the displacement compensation parameter set, capture the characteristic signal intensity of the preset spectral marker points through a multispectral sensor, perform two-dimensional Gaussian fitting using the least squares method for the regional signal, and generate a registered marker point coordinate dataset. S4: Divide the registered marker point coordinate dataset into spatial triangular meshes, calculate the offset between the side length of the triangular unit and the preset geometric layout, map it to a two-dimensional plane coordinate system using interpolation, and generate a spatial registration error distribution map. S5: Input the spatial registration error distribution map into the bilinear interpolation algorithm to perform spatial domain remapping calculation on the acquisition time sequence parameters of the multispectral sensor array, and output the multispectral imaging dataset.

2. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 1, characterized in that, The adjacent band response difference distribution includes distribution characteristics, statistical amplitude, and spectral difference index; the displacement compensation parameter set includes compensation amplitude, compensation direction, and compensation accuracy; the registered marker point coordinate dataset includes marker point spatial coordinates, attribute labels, and signal strength information; the spatial registration error distribution map includes error value distribution, spatial location mapping, and error extreme value distribution; and the multispectral imaging dataset includes remapping acquisition parameters, acquisition timing information, and registered multispectral data.

3. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 1, characterized in that, The specific steps of S1 include: S101: Acquire reflection signals of the target area under multiple bands through a multispectral sensor, collect multi-band floor response data, statistically analyze the numerical distribution of the floor response based on the response signals of each band, establish a complete band response database by collecting the sampled values ​​under different bands, and generate band floor response values. S102: Based on the base response value of the band, a moving average filtering algorithm is used to set a sliding window for the base response values ​​of adjacent bands, and a weighted average of adjacent response data is performed using a convolution processing method. The convolution result is matched with the original band order to generate a filtered average response value. S103: Based on the average response value of the filter, the difference calculation method is used to perform numerical difference on the average response values ​​of each pair of adjacent bands, summarize the difference results between the bands, and statistically analyze the range of difference variation to obtain the distribution of response differences between adjacent bands.

4. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 3, characterized in that, The sliding window width of the moving average filtering algorithm is automatically set based on the signal-to-noise ratio parameters of the multispectral sensor in different bands, and the convolution weights are modified by the parameters of the original acquired data. The band substrate response value is a statistical characteristic value of the target area reflection signal collected by the multispectral sensor in multiple bands, and is a statistical characteristic of the target area reflection signal. The filtered average response value is the result obtained by weighting and smoothing the base response values ​​of adjacent bands using a moving average filtering algorithm, and represents the statistical characteristics of the reflected signal in the target area.

5. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 3, characterized in that, The specific steps of S2 include: S201: Obtain the response difference distribution data of the adjacent bands, set the normalization parameters for normalization processing, select the cross-correlation analysis method to set the operation window and rules, input the normalized pixels point by point into the analysis algorithm, calculate the correlation coefficient according to the pixel sliding window, record the result of each point, and form a cross-correlation coefficient sequence. S202: Based on the cross-correlation coefficient sequence, identify the trend of pixel correlation change, record the original coordinates and the coordinates corresponding to the peak, calculate the coordinate offset, organize the offset values ​​between pixels, and generate a spatial offset vector set; S203: Call the spatial offset vector set, retrieve the corresponding offset vector amplitude for each sensor array channel one by one, adjust the position parameters of the channel pixels in combination with the offset vector amplitude, summarize the adjusted channel pixel compensation data, and establish a displacement compensation parameter set.

6. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 5, characterized in that, The specific steps of S3 include: S301: Call the compensated three-dimensional coordinates in the displacement compensation parameter set, set the sampling path of the multispectral sensor, and collect the characteristic signal intensity at the preset spectral marker points in sequence according to the spatial coordinates. For each spatial coordinate point, the measured raw signal is spatially grouped, the signal distribution data of the marker points is statistically analyzed, and a spatial signal distribution value is generated. S302: Call the spatial signal distribution value, integrate the two-dimensional matrix signal intensity according to the spatial neighborhood for each group of signal distribution areas, calculate the spatial neighborhood weighted mean, fit the two-dimensional signal intensity distribution by the least squares method, and use the two-dimensional Gaussian function to extract the distribution feature parameters of the region to obtain the Gaussian parameter coefficients. S303: Based on the Gaussian parameter coefficients, determine the signal center position of the marker point in the two-dimensional space, match the center coordinates obtained from the positioning with the compensated three-dimensional coordinates, calculate their corresponding Euclidean distance, filter out the three-dimensional points with small Euclidean distances, and obtain the registered marker point coordinate dataset.

7. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 6, characterized in that, The spatial signal distribution value is the signal intensity data collected by the multispectral sensor at the spatial coordinate point, reflecting the spectral signal characteristics of the detected object at the coordinate point; The Gaussian parameter coefficients are characteristic parameters of the center position, amplitude, and diffusion width obtained by fitting the signal distribution based on a two-dimensional Gaussian distribution function. The least squares method is a mathematical method that optimizes parameters by minimizing the sum of squared errors between the fitted model and the actual observed data. The initial values ​​of the parameters are set based on the statistical characteristics of the sampled data. The two-dimensional Gaussian function characterizes the spatial distribution of the signal, and its parameters include amplitude, center coordinates, and standard deviation. The initial value is determined based on the distribution center, extrema, and diffusion range of the sampled signal.

8. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 6, characterized in that, The specific steps of S4 include: S401: Based on the registered marker point coordinate dataset, all marker points are filtered in sequence, and the marker points are grouped and corresponding triangular units are constructed using a spatial triangular mesh partitioning method. Three sides are extracted for each group of triangular units, and the actual side length is calculated for the coordinate data of the marker points at both ends of each side to generate the side length value of the triangular unit. S402: Call the side length value of the triangular unit, extract the corresponding preset geometric layout standard side length for each triangular unit side, calculate the numerical offset one by one according to the actual side length and the standard side length, and classify and statistically analyze the offset data of the triangular unit to obtain the spatial offset of the triangular unit. S403: Based on the spatial offset of the triangular unit, the spatial offset data is mapped to a two-dimensional plane coordinate system using an interpolation method. The error distribution information of multiple points in the two-dimensional plane is allocated through the interpolation scheme to obtain a spatial registration error distribution map.

9. The pattern recognition image acquisition method based on fused multispectral imaging according to claim 8, characterized in that, The specific steps of S5 include: S501: Obtain the spatial registration error distribution map data, and according to the spatial coordinate points of each group of pixels in the distribution map, use the spatial mapping rules of the bilinear interpolation algorithm to perform a remapping calculation on the corresponding relationship of the coordinate points to generate spatial coordinate remapping values. S502: Based on the spatial coordinate remapping value, combined with the original acquisition time series parameters of the multispectral sensor array, for each set of time series parameters and pixel coordinates, the error distribution and the pixels under the time series acquisition parameters are mapped to new spatial positions through the bilinear interpolation algorithm to obtain the multispectral time series spatial mapping value. S503: Based on the multispectral temporal spatial mapping value, the channels of the original data acquired by the multispectral sensor array are reordered in spatial order, the mapping sequence of the channel data is integrated, and a multispectral imaging dataset is established.

10. A pattern recognition image acquisition device integrating multispectral imaging, characterized in that, The device is used to implement the pattern recognition image acquisition method for fused multispectral imaging as described in any one of claims 1-9, and the device comprises: The floor response filtering module is used to acquire multi-band floor response values ​​of the target area through a multispectral sensor, and to perform convolution operation on the floor response values ​​of adjacent bands using a moving average filtering algorithm to generate the differential distribution of responses of adjacent bands, which is then transmitted to the spatial offset compensation module. The spatial offset compensation module is used to call the response difference distribution of adjacent bands, input the difference distribution into the cross-correlation analysis algorithm to extract the spatial offset vector between adjacent bands, perform displacement compensation on the three-dimensional coordinates of the sensor array based on the offset vector amplitude, generate a set of displacement compensation parameters, and pass them to the marker registration module. The marker registration module is used to call the compensated three-dimensional coordinates in the displacement compensation parameter set, capture the characteristic signal intensity of the preset spectral markers in a directional manner, perform two-dimensional Gaussian fitting on the regional signal using the least squares method, generate a registered marker coordinate dataset, and pass it to the error mapping module. The error mapping module is used to divide the registered marker point coordinate dataset into spatial triangular meshes, calculate the offset between the side length of the triangular unit and the preset geometric layout, map it to a two-dimensional plane coordinate system by interpolation, generate a spatial registration error distribution map, and pass it to the temporal remapping module. The temporal remapping module is used to call the spatial registration error distribution map, input the acquisition timing parameters of the multispectral sensor array into the bilinear interpolation algorithm, perform spatial domain pixel coordinate transformation on the timing parameters according to the error distribution, and output the multispectral imaging dataset.

Images