Missile charging process control method and system based on machine vision

Abstract

The application discloses a missile charging process control method and system based on machine vision, and relates to the technical field of image data processing. The method collects a missile solid propellant matte homogeneous charging forming quality detection image, performs multi-scale decomposition on the detection image, establishes a corresponding relationship between a reference image block and each scale image block, extracts a gray level co-occurrence matrix texture feature and forms a texture feature vector; the texture feature vector is taken as a one-dimensional signal, a plurality of modal components are obtained through joint objective function decomposition containing a data fidelity term and an orthogonal constraint penalty term, and a background microtexture modal component and a candidate interface defect modal component are determined according to power spectrum entropy; a block-level defect probability of each reference image block is obtained by a defect discriminator, a pixel-level defect probability matrix is generated through covering mapping and weighted fusion, an interface defect area is determined by combining threshold judgment and connected domain screening, a forming state is determined according to a defect area proportion, and charging process parameters are adjusted.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to a missile propellant loading process control method and system based on machine vision. Background Technology

[0002] Missile solid propellant loading involves casting a high-solids-content slurry into a mold or engine casing and allowing it to solidify. The bonding quality of the propellant interface directly determines the structural integrity and combustion stability of the propellant grain. In multi-layer casting processes, an interlayer bonding interface is formed between the subsequent layer of slurry and the solidified or semi-solidified surface of the previous layer. The formation state of this interface is one of the core concerns in process control.

[0003] Propellant slurries typically have a matte, homogeneous appearance. Their solidified surfaces possess natural, fine textures formed by solid particle filling and matrix shrinkage. Under imaging conditions, this texture manifests as low-contrast, random grayscale fluctuations without clear direction. Ideally, the interlayer bonding regions formed by multi-layer casting should be consistent with the material bulk in density and structure, resulting in a gradual transition from one layer to another in the image, lacking clearly identifiable physical boundaries. The imaging conditions in confined environments further constrain the available light source and acquisition angle, limiting the overall contrast and resulting in a low signal-to-noise ratio in the acquired images.

[0004] Under the aforementioned imaging conditions, conventional machine vision methods for detecting the quality of charge forming typically follow the technical route of texture feature extraction and pattern classification. The general process includes preprocessing the acquired image, constructing a texture descriptor, and determining the target region based on a classification model in a predefined feature space. This type of method has been widely used in situations where there is a clear texture difference between the detected object and the background.

[0005] The limitations of existing technologies include at least the following problems: In the quality inspection of matte homogeneous solid propellant molding for missiles based on machine vision, the propellant grain itself has natural background micro-textures, and the interlayer bonding interface formed by multi-layer casting has a low-contrast gradual transition shape with no clearly discernible physical boundaries. In addition, the imaging signal-to-noise ratio is low in the closed operation scenario. The original texture of the material and the interface defects caused by interlayer debonding and micro-bubbles have similar gray-scale statistical characteristics and spatial scale distribution at the pixel level. The processing logic of existing machine vision for this type of image is based on the differentiation of texture morphology patterns. When the target defect and the material background have a high degree of overlap in the feature space, the classifier, relying on the existing feature definition, finds it difficult to calibrate a stable decision boundary between false alarms from the original texture of the material and missed detections of hidden defects in the gradual interface. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a missile propellant loading process control method and system based on machine vision. This method solves the problems in existing technologies where the background microtextures and interlayer defect image features of missile propellants highly overlap, making it difficult to define stable detection decision boundaries and easily leading to false alarms and missed defects.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a missile propellant loading process control method based on machine vision, comprising the following steps: acquiring images of the matte homogeneous solid propellant molding quality inspection of missiles; performing multi-scale decomposition on the inspection images, dividing the original inspection images into reference image blocks, and determining the multi-scale image block groups corresponding to each reference image block based on the scaling relationship between each scale decomposed image and the original inspection image; calculating the gray-level co-occurrence matrix of each scale image block, extracting specified texture feature parameters and combining them in scale order to generate texture feature vectors corresponding to each reference image block; using the texture feature vectors corresponding to each reference image block as one-dimensional signals to construct a system containing data fidelity terms and orthogonal... The joint objective function of the constraint penalty term obtains multiple modal components through the alternating direction multiplier method, and determines the background microtexture modal component and candidate interface defect modal component based on the power spectral entropy. The candidate interface defect modal components corresponding to each reference image block are input into the trained defect discriminator to obtain the block-level defect probability value of each reference image block. According to the coverage relationship of each reference image block, the block-level defect probability values ​​covering the same pixel position are weighted and fused to generate a pixel-level defect probability matrix, and the interface defect region is determined by threshold judgment and connected component screening. The forming state of the interlayer bonding interface is determined according to the area ratio of the interface defect region, and the charging process parameters of the subsequent casting layer or the next casting cycle are adjusted according to the judgment result.

[0008] Further, the specific steps for determining the multi-scale image block group corresponding to each reference image block and extracting the specified texture feature parameters are as follows: Perform multi-scale Gaussian pyramid decomposition on the detection image to obtain decomposed images at multiple scales; Use a fixed-size sliding window to traverse and divide the original detection image into blocks with a specified step size to obtain a set of reference image blocks; For the edge positions of the interlayer bonding interface detection region that are not covered by the reference image blocks after traversing according to the specified step size, add edge-fitting reference image blocks so that each pixel position in the interlayer bonding interface detection region is covered by at least one reference image block; For each reference... For each image patch, the center coordinates and window size of the reference image patch are proportionally converted according to the scaling ratio of each scale decomposed image relative to the original detection image. The corresponding scale image patch is then extracted from each scale decomposed image to obtain a multi-scale image patch group corresponding to the reference image patch. For each scale image patch in each multi-scale image patch group, the gray-level co-occurrence matrix is ​​calculated in a specified direction. Contrast, correlation, energy, and homogeneity are extracted from the gray-level co-occurrence matrix in each specified direction. The same texture feature parameters extracted in each specified direction are averaged to serve as the texture feature parameters of the corresponding scale image patch.

[0009] Furthermore, for each reference image patch, the step of combining multiple specified texture feature parameters of the corresponding image patches at each scale in scale order to form the texture feature vector corresponding to the reference image patch is as follows: For each reference image patch, the texture feature parameters extracted from each scale image patch in the multi-scale image patch group corresponding to the reference image patch are concatenated end to end in the order from the original scale to the progressively downsampled scale to obtain the texture feature vector corresponding to the reference image patch; and the texture feature vector corresponding to each reference image patch is subjected to L2 norm normalization processing.

[0010] Furthermore, the modal components and their corresponding center frequencies are iteratively updated using the alternating direction multiplier method until the convergence condition is met, resulting in multiple modal components. Specifically, the center frequency constraint parameters are set based on the one-dimensional signal, and the number of modal decompositions is set to the number corresponding to the background microtexture mode and the candidate interface defect mode; each modal component and its corresponding center frequency are initialized; based on the joint objective function, each modal component, its corresponding center frequency, and the Lagrange multiplier are alternately updated; the L2 norm of the difference between the modal component obtained in the current iteration and the modal component obtained in the previous iteration is calculated; when the L2 norm is lower than the preset convergence threshold, or the number of iterations reaches the preset maximum number of iterations, the iteration is stopped, resulting in multiple modal components.

[0011] Furthermore, in the process of variational mode decomposition of a one-dimensional signal, the specific steps for introducing an orthogonal constraint penalty term into the objective function are as follows: calculate the inner product between each pair of modal components; calculate the sum of the L2 norms of each inner product as the orthogonal constraint penalty term; combine the bandwidth constraint term of variational mode decomposition with the weighted data fidelity term and the weighted orthogonal constraint penalty term to form a joint objective function; and perform iterative updates using the alternating direction multiplier method based on the joint objective function.

[0012] Further, the specific steps for generating the pixel-level defect probability matrix are as follows: Candidate interface defect modal components corresponding to each reference image block in the normal charge image are used as normal training samples. Candidate interface defect modal components corresponding to reference image blocks in the charge image containing interface defect markers, where the ratio of the overlapping area of ​​the interface defect marker region to the area of ​​the corresponding reference image block reaches a preset marker ratio, are used as defect training samples. A generative enhancement discriminant network consisting of a sample generator and a defect discriminator is jointly trained. The sample generator takes random noise as input and outputs simulated candidate modal components with the same dimension as the normal training samples. The network parameters are updated based on the maximum mean difference between the simulated candidate modal components and the normal training samples, so that the distribution of the simulated candidate modal components approximates the distribution of the normal training samples. The defect discriminator takes normal training samples, defect training samples, or simulated candidate modal components as input. Normal training samples and simulated candidate modal components... The components correspond to normal labels, and the defective training samples correspond to defective labels. The defect discriminator outputs the block-level defect probability value of the reference image block corresponding to the input modal component containing the interface defect. After training, the defect discriminator is retrieved. The candidate interface defect modal component corresponding to each reference image block is input into the defect discriminator to obtain the block-level defect probability value corresponding to each reference image block. For each pixel position in the interlayer interface detection area, the set of reference image blocks covering the pixel position is determined. The initial fusion weight of each block-level defect probability value is determined according to the reciprocal of the sum of the distance between the pixel position and the center of each reference image block and a preset positive number. The initial fusion weights are normalized, and the block-level defect probability values ​​corresponding to the set of reference image blocks are weighted and fused to obtain the defect probability value corresponding to the pixel position. The defect probability values ​​are combined according to the arrangement relationship of each pixel position in the interlayer interface detection area to generate a pixel-level defect probability matrix.

[0013] Furthermore, during the training of the enhanced discriminant network, the specific steps for introducing a modality feature distribution difference regularization term into the loss function of the defect discriminator are as follows: In the feature extraction layer of the defect discriminator, normal discriminant features corresponding to normal training samples and defect discriminant features corresponding to defective training samples are extracted respectively; the maximum mean difference between normal discriminant features and defective discriminant features is calculated using a linear kernel; when the maximum mean difference is lower than the preset distribution interval, the square of the difference between the preset distribution interval and the maximum mean difference is used as the modality feature distribution difference regularization term; when the maximum mean difference is higher than or equal to the preset distribution interval, the modality feature distribution difference regularization term is set to zero; the block-level cross-entropy loss of the defect discriminator is weighted and summed with the modality feature distribution difference regularization term to obtain the total loss function of the defect discriminator.

[0014] Furthermore, the specific steps for determining the pixel positions with probability values ​​higher than a preset threshold as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface are as follows: Gaussian filtering is applied to the pixel-level defect probability matrix; pixel positions with probability values ​​higher than a preset threshold in the Gaussian-filtered probability matrix are marked as interface defect candidate pixels; connected component analysis is performed on the interface defect candidate pixels, connected components with areas lower than an area threshold are removed, and the pixel positions corresponding to the retained connected components are determined as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface.

[0015] Furthermore, based on the determined area proportion of the interface defect area, the forming state of the bonding interface between the multi-layer casting layers of the missile charge is determined, and the specific steps for adjusting the charge process parameters according to the determination results are as follows: Calculate the ratio of the total area of ​​the determined interface defect area to the total area of ​​the bonding interface between the multi-layer casting layers; compare the calculated ratio with the permissible ratio specified in the process standard; when the ratio exceeds the permissible ratio, adjust the casting speed parameters and casting temperature parameters of the subsequent casting layer or the next casting cycle; when the ratio is within the permissible ratio, maintain the currently set casting speed parameters and casting temperature parameters of the subsequent casting layer or the next casting cycle unchanged.

[0016] A machine vision-based missile propellant loading process control system includes: an image acquisition unit for acquiring images of the matte homogeneous solid propellant molding quality inspection of missiles; a multi-scale texture feature extraction unit for multi-scale decomposition of the inspection images, dividing the original inspection images into reference image blocks, determining the multi-scale image block groups corresponding to each reference image block based on the scaling relationship of each scale decomposed image relative to the original inspection image, calculating the gray-level co-occurrence matrix of each scale image block, and extracting specified texture feature parameters; a feature vector construction unit for combining the specified texture feature parameters of each scale image block in the multi-scale image block group corresponding to each reference image block according to the scale order to generate the texture feature vector corresponding to each reference image block; and a mode decomposition unit for treating the texture feature vector corresponding to each reference image block as a one-dimensional signal, setting center frequency constraint parameters and the number of mode decompositions, constructing a joint objective function including data fidelity terms and orthogonal constraint penalty terms, iteratively updating each modal component and the center frequency corresponding to each modal component through the alternating direction multiplier method, and calculating the power spectral entropy of each modal component. The system determines the background micro-texture modal components and candidate interface defect modal components based on the comparison between the power spectral entropy and a preset entropy threshold. A defect discrimination unit inputs the candidate interface defect modal components corresponding to each reference image block into the defect discriminator of the trained generative enhancement discriminator network, which outputs the block-level defect probability value indicating that the reference image block contains interface defects. A probability mapping unit maps the block-level defect probability value of each reference image block to the pixel positions it covers, based on the coverage area of ​​each reference image block in the original detection image. The system performs weighted fusion of multiple block-level defect probability values ​​corresponding to the same pixel location to generate a pixel-level defect probability matrix. The logic judgment and process control unit is used to determine the pixel locations in the pixel-level defect probability matrix with probability values ​​higher than a preset threshold as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface. Based on the area ratio of the determined interface defect areas, the system determines the forming state of the interlayer bonding interface of the missile charge multi-layer casting and adjusts the charge process parameters of the subsequent casting layer or the next casting cycle according to the determination results.

[0017] The present invention has the following beneficial effects:

[0018] (1) The missile loading process control method based on machine vision divides the original detection image into overlapping reference image blocks, and maps the center coordinates and coverage of each reference image block to the decomposed image at each scale according to the scaling relationship of each scale decomposed image relative to the original detection image, forming a multi-scale image block group covering the same original spatial region. Then, based on the gray-level co-occurrence matrix of each scale image block, the contrast, correlation, energy and homogeneity are extracted, so that the texture feature vector formed by the combination maintains the multi-scale texture information and its spatial correspondence at the same time, providing a texture feature basis for the separation of the background micro-texture modal component and the candidate interface defect modal component.

[0019] (2) The missile loading process control method based on machine vision constructs a joint objective function containing data fidelity term and orthogonal constraint penalty term by using texture feature vector as a one-dimensional signal, and uses the alternating direction multiplier method to update each modal component and its center frequency to reduce the correlation between different modal components. Then, the background microtexture modal component and candidate interface defect modal component are determined based on the power spectrum entropy to reduce the interference of background microtexture on interface defect discrimination.

[0020] (3) The missile propellant loading process control method based on machine vision outputs the block-level defect probability value of each reference image block by generating a defect discriminator of the enhanced discriminant network. According to the coverage relationship of the reference image blocks, the block-level defect probability value is mapped to the corresponding pixel position. Multiple block-level defect probability values ​​covering the same pixel position are weighted and fused to generate a pixel-level defect probability matrix. The interface defect area is determined by Gaussian filtering, threshold judgment and connected component screening, so that the defect judgment result can be used for the interlayer bonding interface forming state judgment and the adjustment of the propellant loading process parameters of the subsequent casting layer or the next casting cycle.

[0021] (4) The missile propellant loading process control system based on machine vision, through the image acquisition unit, multi-scale texture feature extraction unit, feature vector construction unit, modal decomposition unit, defect discrimination unit, probability mapping unit, and logic judgment and process control unit, sequentially completes the detection image acquisition, multi-scale image block correspondence, texture feature vector generation, candidate interface defect modal component extraction, block-level defect probability judgment, pixel-level defect probability mapping, interface defect area determination, and propellant loading process parameter adjustment, thereby realizing the data connection between missile propellant forming quality inspection and subsequent casting process control. Attached Figure Description

[0022] Figure 1 This is a flowchart of a missile propellant loading process control method based on machine vision according to the present invention.

[0023] Figure 2This is a flowchart illustrating the specific steps in a machine vision-based missile propellant loading process control method of the present invention to determine the pixel positions with probability values ​​higher than a preset threshold as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface.

[0024] Figure 3 This is a block diagram of a missile propellant loading process control system based on machine vision according to the present invention. Detailed Implementation

[0025] Please see Figure 1 This invention provides a technical solution: a missile propellant loading process control method based on machine vision, comprising the following steps: acquiring images of the matte homogeneous solid propellant molding quality inspection of missiles; performing multi-scale decomposition on the inspection images, dividing the original inspection images into reference image blocks, and determining the multi-scale image block group corresponding to each reference image block according to the scaling relationship of each scale decomposed image relative to the original inspection image; calculating the gray-level co-occurrence matrix of each scale image block, extracting specified texture feature parameters and combining them in scale order to generate texture feature vectors corresponding to each reference image block; using the texture feature vectors corresponding to each reference image block as one-dimensional signals to construct a system containing data fidelity terms and orthogonal constraint penalties. The joint objective function of the term is used to obtain multiple modal components through the alternating direction multiplier method. The background microtexture modal component and the candidate interface defect modal component are determined based on the power spectral entropy. The candidate interface defect modal components corresponding to each reference image block are input into the trained defect discriminator to obtain the block-level defect probability value of each reference image block. According to the coverage relationship of each reference image block, the block-level defect probability values ​​covering the same pixel position are weighted and fused to generate a pixel-level defect probability matrix. The interface defect region is determined by threshold judgment and connected component screening. The forming state of the interlayer bonding interface is determined according to the area ratio of the interface defect region, and the charging process parameters of the subsequent casting layer or the next casting cycle are adjusted according to the judgment result.

[0026] The acquired images of the matte homogeneous solid propellant molding quality inspection for missiles must completely cover the interlayer bonding interface of the multi-layer casting and a certain range of the surrounding propellant grain area. This ensures that the images include the natural background micro-texture of the propellant grain itself, the low-contrast gradient transition area of ​​the interlayer bonding interface, and the interface defects caused by possible interlayer debonding or micro-bubbles. In one embodiment, the acquired image resolution is not less than 1024×768 to ensure clear capture of pixel-level features, and to avoid stray light interference during the acquisition process.

[0027] Specifically, the steps for determining the multi-scale image block groups corresponding to each reference image block and extracting specified texture feature parameters are as follows: The detection image is decomposed into a multi-scale Gaussian pyramid to obtain decomposed images at multiple scales. Specifically, Gaussian filtering is used to smooth the original detection image, and then downsampling is performed to obtain decomposed images at different scales, forming a Gaussian pyramid structure. The kernel size of the Gaussian filter is set according to the noise level of the detection image. Downsampling uses an interlaced sampling method, so that the size of the decomposed images at adjacent scales decreases according to a preset scaling ratio.

[0028] In one implementation, the number of scales for Gaussian pyramid decomposition is set to 3 to 5. The first layer is the original detection image, the second layer is the original detection image downsampled after being filtered by a 3×3 Gaussian filter, the third layer is the decomposed image of the second layer downsampled after being filtered by a 5×5 Gaussian filter, and so on. Each scale of decomposed image covers the interlayer interface detection area.

[0029] A fixed-size sliding window is used to traverse and divide the original detection image into blocks with a specified step size, resulting in a set of reference image blocks. The size of the sliding window is set according to the resolution and texture detail density of the original detection image, and the sliding step size is set to 1 / 2 to 2 / 3 of the window size, so that overlapping areas are formed between adjacent reference image blocks.

[0030] For the edge positions of the interlayer bonding interface detection area that are not covered by the reference image block after traversing according to the specified step size, add a reference image block at the corresponding edge position so that each pixel position in the interlayer bonding interface detection area is covered by at least one reference image block.

[0031] For each reference image block, record the center coordinates and coverage area of ​​the reference image block in the original detection image. Based on the scaling ratio of each scale decomposition image relative to the original detection image, perform proportional conversion on the center coordinates and window size of the reference image block. Extract the corresponding scale image blocks from each scale decomposition image to form a multi-scale image block group corresponding to the reference image block.

[0032] In one implementation, the resolution of the original detection image is 1024×768, the window size of the reference image block is 64×64, and the sliding step size is 32; the scaling ratio of the second scale decomposition image relative to the original detection image is 1 / 2, the center coordinates of the reference image block are reduced to 1 / 2 of the original center coordinates, and a 32×32 window is used to crop the second scale image block; the scaling ratio of the third scale decomposition image relative to the original detection image is 1 / 4, the center coordinates of the reference image block are reduced to 1 / 4 of the original center coordinates, and a 16×16 window is used to crop the third scale image block.

[0033] The 64×64 image block, 32×32 image block, and 16×16 image block obtained by mapping from the same reference image block correspond to the same spatial region in the original detection image and together form the multi-scale image block group corresponding to the reference image block.

[0034] For each scale image patch in each multi-scale image patch group, the gray-level co-occurrence matrix is ​​calculated in a specified direction. The specified directions include 0°, 45°, 90°, and 135°, which are used to extract gray-level spatial distribution features in different directions.

[0035] When calculating the gray-level co-occurrence matrix (GLCM), the gray values ​​of image blocks at each scale are normalized to map them to a uniform gray level range. The frequency of co-occurrence of different gray value combinations under a specified direction and pixel stride is counted to form a GLCM for the corresponding direction. In one implementation, the pixel stride is set to 1, and the gray level is normalized to 16 levels. 16×16 GLCMs are calculated for each scale image block in the 0°, 45°, 90°, and 135° directions. The elements in the GLCM... This indicates that the gray values ​​of two pixels in the corresponding scale image patch, along a specified direction and separated by one pixel step, are respectively and The co-occurrence frequency of elements is calculated; contrast, correlation, energy, and homogeneity are extracted from the gray-level co-occurrence matrix for each specified direction. Contrast reflects the significance of gray-level differences in an image patch; a higher contrast indicates a more pronounced local gray-level contrast. Correlation reflects the degree of linear correlation between gray-level values ​​of adjacent pixels; a higher absolute value indicates a stronger correlation between gray-level values ​​of adjacent pixels. Energy reflects the concentration of elements in the gray-level co-occurrence matrix; a higher energy indicates a more uniform texture distribution in the image patch. Homogeneity reflects the smoothness of local gray-level changes; a higher homogeneity indicates a smaller gray-level difference between adjacent pixels.

[0036] Calculate the arithmetic mean of the same texture feature parameters in each specified direction to obtain the average contrast, average correlation, average energy, and average homogeneity of the image patch at the corresponding scale.

[0037] In one implementation, the contrast ratios in the four specified directions are 25.3, 24.8, 25.1 and 24.9, respectively, and their arithmetic mean is 25.0. 25.0 is used as the contrast feature parameter of the image patch at this scale, and the remaining texture feature parameters are calculated in the same way.

[0038] In this implementation scheme, decomposed images at different scales are obtained through Gaussian pyramids. The reference image blocks in the original detection image are used as a unified spatial reference. The center coordinates and coverage of each reference image block are mapped proportionally to the decomposed images at each scale, forming a multi-scale image block group corresponding to the same spatial region in the original detection image. By calculating the gray-level co-occurrence matrix of multiple specified directions for each scale image block and averaging the same texture feature parameters, the spatial correspondence between image blocks at different scales is maintained, and the influence of a single texture direction on the feature extraction results is reduced.

[0039] Specifically, for each reference image block, the specified texture feature parameters of each scale image block in its corresponding multi-scale image block group are combined to form the texture feature vector corresponding to the reference image block. The specific steps are as follows: For each reference image block, the contrast, correlation, energy and homogeneity extracted from each scale image block in the multi-scale image block group corresponding to the reference image block are concatenated end to end in the order from the original scale to the progressively downsampled scale to obtain the texture feature vector corresponding to the reference image block.

[0040] The dimension of the texture feature vector is equal to the product of the number of scales and the number of texture feature parameters at each scale, so that the texture feature parameters extracted from the image patch at each scale are all retained in the texture feature vector.

[0041] In one implementation, the Gaussian pyramid comprises three scales, each extracting four texture feature parameters: contrast, relevance, energy, and homogeneity. The texture feature vector has a dimension of 12. The texture feature parameters are arranged in the following order: contrast, relevance, energy, and homogeneity at the first scale; contrast, relevance, energy, and homogeneity at the second scale; and contrast, relevance, energy, and homogeneity at the third scale.

[0042] The texture feature vector corresponding to each baseline image patch is subjected to L2 norm normalization. L2 norm normalization is used to reduce the influence of differences in the numerical range of different texture feature parameters on the texture feature vector, and to avoid texture feature parameters with large numerical ranges having excessive weights in subsequent mode decomposition. Let the texture feature vector before normalization be: The normalized texture feature vector is: ; the normalized first Each texture feature vector element is calculated according to the following formula: ;in, The first part of the texture feature vector before normalization One element, The first element in the normalized texture feature vector One element, The first part of the texture feature vector before normalization One element, is the dimension of the texture feature vector.

[0043] In this implementation scheme, the texture feature parameters extracted from image patches at each scale are stitched together in order from the original scale to the progressively downsampled scale, so that the texture feature vector retains the feature arrangement relationship at different scales; and the influence of the difference in the numerical range of different texture feature parameters on the subsequent mode decomposition results is reduced by L2 norm normalization.

[0044] Specifically, the steps for separating the background micro-texture modal components and candidate interface defect modal components from the texture feature vectors corresponding to each reference image block are as follows: The texture feature vector corresponding to each reference image block is treated as a one-dimensional signal. A center frequency constraint parameter is set, and the number of modal decompositions is set. Specifically, since the texture feature vector is a high-dimensional vector formed by splicing texture parameters of multiple scales, it can be regarded as a one-dimensional discrete signal, and the length of the signal is equal to the dimension of the feature vector. The center frequency constraint parameter is set according to the background micro-texture characteristics of the matte homogeneous solid propellant of the missile, and is used to constrain the center frequency range of the modal decomposition to avoid decomposing irrelevant modal components. The number of modal decompositions is set to 2, corresponding to the background micro-texture modal components and the candidate interface defect modal components, respectively. In one embodiment, the center frequency constraint parameter is set to... The frequency of the one-dimensional signal is 0.2-0.8 times the dominant frequency, and the number of mode decompositions is fixed at 2 to ensure that the decomposition result contains only two target mode components and does not generate redundant modes. Variational mode decomposition is performed on the one-dimensional signal, and the center frequency of each mode component and its corresponding center frequency are iteratively updated by the alternating direction multiplier method until the convergence condition is met. Specifically, the core of variational mode decomposition is to construct and solve a variational problem to achieve adaptive separation of mode components. The alternating direction multiplier method is used to iteratively update the mode components and center frequencies to accelerate the convergence speed. During the iteration process, the difference between the current mode component and the mode component of the previous iteration is calculated after each update. When the L2 norm of the difference is less than the preset convergence threshold, or the number of iterations reaches the preset maximum number of iterations, the iteration stops. In one embodiment, the preset convergence threshold is set to 1×10. -5The maximum number of iterations is preset to 200. During the iteration process, after each update of a modal component, the difference between it and the previous modal component is calculated. If the L2 norm of the difference is less than the convergence threshold for three consecutive iterations, the convergence condition is met, and the iteration stops. After convergence, multiple modal components are obtained. Specifically, since the number of modal decompositions is set to 2, two modal components are obtained after convergence. The lengths of the two modal components are consistent with the length of the input one-dimensional signal, and the error between the superposition result of the two modal components and the original one-dimensional signal is within a preset range. One modal component mainly contains relatively smooth and low-frequency background micro-texture features; the other modal component mainly contains relatively drastic gray-level changes and high-frequency candidate interface defect features. In one embodiment... After convergence, modal components A and B are obtained. The element values ​​of modal component A fluctuate less and tend to be stable overall, corresponding to the background micro-texture features. The element values ​​of modal component B fluctuate more and have relatively obvious peaks, corresponding to the candidate interface defect features. The power spectral entropy of each modal component is calculated as follows: the power spectral entropy is used to measure the complexity of the modal component. The lower the power spectral entropy value, the more concentrated the frequency distribution of the modal component and the lower the complexity, corresponding to the gently changing background micro-texture. The higher the power spectral entropy value, the more dispersed the frequency distribution of the modal component and the higher the complexity, corresponding to the drastically changing interface defect. The calculation steps are: perform Fourier transform on the modal component to obtain its power spectrum; normalize the power spectrum to obtain the power spectral probability distribution, where the th The normalized power spectral probability corresponding to each frequency point is calculated according to the following formula: Calculate the power spectral entropy based on the power spectral probability distribution: ;in, For power spectral entropy, For the first The power spectral density values ​​corresponding to each frequency point For the first The power spectral density values ​​corresponding to each frequency point For the first Normalized power spectrum probability corresponding to each frequency point The power spectrum contains the number of frequency points. In one implementation, the power spectral entropy of mode component A is 0.32, and the power spectral entropy of mode component B is 0.87, which is significantly higher than that of mode component A. Mode components with power spectral entropy values ​​below a preset entropy threshold are marked as background micro-texture mode components, and mode components with power spectral entropy values ​​above or equal to the preset entropy threshold are marked as candidate interface defect mode components. Specifically, the preset entropy threshold is determined based on a large number of sample training. By statistically analyzing the power spectral entropy of mode components in normal charge samples and samples with defects, it is determined that the two modes can be effectively distinguished. The threshold of the components; in one embodiment, by statistically analyzing the modal components of 1000 normal charge samples and 1000 samples with defects, a preset entropy threshold of 0.5 is determined. Modal components with power spectral entropy below 0.5 are marked as background microtexture modal components, and those with power spectral entropy above or equal to 0.5 are marked as candidate interface defect modal components. If the power spectral entropy of the two modal components obtained after convergence is 0.32 and 0.87 respectively, then the modal component corresponding to 0.32 is marked as a background microtexture modal component, and the modal component corresponding to 0.87 is marked as a candidate interface defect modal component.

[0045] In the process of variational mode decomposition of a one-dimensional signal, an orthogonality constraint penalty term is introduced into the objective function. The specific steps are as follows: Calculate the inner product between each pair of modal components. Specifically, since the number of mode decompositions is 2, only the inner product between two modal components needs to be calculated. The inner product is used to measure the degree of orthogonality between two modal components. The closer the inner product value is to 0, the better the orthogonality between the two modal components and the better the separation effect. Let the two modal components be... and The inner product between two modal components is calculated according to the following formula: ;in, For the first modal component, For the second modal component, The first modal component One element, The first of the second modal components One element, The number of elements contained in each modal component; the sum of the L2 norms of each inner product is calculated as the orthogonality constraint penalty term, specifically: for multiple modal components obtained from modal decomposition, the orthogonality constraint penalty term is calculated according to the following formula: ;in, This is an orthogonal constraint penalty term. For the first One modal component, For the first One modal component, The number of mode decompositions is 2; when the number of mode decompositions is 2, the orthogonal constraint penalty term is calculated according to the following formula: ;in, and These are the first and second modal components, respectively. In one embodiment, the inner product of the two modal components is 0.023, and their L2 norm is 0.023, meaning the orthogonality constraint penalty term is 0.023. The bandwidth constraint term of the variational mode decomposition is combined with the weighted data fidelity term and the weighted orthogonality constraint penalty term to form a joint objective function, specifically: the data fidelity term is used to characterize the difference between the superposition result of each modal component after decomposition and the original one-dimensional signal, and the data fidelity term is calculated according to the following formula: ;in, For data fidelity items, The original one-dimensional signal, For the first One modal component, The number of mode decompositions; the joint objective function of the variational mode decomposition is constructed as follows: in, For the first One modal component, For the first The center frequency corresponding to each modal component The number of modal decompositions, The weighting coefficients for data fidelity items. The weight coefficients for the orthogonal constraint penalty term are... For time variables The derivative operator, * for convolution, The imaginary unit is used; in one implementation, the weight coefficient of the data fidelity term is set to 1000, and the weight coefficient of the orthogonal constraint penalty term is set according to the correlation between the modal components. While maintaining the superposition result of each modal component close to the original one-dimensional signal, the feature aliasing between different modal components is reduced; the alternating direction multiplier method is used for iterative update based on the joint objective function, specifically: fixing the center frequency and Lagrange multiplier corresponding to each modal component, and updating each modal component; fixing each modal component and Lagrange multiplier, and updating the center frequency corresponding to each modal component; root The Lagrange multipliers are updated based on the reconstruction difference between the superposition results of each modal component and the original one-dimensional signal; the orthogonal constraint penalty term is recalculated based on the updated modal components; after each iteration, the L2 norm of the difference between the modal component obtained in the current iteration and the modal component obtained in the previous iteration is calculated. Iteration is stopped when the L2 norm is lower than a preset convergence threshold or when the number of iterations reaches a preset maximum number of iterations. In one embodiment, when the iteration reaches 86, the L2 norm of the difference between the modal component obtained in the current iteration and the modal component obtained in the previous iteration is 8.2 × 10⁻⁶. -6 The convergence threshold is 1×10⁻⁶. -5 Stop the iteration. At this point, the difference between the superposition result of each modal component and the original one-dimensional signal is within the preset range.

[0046] In this implementation scheme, the texture feature vector is used as a one-dimensional signal for variational mode decomposition. The difference between the superposition result of each modal component and the original one-dimensional signal is limited by the data fidelity term. The correlation between different modal components is limited by the orthogonal constraint penalty term. The background microtexture modal component and the candidate interface defect modal component are determined based on the comparison result of the power spectrum entropy and the preset entropy threshold.

[0047] Specifically, the steps for generating the pixel-level defect probability matrix are as follows: both normal training samples and defect training samples are derived from actual missile solid propellant loading imaging samples. For normal loading images, the candidate interface defect mode components corresponding to each reference image block are used as normal training samples; for loading images containing interface defect markers, the ratio of the overlap area between each reference image block and the interface defect marker area to the area of ​​the corresponding reference image block is calculated, and the candidate interface defect mode components corresponding to the reference image blocks whose ratio reaches the preset marker ratio are used as defect training samples.

[0048] Normal training samples and defective training samples have the same data dimensions, and are labeled with normal and defective labels respectively.

[0049] In one implementation, a preset labeling ratio is set to 50%. Candidate interface defect modal components corresponding to 5000 reference image blocks are selected from normal charge images, and another 5000 candidate interface defect modal components corresponding to reference image blocks meeting the preset labeling ratio are selected from charge images containing interface defect labels. These are used to form a training dataset. The training and validation sets are divided according to the original detection images, with 80% of the training samples corresponding to the original detection images used for training and 20% used for validation. This avoids reference image blocks corresponding to the same original detection image being included in both the training and validation sets simultaneously.

[0050] The generative enhancement discriminant network consists of a sample generator and a defect discriminator. The sample generator adopts a fully connected network structure, takes a random noise vector as input, and outputs simulated candidate mode components with the same dimensions as the normal training samples.

[0051] The maximum mean difference between the simulated candidate modal component set and the normal training sample set is calculated, and this maximum mean difference is used as the generation loss of the sample generator. The network parameters of the sample generator are updated according to the generation loss, so that the distribution of the simulated candidate modal components gradually approximates the distribution of the normal training samples. The maximum mean difference can be calculated using the maximum mean difference calculation formula described later, where the two input sets are the simulated candidate modal component set and the normal training sample set, respectively.

[0052] The defect discriminator includes a fully connected feature extraction layer and a classification output layer, which takes normal training samples, defective training samples or simulated candidate modal components as input. Normal training samples and simulated candidate modal components correspond to normal labels, and defective training samples correspond to defective labels.

[0053] The defect discriminator generates corresponding discriminative feature vectors through a fully connected feature extraction layer, and outputs block-level defect probability values ​​by a classification output layer. The closer the block-level defect probability value is to 1, the higher the probability that the corresponding reference image block contains interface defects; the closer the block-level defect probability value is to 0, the lower the probability that the corresponding reference image block contains interface defects. After training, the defect discriminator is extracted, and the candidate interface defect modal components corresponding to each reference image block are input into the defect discriminator to obtain the block-level defect probability value of the corresponding reference image block.

[0054] Based on the coverage of each reference image block in the original detection image, the block-level defect probability value of each reference image block is mapped to the pixel positions covered by that reference image block.

[0055] For each pixel location in the interlayer bonding interface detection region, a set of reference image blocks covering that pixel location is determined. The distance between the pixel location and the center of each reference image block is calculated, and each distance is added to a preset positive number. The reciprocal of the resulting value is used as the initial fusion weight for the corresponding block-level defect probability value.

[0056] Divide each initial fusion weight by the sum of all initial fusion weights to obtain the normalized fusion weight; then, sum the block-level defect probability values ​​covering the same pixel location according to the normalized fusion weight to obtain the defect probability value corresponding to that pixel location.

[0057] Arrange the defect probability values ​​according to the row and column positions of each pixel in the interlayer interface detection area to generate a pixel-level defect probability matrix.

[0058] During the training of the enhanced discriminant network, a modality feature distribution difference regularization term is introduced into the loss function of the defect discriminator. The specific steps are as follows: In the fully connected feature extraction layer of the defect discriminator, the normal discriminant feature vectors corresponding to normal training samples and the defect discriminant feature vectors corresponding to defect training samples are extracted respectively, and normal discriminant feature sets and defect discriminant feature sets are formed respectively.

[0059] The maximum mean difference between the normal discriminant feature set and the defective discriminant feature set is calculated using a linear kernel. The mean vectors of the normal and defective feature sets are calculated separately, and the L2 norm between the two mean vectors is also calculated to obtain the maximum mean difference. Specifically, the calculation is as follows: ;in, The maximum mean difference between the normal discriminant feature set and the defect discriminant feature set. For normal discriminative feature set, For the set of features for defect discrimination, , These are the mean vectors of the normal discriminant feature set and the defect discriminant feature set, respectively. The maximum mean difference is compared with a preset distribution interval. When the maximum mean difference is lower than the preset distribution interval, the square of the difference between the preset distribution interval and the maximum mean difference is used as the modal feature distribution difference regularization term; when the maximum mean difference is higher than or equal to the preset distribution interval, the modal feature distribution difference regularization term is set to zero.

[0060] The block-level cross-entropy loss of the defect discriminator measures the error between the block-level defect probability value and the corresponding training label of the benchmark image block. The total loss function of the defect discriminator is obtained by weighted summing the block-level cross-entropy loss with the modal feature distribution difference regularization term.

[0061] In one implementation, the weight coefficient of the modal feature distribution difference regularization term is set to 0.1, and the network parameters of the defect discriminator are updated according to the total loss function of the defect discriminator, so that the normal discriminant features and the defect discriminant features maintain a preset distribution interval.

[0062] In this implementation scheme, block-level training samples are constructed using candidate interface defect modal components corresponding to each reference image block in normal charge images and charge images containing interface defect markers. Normal training samples are supplemented by a sample generator, and the block-level defect probability value of each reference image block is obtained using a defect discriminator. Then, the block-level defect probability value is mapped and weighted according to the coverage relationship of the reference image blocks to generate a pixel-level defect probability matrix.

[0063] Specifically, such as Figure 2As shown, the specific steps for determining pixel locations with probability values ​​higher than a preset threshold as interface defects caused by interlayer debonding and microbubbles in the interlayer bonding interface are as follows: Gaussian filtering is applied to the pixel-level defect probability matrix. Specifically, the purpose of Gaussian filtering is to eliminate noise interference in the probability matrix, smooth fluctuations in probability values, and avoid misjudgments caused by noise. A two-dimensional Gaussian filter kernel is used, with the kernel size set according to the size of the probability matrix. The standard deviation of the kernel is adaptively adjusted according to the noise intensity to ensure that details of the interface defect area are not lost while smoothing noise. The row and column dimensions of the pixel-level defect probability matrix are consistent with the row and column dimensions of the interlayer bonding interface detection area. In one embodiment, the original detection image... The image resolution is 1024×768. When the interlayer combined interface detection area completely covers the original detection image, the size of the pixel-level defect probability matrix is ​​1024×768. A 3×3 Gaussian filter kernel is used to filter the pixel-level defect probability matrix, and the standard deviation of the filter kernel is set to 1.0. During the filtering process, the defect probability value of each pixel position is obtained by Gaussian weighted averaging of its own and surrounding pixel positions to reduce local probability fluctuations in the pixel-level defect probability matrix. Pixel positions with probability values ​​higher than a preset threshold in the Gaussian filtered probability matrix are marked as interface defect candidate pixels. Specifically, the preset threshold is determined based on a large number of sample statistics. By analyzing the probability value distribution of normal pixels and defective pixels, A threshold is determined to effectively distinguish between the two types of pixels, balancing the false alarm rate and the false negative rate. In one implementation, the probability values ​​of 1000 samples are statistically analyzed, and a preset threshold of 0.3 is determined based on the probability value distribution of normal pixels and defective pixels. Pixels with a probability value higher than 0.3 after Gaussian filtering are marked as candidate pixels for interface defects. Connectivity analysis is performed on the candidate pixels for interface defects, and connected components with an area lower than the area threshold are removed. The pixel positions corresponding to the retained connected components are determined to be interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface. Specifically, connected component analysis is used to remove isolated false candidate pixels (caused by noise) and retain continuous real defect areas. Connectivity is defined as adjacent... The region is composed of candidate pixels (4-neighbor or 8-neighbor), and the area of ​​each connected region (i.e., the number of pixels in the connected region) is calculated. The area threshold is set according to the minimum size of microbubbles and interlayer debonding to ensure that the smallest real defect area is retained while isolated false pixels are removed. In one implementation, 8-neighbor connected region analysis is used, and the area threshold is set to 5 pixels. That is, connected regions with fewer than 5 pixels are judged as false regions and removed, and connected regions with more than or equal to 5 pixels are judged as real defect areas. For example, a connected region containing 3 pixels has an area lower than the threshold and is removed; a connected region containing 8 pixels has an area higher than the threshold and is retained. The corresponding pixel position is the interface defect area.

[0064] In this implementation scheme, Gaussian filtering is applied to the pixel-level defect probability matrix to reduce local probability value fluctuations; candidate pixels for interface defects are extracted according to a preset threshold, and candidate regions with smaller areas are eliminated by a connected component area threshold to obtain the interface defect region used for determining the forming state.

[0065] Specifically, the following steps are taken to determine the forming state of the interlayer bonding interface of the missile propellant multi-layer casting based on the determined area proportion of the interface defect area, and to adjust the propellant process parameters according to the determination result: The ratio of the total area of ​​the determined interface defect area to the total area of ​​the interlayer bonding interface of the multi-layer casting is calculated. Specifically, the pixel area in the image is converted to the actual physical area through the correspondence between image pixel size and actual physical size; the number of pixels in all retained connected components (interface defect areas) is counted, multiplied by the actual area corresponding to a single pixel, to obtain the total area of ​​the interface defect area; the pre-determined interlayer bonding interface detection area in the detection image is used as the statistical range for the area proportion, the number of pixels contained in the detection area is counted, and multiplied by the actual area corresponding to a single pixel, to obtain the total area of ​​the interlayer bonding interface; the ratio of the two is calculated, i.e., the defect area proportion; in one embodiment, the actual area corresponding to a single pixel is 0.01 mm², and the total number of pixels in the interface defect area is... The total area is 2.5 mm², and the total number of pixels at the interlayer bonding interface is 10,000, with a total area of ​​100 mm². Therefore, the defect area ratio is 2.5%. The calculated ratio is compared with the permissible ratio specified in the process standard. Specifically, the permissible ratio specified in the process standard is set according to the performance requirements of missile solid propellant charge and is divided into a qualified permissible ratio and a limit permissible ratio. The qualified permissible ratio is the maximum acceptable defect area ratio for normal production, and the limit permissible ratio is the minimum unacceptable defect area ratio. During the comparison, the calculated defect area ratio is compared with the two permissible ratios to determine the forming state level. In one embodiment, the qualified permissible ratio specified in the process standard is 3%, and the limit permissible ratio is 5%. If the calculated defect area ratio is 2.5%, it is lower than the qualified permissible ratio and is judged to be in a qualified forming state. If the ratio is 3.8%, it is higher than the qualified permissible ratio and lower than the limit permissible ratio and is judged to be slightly unqualified. If the ratio is 5.A defect rate of 2% is considered above the permissible limit and is deemed a serious non-compliance. When the ratio exceeds the permissible limit, the casting speed and temperature parameters for subsequent casting layers or the next casting cycle are adjusted. Specifically, the process parameters are adjusted in stages based on the degree to which the defect area exceeds the permissible limit. If the percentage is higher than the acceptable permissible ratio but lower than the permissible limit (minor non-compliance), the parameters are slightly adjusted, reducing the casting speed and fine-tuning the casting temperature to reduce bubbles and debonding during slurry flow. If the percentage is higher than the permissible limit (serious non-compliance), the parameters are significantly adjusted, reducing the casting speed and adjusting the casting temperature to the optimal range. Simultaneously, the casting equipment is inspected to prevent further expansion of the defect. In one embodiment, for minor defects, the casting speed is reduced by 10%, and the casting temperature is adjusted to the preset optimal temperature ±2℃; for severe defects, the casting speed is reduced by 20%, and the casting temperature is adjusted to the preset optimal temperature, while simultaneously checking the discharge port and temperature control system of the casting equipment; when the ratio is within the permissible ratio, the currently set casting speed and temperature parameters for subsequent casting layers or the next casting cycle remain unchanged. Specifically, if the proportion of defective area is lower than the permissible ratio, it indicates that the proportion of defective area at the current interlayer interface is within the permissible range, and the currently set casting speed and temperature parameters for subsequent casting layers or the next casting cycle remain unchanged.

[0066] Record the texture features, defect area ratio, casting speed parameters, and casting temperature parameters corresponding to the current detected image to form a record of the charging process parameters.

[0067] In one implementation, the defect area ratio is 2.5%, which is lower than the acceptable ratio of 3%. The casting speed and casting temperature parameters currently set for subsequent casting layers or the next casting cycle are kept unchanged, and the corresponding texture features, defect area ratio, casting speed parameters and casting temperature parameters are recorded.

[0068] In this implementation scheme, the total area of ​​the interface defect region and the total area of ​​the interlayer bonding interface are calculated based on the correspondence between image pixel size and actual physical size, and the defect area ratio is obtained. The molding state is determined based on the comparison result of the defect area ratio and the allowable ratio, and then the casting speed parameters and casting temperature parameters are selected to be maintained or adjusted.

[0069] Please see Figure 3This invention provides a technical solution: a missile propellant loading process control system based on machine vision, comprising: an image acquisition unit for acquiring images of the matte homogeneous solid propellant molding quality inspection of missiles; a multi-scale texture feature extraction unit for performing multi-scale decomposition on the inspection images, dividing the original inspection images into reference image blocks, determining the multi-scale image block groups corresponding to each reference image block according to the scaling relationship of each scale decomposed image relative to the original inspection image, calculating the gray-level co-occurrence matrix of each scale image block and extracting specified texture feature parameters; a feature vector construction unit for combining the specified texture feature parameters of each scale image block in the multi-scale image block group corresponding to each reference image block according to the scale order to generate texture feature vectors corresponding to each reference image block; and a mode decomposition unit for using the texture feature vector corresponding to each reference image block as a one-dimensional signal, setting center frequency constraint parameters and the number of mode decompositions, constructing a joint objective function containing data fidelity terms and orthogonal constraint penalty terms, iteratively updating each modal component and the center frequency corresponding to each modal component through the alternating direction multiplier method until the convergence condition is met, and obtaining multiple modal components. The system calculates the power spectral entropy of each modal component and determines the background micro-texture modal component and candidate interface defect modal component based on the comparison between the power spectral entropy and a preset entropy threshold. A defect discrimination unit inputs the candidate interface defect modal component corresponding to each reference image block into the defect discriminator of the trained generative enhancement discriminant network, which outputs the block-level defect probability value indicating that the reference image block contains interface defects. A probability mapping unit maps the block-level defect probability value of each reference image block to a map based on the coverage area of ​​each reference image block in the original detection image. It covers the pixel locations and performs weighted fusion of multiple block-level defect probability values ​​corresponding to the same pixel location to generate a pixel-level defect probability matrix; the logic judgment and process control unit is used to determine the pixel locations in the pixel-level defect probability matrix with probability values ​​higher than a preset threshold as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface. Based on the area ratio of the determined interface defect areas, it makes a judgment on the forming state of the interlayer bonding interface of the multi-layer casting of missile charge, and adjusts the charge process parameters of subsequent casting layers or the next casting cycle according to the judgment results.

[0070] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0071] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A missile propellant loading process control method based on machine vision, characterized in that, Includes the following steps: Acquire images of the matte homogeneous solid propellant charge forming quality inspection for missiles; The detection image is decomposed into multiple scales, and reference image blocks are divided in the original detection image. The multi-scale image block group corresponding to each reference image block is determined according to the scaling relationship of each scale decomposed image relative to the original detection image. The gray-level co-occurrence matrix of image patches at each scale is calculated, specified texture feature parameters are extracted and combined in scale order to generate texture feature vectors corresponding to each reference image patch. The texture feature vectors corresponding to each reference image patch are used as one-dimensional signals to construct a joint objective function containing data fidelity terms and orthogonal constraint penalty terms. Multiple modal components are obtained through the alternating direction multiplier method, and the background micro-texture modal components and candidate interface defect modal components are determined based on the power spectral entropy. The candidate interface defect modal components corresponding to each reference image patch are input into a trained defect discriminator to obtain the block-level defect probability value of each reference image patch. According to the coverage relationship of each reference image patch, the block-level defect probability values ​​covering the same pixel position are weighted and fused to generate a pixel-level defect probability matrix. The interface defect region is determined by threshold judgment and connected component screening. The forming state of the interlayer bonding interface is determined based on the area ratio of the interface defect area, and the charging process parameters of the subsequent casting layer or the next casting cycle are adjusted according to the determination result.

2. The missile propellant loading process control method based on machine vision according to claim 1, characterized in that, The specific steps for determining the multi-scale image block group corresponding to each benchmark image block and extracting the specified texture feature parameters are as follows: Perform multi-scale Gaussian pyramid decomposition on the detection image to obtain decomposed images at multiple scales; use a fixed-size sliding window with a specified step size to traverse and divide the original detection image into blocks to obtain the benchmark image block set. For the edge positions of the interlayer bonding interface detection region that are not covered by the reference image block after traversing according to the specified step size, a new edge-fitting reference image block is added so that each pixel position in the interlayer bonding interface detection region is covered by at least one reference image block; for each reference image block, the center coordinates and window size of the reference image block are proportionally converted according to the scaling ratio of each scale decomposition image relative to the original detection image, and the corresponding scale image block is extracted from each scale decomposition image to obtain the multi-scale image block group corresponding to the reference image block. For each scale image patch in each multi-scale image patch group, calculate the gray-level co-occurrence matrix in the specified direction; Contrast, correlation, energy, and homogeneity are extracted from the gray-level co-occurrence matrix in each specified direction. The same texture feature parameters extracted from each specified direction are averaged to serve as the texture feature parameters of the corresponding scale image patch.

3. The missile propellant loading process control method based on machine vision according to claim 1, characterized in that, For each reference image patch, the steps for combining multiple specified texture feature parameters of the corresponding image patches at each scale in scale order to form the texture feature vector corresponding to the reference image patch are as follows: For each reference image patch, the texture feature parameters extracted from each scale image patch in the multi-scale image patch group corresponding to the reference image patch are concatenated end to end in the order from the original scale to the progressively downsampled scale to obtain the texture feature vector corresponding to the reference image patch; and the texture feature vector corresponding to each reference image patch is subjected to L2 norm normalization.

4. The missile propellant loading process control method based on machine vision according to claim 1, characterized in that, The modal components and their corresponding center frequencies are iteratively updated using the alternating direction multiplier method until the convergence condition is met, resulting in multiple modal components. Specifically, the center frequency constraint parameters are set based on the one-dimensional signal, and the number of modal decompositions is set to the number corresponding to the background microtexture mode and the candidate interface defect mode; each modal component and its corresponding center frequency are initialized; based on the joint objective function, each modal component, its corresponding center frequency, and the Lagrange multiplier are alternately updated; the L2 norm of the difference between the modal component obtained in the current iteration and the modal component obtained in the previous iteration is calculated; when the L2 norm is lower than the preset convergence threshold, or the number of iterations reaches the preset maximum number of iterations, the iteration is stopped, resulting in multiple modal components.

5. The missile propellant loading process control method based on machine vision according to claim 4, characterized in that, The specific steps for introducing an orthogonal constraint penalty term into the objective function during variational mode decomposition of a one-dimensional signal are as follows: calculate the inner product between each pair of modal components; calculate the sum of the L2 norms of each inner product as the orthogonal constraint penalty term; combine the bandwidth constraint term of variational mode decomposition with the weighted data fidelity term and the weighted orthogonal constraint penalty term to form a joint objective function; and perform iterative updates using the alternating direction multiplier method based on the joint objective function.

6. The missile propellant loading process control method based on machine vision according to claim 1, characterized in that, The specific steps for generating the pixel-level defect probability matrix are as follows: the candidate interface defect modal components corresponding to each reference image block in the normal drug loading image are used as normal training samples, and the candidate interface defect modal components corresponding to the reference image blocks in the drug loading image containing interface defect markers whose ratio of the overlapping area of ​​the interface defect marker area to the area of ​​the corresponding reference image block reaches the preset marker ratio are used as defect training samples. A generative enhancement discriminant network consisting of a sample generator and a defect discriminator is jointly trained. The sample generator takes random noise as input and outputs simulated candidate modal components with the same dimensions as normal training samples, updating network parameters based on the distribution differences between the simulated candidate modal components and normal training samples. The defect discriminator takes normal training samples, defective training samples, or simulated candidate modal components as input. Normal training samples and simulated candidate modal components correspond to normal labels, while defective training samples correspond to defect labels. The defect discriminator outputs the block-level defect probability value of the reference image patch corresponding to the input modal component, indicating whether it contains interface defects. After training, the defect discriminator is removed; each reference image patch is then compared with... The corresponding candidate interface defect modal components are input into the defect discriminator to obtain the block-level defect probability value corresponding to each reference image block. For each pixel position in the interlayer bonding interface detection area, the set of reference image blocks covering the pixel position is determined. The initial fusion weight of each block-level defect probability value is determined according to the reciprocal of the sum of the distance between the pixel position and the center of each reference image block and a preset positive number. The initial fusion weights are normalized, and the block-level defect probability values ​​corresponding to the set of reference image blocks are weighted and fused to obtain the defect probability value corresponding to the pixel position. The defect probability values ​​are combined according to the arrangement relationship of each pixel position in the interlayer bonding interface detection area to generate a pixel-level defect probability matrix.

7. The missile propellant loading process control method based on machine vision according to claim 6, characterized in that, The specific steps for introducing a modality feature distribution difference regularization term into the loss function of the defect discriminator during the training of the enhanced discriminator are as follows: In the feature extraction layer of the defect discriminator, extract the normal discriminant features corresponding to the normal training samples and the defect discriminant features corresponding to the defective training samples respectively. A linear kernel is used to calculate the maximum mean difference between normal discriminant features and defect discriminant features. When the maximum mean difference is lower than the preset distribution interval, the square of the difference between the preset distribution interval and the maximum mean difference is used as the modal feature distribution difference regularization term. When the maximum mean difference is higher than or equal to the preset distribution interval, the modal feature distribution difference regularization term is set to zero. The block-level cross-entropy loss of the defect discriminator is weighted and summed with the modal feature distribution difference regularization term to obtain the total loss function of the defect discriminator.

8. The missile propellant loading process control method based on machine vision according to claim 1, characterized in that, The specific steps for determining the pixel positions with probability values ​​higher than a preset threshold as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface are as follows: Gaussian filtering is applied to the pixel-level defect probability matrix; pixel positions with probability values ​​higher than the preset threshold in the Gaussian-filtered probability matrix are marked as interface defect candidate pixels; connected component analysis is performed on the interface defect candidate pixels, and connected components with areas lower than the area threshold are removed, and the pixel positions corresponding to the retained connected components are determined as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface.

9. The missile propellant loading process control method based on machine vision according to claim 1, characterized in that, Based on the determined area proportion of the interface defect region, the forming state of the bonding interface between the multi-layer casting layers of the missile charge is determined, and the specific steps for adjusting the charge process parameters according to the determination results are as follows: Calculate the ratio of the total area of ​​the determined interface defect region to the total area of ​​the bonding interface between the multi-layer casting layers; compare the calculated ratio with the permissible ratio specified in the process standard; when the ratio exceeds the permissible ratio, adjust the casting speed parameters and casting temperature parameters of the subsequent casting layer or the next casting cycle; when the ratio is within the permissible ratio, maintain the currently set casting speed parameters and casting temperature parameters of the subsequent casting layer or the next casting cycle unchanged.

10. A missile propellant loading process control system based on machine vision, employing the missile propellant loading process control method based on machine vision as described in any one of claims 1-9, characterized in that, include: The image acquisition unit is used to acquire images of the matte homogeneous solid propellant molding quality inspection of missiles; the multi-scale texture feature extraction unit is used to decompose the inspection image into multiple scales, divide the original inspection image into reference image blocks, determine the multi-scale image block group corresponding to each reference image block according to the scaling relationship of each scale decomposed image relative to the original inspection image, calculate the gray-level co-occurrence matrix of each scale image block and extract the specified texture feature parameters. The feature vector construction unit is used to combine the specified texture feature parameters of each scale image block in the multi-scale image block group corresponding to each reference image block according to the scale order to generate the texture feature vector corresponding to each reference image block. The modal decomposition unit is used to treat the texture feature vector corresponding to each reference image block as a one-dimensional signal, set the center frequency constraint parameter and the number of modal decompositions, construct a joint objective function including data fidelity term and orthogonal constraint penalty term, iteratively update each modal component and the center frequency corresponding to each modal component through the alternating direction multiplier method, calculate the power spectral entropy of each modal component, and determine the background micro-texture modal component and candidate interface defect modal component based on the comparison result of the power spectral entropy and the preset entropy threshold. The defect discrimination unit is used to input the candidate interface defect modal components corresponding to each reference image block into the defect discriminator of the trained generative enhancement discriminator network, and the defect discriminator outputs the block-level defect probability value of the reference image block containing interface defects. The probability mapping unit is used to map the block-level defect probability value of each reference image block to the pixel position it covers based on the coverage area of ​​each reference image block in the original detection image, and to perform weighted fusion of multiple block-level defect probability values ​​corresponding to the same pixel position to generate a pixel-level defect probability matrix. The logic judgment and process control unit is used to determine the pixel positions with probability values ​​higher than a preset threshold in the pixel-level defect probability matrix as interface defect areas caused by interlayer debonding and microbubbles in the interlayer bonding interface. Based on the area ratio of the determined interface defect areas, the forming state of the interlayer bonding interface of the missile charge multi-layer casting is determined, and the charging process parameters of the subsequent casting layer or the next casting cycle are adjusted according to the judgment results.

Images