A deep learning-based suction powder image quality evaluation method and system

Abstract

The present application relates to the technical field of image pattern recognition and deep learning, and particularly relates to a kind of suction powder image quality evaluation method and system based on deep learning, the method comprises: obtaining suction powder image, for each pixel point, extract multi-scale relative darkness and neighborhood gradient direction consistency to determine structure characteristic value;Extract orthogonal spectral residual and inverse weight to determine spectral heterogeneity, fuse the above features to generate physical saliency matrix;Construct double-flow convolutional neural network, map physical saliency matrix to spatial attention weight, use affine transformation to force modulation intermediate layer feature, generate physically enhanced high-dimensional semantic feature data, globally pool high-dimensional semantic feature data and classify, output quality evaluation result.The method avoids the defect feature dilution effect caused by global average pooling through the feature modulation architecture guided by physical prior, improves the accuracy of suction powder quality evaluation result.

Description

Technical Field

[0001] This invention relates to the fields of image pattern recognition and deep learning technology. Specifically, it relates to a method and system for evaluating the image quality of suction powder based on deep learning. Background Technology

[0002] In the automated production system of the grain and oil processing industry, the quality and purity of the suction powder, a key byproduct collected by the air separation system, is a core indicator for measuring the technological level of the production line and the grade of the final product. To ensure consistent quality and improve detection efficiency during large-scale production, the industry widely adopts automated image detection technology based on computer vision. This involves using high-resolution industrial cameras to acquire real-time images of the suction powder under specific lighting conditions, and then analyzing its color, particle size, and morphological characteristics through image processing algorithms to achieve online monitoring and intelligent grading of impurities, bran, or dust agglomerates.

[0003] In existing technologies, the mainstream solution is to use a standard convolutional neural network classification model based on deep learning, such as the ResNet or MobileNet architecture. The basic working principle of this type of method is as follows: First, the collected images of the suction powder are input into a pre-trained backbone network to extract high-dimensional feature maps containing rich semantic information; then, a global average pooling layer is used to compress the feature maps with spatial dimensions into a single feature vector to achieve dimensionality reduction; finally, the feature vector is input into a fully connected layer for classification calculation to determine whether the current batch of suction powder is qualified.

[0004] However, this existing technology based on global average pooling has significant fundamental flaws when dealing with specific defects in suction powder. The main flaw is the dilution effect of the background signal on the defect features. The background signal refers to the area formed by countless tiny flour particles stacked together, while the defect features are generally divided into two categories: one is color defects such as charred particles or bran that are significantly darker than the background; the other is structural defects such as dust clumps that are similar in brightness to the background but have abnormal texture.

[0005] Because the global average pooling layer adopts an indiscriminate averaging strategy for all spatial locations of the feature map, for sparsely distributed color defects, their dark features will be neutralized by large areas of bright background features; for structural defects with similar brightness, their local texture fluctuation features will also be suppressed by the smooth background signal. This mechanism makes it difficult for the model to effectively separate weak defect signals from strong background noise, thus frequently misjudging unqualified products containing trace impurities or clumps as qualified, reducing the accuracy of the suction powder quality assessment results. Summary of the Invention

[0006] To address the problem that weak defect features are diluted by background signals, leading to a decrease in the accuracy of the air-suction powder quality assessment results, this invention proposes a deep learning-based method and system for air-suction powder image quality assessment.

[0007] In a first aspect, the present invention provides a deep learning-based method for evaluating the quality of suction powder images, comprising:

[0008] Obtain images of the suction powder to be evaluated;

[0009] For each pixel in the suction powder image, the multi-scale relative darkness and neighborhood gradient direction consistency of the pixel are obtained based on a multi-scale search strategy. The structural feature value of the pixel is determined based on the multi-scale relative darkness and neighborhood gradient direction consistency. The orthogonal spectral residual vector of the pixel in the RGB color space is calculated using the vector orthogonal projection method. The inverse weight is determined based on the brightness of the pixel and the average brightness of all pixels. The spectral heterogeneity of the pixel is obtained by multiplying the magnitude of the orthogonal spectral residual vector with the inverse weight.

[0010] The physical saliency of each pixel is determined by fusing its structural feature values ​​and spectral heterogeneity to generate a physical saliency matrix. A two-stream convolutional neural network containing a backbone feature extraction stream and a physical prior modulation stream is constructed. The image of the suction powder is input into the backbone feature extraction stream to extract intermediate layer feature data, and the physical saliency matrix is ​​input into the physical prior modulation stream. The physical saliency matrix is ​​mapped to a spatial attention weight matrix through the convolutional layer of the physical prior modulation stream. The spatial attention weight matrix is ​​used to perform affine transformation modulation on the intermediate layer feature data to generate physically enhanced high-dimensional semantic feature data. Global pooling and classification are performed on the high-dimensional semantic feature data to output the quality evaluation result of the suction powder.

[0011] This invention constructs a data-driven evaluation architecture guided by physical priors. It corrects the implicit attention bias of deep neural networks through explicit physical feature calculation. First, at the feature perception level, multi-scale relative darkness and orthogonal spectral residuals are used to decouple weak impurity signals in the suction powder from both geometric and spectral physical dimensions, constructing a high signal-to-noise ratio physical saliency matrix. Then, at the feature cognition level, a dual-stream feature modulation architecture is adopted. By mapping physical saliency to a spatial attention weight matrix and using affine transformation to force modulation of the intermediate layer features of the backbone network, this mechanism no longer relies on the network's random convergence but uses strong mathematical constraints to force convolutional kernels to focus on physically significant defect areas when extracting features and suppress large-area background responses. This enables accurate identification of bran, black spots, and homogeneous clumps in suction powder images, improving the accuracy of suction powder quality assessment results.

[0012] Preferably, the multi-scale relative darkness is determined as follows: For each preset scale, a neighborhood window for each pixel is divided at that scale, the average brightness of all pixels within the neighborhood window is calculated as the background brightness reference for that neighborhood window, the downward deviation of the pixel's brightness relative to the background brightness reference is calculated, the ratio of the downward deviation to the background brightness reference is determined as the relative darkness at that scale, and if the brightness of the pixel is not less than the background brightness reference, the relative darkness is directly set to zero; the maximum value of the relative darkness at all scales is selected as the multi-scale relative darkness of the pixel, and the neighborhood window corresponding to the maximum value is used as the adaptation neighborhood window of the pixel.

[0013] This technical solution determines relative darkness by the ratio of the degree of downward deviation of pixel brightness relative to the local background reference to the background brightness reference. Essentially, it extracts local contrast features, enabling the algorithm to automatically adapt to the overall drift of light source intensity. This ensures that as long as the impurity is darker than the surrounding background, it can be stably detected regardless of the overall brightness of the environment.

[0014] Preferably, the neighborhood gradient direction consistency is determined as follows: using the edge detection operator, the gradient vector of each pixel is calculated based on the brightness of each pixel. The magnitude of the vector sum of the gradient vectors of all pixels within the adapted neighborhood window of the pixel is taken as the synthetic gradient magnitude. The sum of the magnitudes of the gradient vectors of all pixels is taken as the absolute gradient sum. The ratio of the synthetic gradient magnitude to the absolute gradient sum is determined as the neighborhood gradient direction consistency of the pixel.

[0015] Preferably, the structural eigenvalues ​​are determined based on the following relationship:

[0016] ;

[0017] in, For the suction powder image located in The structural feature value of the pixel at that location, For located The gradient direction of the neighborhood of the pixel is consistent. The variance of the multi-scale relative darkness of all pixels within the adaptation neighborhood window for this pixel. For located The background brightness reference of the neighboring window for the pixel at that location. This represents the average brightness of all pixels in the image of the suction powder. Let V be the variance of the brightness of all pixels. This is a preset parameter to prevent the denominator from being zero. It is a natural exponential function.

[0018] This technical solution organically combines neighborhood gradient direction consistency, multi-scale dark variance, and inverse Gaussian gating coefficients to construct a strict AND gate logic. Only when three conditions are met simultaneously—directional edge, severe local dark fluctuation, and significant deviation from the background statistical distribution—will the structural feature value generate a high response. This multi-dimensional verification mechanism can accurately distinguish between dust agglomerates with specific physical structures and randomly distributed sensor noise.

[0019] Preferably, the orthogonal spectral residual vector of the pixel in the RGB color space is calculated using the orthogonal projection method. The inverse weight is determined based on the brightness of the pixel and the average brightness of all pixels. This includes: pre-setting a reference color vector in the RGB color space to characterize the pure background color of the air-absorbing powder; obtaining the color vector of each pixel based on its value in the RGB color space; projecting this color vector onto the reference color vector to obtain a projection vector; and calculating the difference vector between the color vector and the projection vector as the orthogonal spectral residual vector. The inverse weight satisfies the following relationship: ,in, For the suction powder image located in The brightness value of the pixel at that location. The average brightness of all pixels in the suction powder image. It is a natural exponential function.

[0020] This technical solution utilizes the principle of geometric projection to decompose the color of a pixel into a background principal axis component and an orthogonal residual component, thereby eliminating the interference of light intensity on color judgment. On this basis, the introduced inverse brightness weight is a nonlinear amplifier for dark targets, which gives extremely high weight gain to low brightness areas. This combination strategy enables the algorithm not only to accurately identify color impurities, but also to highlight tiny black spots that are not obvious in the conventional color space due to low brightness.

[0021] Preferably, physical salience is determined based on the following relationship:

[0022] ;in, For the suction powder image located in The physical salience of the pixel at that location. For the suction powder image located in The structural feature value of the pixel at that location, For the suction powder image located in Spectral heterogeneity of pixels at that location It is a natural exponential function.

[0023] This technical solution adopts a nonlinear fusion model based on inverse exponential mapping. The squared term first performs quadratic compression and suppression on low-amplitude background noise, widening the gap between signal and noise. The exponential function ensures that as long as any one of the structural features or spectral features shows high significance, the fusion result will quickly approach the saturation value, effectively avoiding missed detections caused by the lack of a single feature, while ensuring the numerical stability of the fusion result in the range of 0 to 1.

[0024] Preferably, after generating the physical saliency matrix, the method further includes: calculating the mean of the physical saliency of all pixels within the adaptation neighborhood window of each pixel, using it as a coherence factor, multiplying the physical saliency value of the pixel by the coherence factor to update the physical saliency of the pixel, and obtaining the updated physical saliency matrix.

[0025] This technical solution introduces neighborhood coherence modulation, which is essentially a verification mechanism based on spatial statistics. By calculating the neighborhood mean as the coherence factor, it requires that each high-response point must receive the votes of its surrounding neighborhood pixels. This mechanism achieves adaptive filtering of random speckle noise without losing edge details, thus improving the signal-to-noise ratio of saliency.

[0026] Preferably, the intermediate layer feature data is modulated using an affine transformation using a spatial attention weight matrix to generate physically enhanced high-dimensional semantic features. This includes: mapping the input physical saliency matrix to parameters used to construct the attention weight matrix via a lightweight convolutional network contained in the physical prior modulation stream: a scale scaling factor matrix and a bias factor matrix; and performing affine transformation modulation based on the following relationship: ;in, This is the high-dimensional semantic feature data output after modulation. The intermediate layer feature data is extracted from the main feature extraction stream. For element-wise multiplication, This is the scale scaling factor matrix contained in the spatial attention weight matrix. This is the bias factor matrix contained in the spatial attention weight matrix.

[0027] Preferably, the high-dimensional semantic feature data is globally pooled and classified to output the quality evaluation result of the suction powder, including: using a global average pooling layer to compress the spatial dimension of the high-dimensional semantic feature data, transforming the two-dimensional feature plane into a one-dimensional global feature vector, inputting the global feature vector into a fully connected layer for linear mapping, and using the Softmax activation function to calculate the probability distribution of the suction powder belonging to each preset quality level, and selecting the quality level corresponding to the maximum value in the probability distribution as the quality evaluation result.

[0028] Secondly, the present invention also provides a deep learning-based image quality assessment system for suction powder, the image quality assessment system for suction powder includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the suction powder image quality assessment methods.

[0029] The present invention has the following effects:

[0030] This invention overcomes the feature dilution effect caused by global average pooling in traditional deep learning by constructing a data-driven evaluation architecture guided by physical priors. It achieves enhancement and decoupling of defect features at the pixel level. By utilizing multi-scale relative darkness and gradient consistency, it extracts the geometric information of structural defects. At the same time, it extracts the chromaticity information of color defects by using vector orthogonal projection and dark area gain weights. Then, it constructs a high-confidence physical saliency matrix, which is transformed into spatial attention weights during deep learning. Affine transformation is used to force the modulation of the intermediate layer features of the backbone network. With strong mathematical constraints, the network is forced to focus on physically significant defect regions, effectively suppressing background noise response. This enables robust identification of weak, same-color, or different-color defects in complex particle backgrounds, improving the accuracy of the air-suction powder quality evaluation results. Attached Figure Description

[0031] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0033] Reference Figure 1 A deep learning-based method for evaluating the quality of suction powder images includes the following steps:

[0034] S1: Obtain the image of the suction powder to be evaluated.

[0035] In modern flour production systems, suction flour, as a key byproduct of the air separation system, has a complex composition. In the image compilation of suction flour, the background is composed of countless tiny flour particles stacked together, not a smooth plane, but a highly complex random texture of micro-particle stacking. This stacking structure results in the image being filled with microscopic gaps between particles and random texture fluctuations. The defects mixed in are often extremely small and have weak physical characteristics. For example, caramelized particles are extremely small and dark, and are easily confused with the particle gaps; the color of bran is very similar to that of flour, and is easily submerged in the stacked texture.

[0036] Therefore, the main components of suction powder include: ordinary suction powder, structural defects, and color defects. Ordinary suction powder is composed of randomly stacked tiny particles, characterized by high brightness and chaotic local texture gradient directions. Structural defects (such as bran and clumps) have brightness similar to ordinary suction powder, with relatively coherent geometric edges or clump structures, that is, good consistency in local texture gradient directions. Color defects (such as tiny charred particles) have very low brightness, no obvious structure, and different hues.

[0037] Therefore, the suction powder image to be evaluated mainly consists of three types of pixels: the vast majority of normal suction powder particles, and a very small number of structural defect particles, such as bran and dust clumps, and color defect particles, such as tiny coking particles.

[0038] When actually acquiring images, due to the non-Lambertian characteristics of the suction powder surface, the pixels of normal suction powder particles will produce microscopic shadows and random texture fluctuations when stacked. This high-frequency background noise can easily drown out weak defect signals. Traditional deep learning networks are prone to overfitting to the pixel features of a large number of normal suction powder particles under weak supervision, resulting in poor detection generalization ability.

[0039] This step takes into account that the brightness difference between the pixels of structural defect particles and normal suction powder particles is minimal under conventional monochromatic light, but the chromaticity difference can be captured under the full spectrum. Therefore, a high-resolution industrial linear scan camera with a high color rendering index ring LED light source is used to acquire real-time suction powder images above the closed powder flow channel. It is necessary to construct a dark box environment to shield external stray light and adjust the light source angle to reduce specular reflection in order to create a relatively stable optical physical environment. This ensures that the acquired images can reflect the physical properties of the three types of pixels to the greatest extent, providing a high-quality data foundation for subsequent feature decoupling.

[0040] The acquired RGB image is converted to the HSV color space to obtain the brightness, saturation, and chroma of each pixel. The average brightness of all pixels in the suction powder image is then obtained and denoted as . The variance of the brightness of all pixels is denoted as . Since the suction powder is essentially tiny particles of flour, and standard flour is white with a slight yellow tinge, we first need to obtain a pure image of flour free of any impurities. Then, we calculate the mean values ​​of the R, G, and B channels for all pixels in this pure flour image. Based on these three means, we construct a baseline color vector, denoted as [vector name missing]. .

[0041] S2: For each pixel in the suction powder image, obtain the multi-scale relative darkness and neighborhood gradient direction consistency of the pixel based on a multi-scale search strategy, and determine the structural feature value.

[0042] After acquiring the original image, the first step is to process the brightness fluctuations and texture noise of the normal suction powder particles in the suction powder image caused by random stacking. This complex background feature cannot be distinguished by directly using the brightness threshold from the "pixels of normal suction powder particles in the stacked shadow" and the "pixels corresponding to the real structural / color defect particles". Therefore, this step first starts from the geometric structure dimension to construct features independent of the background stacking state.

[0043] Therefore, this step first starts from the geometric dimension to construct structural features independent of background brightness fluctuations in order to solve the interference problem caused by background particle stacking.

[0044] S21: Determine the relative darkness at multiple scales.

[0045] The pixels of structural / color defective particles are physically darker than the pixels of their adjacent normal suction powder particles. However, due to the unevenness of the stacking of suction powder particles, the particle density and stacking thickness are different in different areas of the suction powder image, resulting in uneven overall brightness distribution of the image, with thicker stacked areas appearing darker overall.

[0046] Therefore, for each preset scale ( , , to ), divide the neighborhood window of each pixel, in order to Taking scale as an example, with each pixel as the center, the surrounding area... The range formed by a number of pixels is used as the range of that pixel. For a neighborhood window at a given scale, the neighborhood window contains a total of 9 pixels. The average brightness of these 9 pixels is used as the background brightness reference for the neighborhood window. The degree of downward deviation of the brightness of the pixel (located at the center of the neighborhood window) relative to the background brightness reference is calculated. The ratio of the downward deviation to the background brightness reference is determined as the relative darkness at this scale. If the brightness of the pixel is not less than the background brightness reference, the relative darkness is directly set to zero.

[0047] Specifically, it satisfies the following relationship:

[0048] Similarly, obtain the pixel value in sequence. , to The neighborhood window at each scale is calculated, and the corresponding relative darkness is obtained. Finally, the maximum value of the relative darkness at all scales is selected as the multi-scale relative darkness of the pixel, and the neighborhood window corresponding to the maximum value is used as the adaptation neighborhood window of the pixel.

[0049] Through this operation, the calculated multi-scale relative darkness is essentially a variant of local Weber contrast. By calculating the ratio, the algorithm removes the influence of the local stacking thickness of normal suction powder particles on the brightness benchmark. Regardless of the background brightness, as long as a pixel exhibits a darker characteristic relative to the surrounding normal suction powder particles, the pixel is more likely to be an impurity particle. This strategy effectively avoids misjudging the natural shadows between particles as impurities.

[0050] S22: Determine the consistency of gradient directions in the neighborhood.

[0051] The background of the suction powder is composed of a large number of discrete normal suction powder particles stacked together. Its gradient direction at the microscopic level is random (isotropic), that is, in any tiny region, the gradient direction of normal flour particles is randomly diverging, and they will cancel each other out when vectors are superimposed. Conversely, the pixels of structural defect particles (such as bran and lumps) usually have relatively coherent physical edges or clumping structures, and their gradient direction has local consistency, which will produce an enhancement effect when vectors are superimposed.

[0052] Therefore, the edge detection operator is used to calculate the gradient vector of each pixel based on the brightness of each pixel. The magnitude of the vector sum of the gradient vectors of all pixels within the adaptation neighborhood window of the pixel is used as the synthetic gradient magnitude. The sum of the magnitudes of the gradient vectors of all pixels is used as the absolute gradient sum. The ratio of the synthetic gradient magnitude to the absolute gradient sum is determined as the neighborhood gradient direction consistency of the pixel.

[0053] Specifically, it satisfies the following relationship:

[0054] Through this operation, the metric acts as a texture entropy filter. For pixels with normal suction powder particles, the value approaches 0; for pixels with structural defect particles, the value approaches 1. This effectively filters out a large amount of interference from irrelevant background (normal suction powder particles) without losing subtle edge details.

[0055] S23: Determine the structural characteristic values.

[0056] To accurately distinguish the pixels of structurally defective particles from those of normal suction powder particles, a rigorous multi-dimensional verification mechanism is needed. Therefore, structural feature values... Determined based on the following relation:

[0057]

[0058] in, For the suction powder image located in The structural feature value of the pixel at that location, For located The gradient direction of the neighborhood of the pixel is consistent. The variance of the multi-scale relative darkness of all pixels within the adaptation neighborhood window for this pixel. For located The background brightness reference of the neighboring window for the pixel at that location. This represents the average brightness of all pixels in the image of the suction powder. Let V be the variance of the brightness of all pixels. The parameter, which is preset to prevent the denominator from being zero, is usually set to a very small positive number. , It is a natural exponential function.

[0059] Through this operation, the formula constructs a rigorous three-dimensional physical feature product gating model, where each term is a logic switch: As the first directional gate, if the gradient around a pixel is messy and the pixel features match those of a normal suction powder particle, this term approaches 0 and the output is blocked. As an activity gating mechanism, the variance of multi-scale relative darkness reflects the intensity of fluctuations in local features. For pixels with normal suction powder particles, the variance of multi-scale relative darkness is extremely low, suppressing output; inverse Gaussian term As a statistical outlier gate, it measures the degree of deviation between the local brightness of the current pixel and the global brightness of the pixel of a normal suction powder particle.

[0060] A pixel is more likely to be a structural defect particle only when it simultaneously meets three physical conditions: having a directional edge, experiencing severe local brightness fluctuations, and significantly deviating from the global background distribution.

[0061] S3: Calculate the orthogonal spectral residual vector of the pixel in the RGB color space using the orthogonal projection method. Determine the inverse weight based on the brightness of the pixel and the average brightness of all pixels. Multiply the magnitude of the orthogonal spectral residual vector by the inverse weight to obtain the spectral heterogeneity of the pixel.

[0062] After effectively extracting the pixels of structural defect particles with obvious geometric structures, it is further considered that the physical features of color defect particles (such as the pixels corresponding to fine coke powder particles) may be overwhelmed by the complex texture of normal suction powder particles. Therefore, it is necessary to further distinguish them from other dimensions.

[0063] This step takes into account that the core feature of the pixels of color-type defective particles is that their spectral composition is abnormal compared with that of normal suction powder particles. Therefore, this step further decouples the features from the spectral physical dimension in order to accurately identify the pixels of color-type defective particles.

[0064] S31: Calculate the orthogonal spectral residual vector.

[0065] In the RGB space, the brightness of pixels in normal suction powder particles will vary due to different stacking angles, mainly reflected in the scaling of the color vector, which is a geometric interference; while the brightness of pixels in color defective particles is reflected in the directional deflection of the color vector.

[0066] Therefore, a reference vector representing the color of a pixel in normal suction powder particles is predefined. The color vector of each pixel is obtained based on its value in the RGB color space. This pixel's color vector is then projected onto the reference vector, and the difference between the original and projected vectors is calculated, resulting in the orthogonal spectral residual vector. In practice, this is decomposed using the Gram-Schmidt orthogonalization principle from linear algebra.

[0067] This operation uses orthogonal projection to decompose pixel color into a background homogeneous component (projected part) and an orthogonal heterogeneous component (residual part). The residual vector is perpendicular to the background main color and only represents the pure color difference between the pixel and the normal suction powder particles, thus decoupling the brightness interference caused by particle stacking.

[0068] S32: Determine the inverse weighting and spectral heterogeneity.

[0069] The pixels of color-defective particles are usually darker than the pixels of normal suction powder particles. Therefore, based on the difference between the brightness of the pixel and the average brightness of all pixels, the inverse weight of the pixel is determined, and the magnitude of the orthogonal spectral residual vector is weighted by the inverse weight to determine the spectral heterogeneity of the pixel.

[0070] Specifically, spectral heterogeneity satisfies the following relationship:

[0071]

[0072] in, For the suction powder image located in Spectral heterogeneity of pixels at that location For the suction powder image located in The orthogonal spectral residual vector of the pixel at that location in the RGB color space. To determine the sign of the modulus, Partially represents the inverse weight of that pixel. For the suction powder image located in The brightness value of the pixel at that location. The average brightness of all pixels in the suction powder image. It is a natural exponential function.

[0073] This relation solves the challenge of color defect identification through a dual mechanism of geometric decoupling and adaptive gain. First, the vector orthogonal projection method uses linear algebra principles to decompose pixel color into background homogeneous components and orthogonal residual components. The orthogonal residual vector eliminates interference from changes in illumination intensity, extracting pure chromaticity differences. Then, the inverse weight, acting as an adaptive amplifier based on statistical differences in brightness, is used in conjunction with… The modulus is linked, utilizing the non-linear saturation characteristic of the Sigmoid function, when the brightness of a pixel... Below the global average That is, when exhibiting dark characteristics, the exponential term As the denominator rapidly decreases, the inverse weight approaches its maximum value, thus preserving or amplifying the chromaticity signal of dark impurities. Conversely, for background areas with brightness higher than the mean, the weight rapidly decays to near zero, thereby suppressing noise interference in bright backgrounds. This design utilizes the statistical distribution characteristics of pixel brightness to automatically enhance the visual saliency of dark impurities in light backgrounds. It prioritizes amplifying pixels that have both color difference (larger modulus of orthogonal spectral residual vector) and are relatively dark (higher inverse weight), i.e., pixels of color-type defect particles, making them significantly brightened in the feature map.

[0074] S4: Combine the structural feature values ​​and spectral heterogeneity of each pixel to determine the physical saliency of that pixel, and generate a physical saliency matrix.

[0075] Structural feature values ​​of pixels representing structural defect particles and spectral heterogeneity of pixels representing color defect particles were obtained. These two features are two independent and complementary sources of evidence describing the quality of the suction powder. However, these two features belong to completely different physical dimensions, with significant differences in their numerical magnitude and dynamic range. If they are directly fused mathematically, the feature with the larger value will dominate the fusion result, causing the feature of the other dimension to become invalid. Therefore, this step must first introduce a statistically based adaptive normalization mechanism to align the features, then perform multimodal fusion, and utilize spatial context information for self-correction.

[0076] S41: Percentile-based statistical normalization.

[0077] Traditional maximum value normalization is easily affected by individual extreme noise points, leading to excessive compression of the overall effective feature values. The number of such bad pixels in the image may far exceed the actual defects. Therefore, the structural feature values ​​and spectral heterogeneity of all pixels are statistically analyzed, sorted in ascending order, and the 95th percentile of each sorted pixel is selected using the percentile method. These 95th percentiles are used as the first and second normalization feature scales, respectively. The structural feature values ​​and spectral heterogeneity of each pixel are then normalized by division using both the first and second normalization feature scales, yielding statistically normalized values ​​for the pixel's structural feature values ​​and spectral heterogeneity.

[0078] This operation, which uses the percentile method to determine the normalized denominator, is essentially about finding the statistical upper limit of the effective signal. This method has extremely strong statistical robustness. It eliminates the interference of extreme outliers and normalizes the feature indices of the pixels of structural defect particles and the pixels of color defect particles to the same dimensional range.

[0079] S42: Nonlinear fusion of physical saliency.

[0080] To achieve complementarity and enhancement of multimodal features and ensure detection even when a single feature is missing, the physical saliency of a pixel is determined based on the following relationship:

[0081]

[0082] in, For the suction powder image located in The physical salience of the pixel at that location. For the suction powder image located in The structural feature value of the pixel at that location, For the suction powder image located in Spectral heterogeneity of pixels at that location It is a natural exponential function.

[0083] Through this operation, the fusion mechanism adopts the form of the sum of squares of energy. Matching saturation function It not only achieves multimodal complementarity, enabling the detection of both same-color clumps and dissimilar impurity particles, but also exhibits nonlinear signal-to-noise ratio enhancement for pixels containing normal suction powder particles. and Both are very low, because the square effect is further compressed and suppressed, leading to Very small; for pixels with structural defect particles, it appears... Very tall The very low performance characteristic is evident in pixels with color-related defects. and These highly characteristic features were quickly brought to saturation. It's very big.

[0084] S43: Update based on neighborhood coherence.

[0085] A high response at a single point may be a sensor defect or transient noise, while real physical defects usually occupy a certain spatial area. Therefore, the mean of the physical saliency of all pixels within the adaptation neighborhood window of a pixel is calculated as a coherence factor. The coherence factor is multiplied by the physical saliency of the pixel to achieve coherent modulation operation and obtain the updated physical saliency. Based on the updated physical saliency of all pixels, a physical saliency matrix is ​​constructed.

[0086] This operation introduces spatial coherence constraints, eliminating isolated noise (usually sensor noise) and retaining pixels representing clusters of structural / color defect particles. Isolated noise is eliminated due to its low neighborhood mean, while real physical defects are enhanced by mutual support within their neighborhoods, thus generating a high-confidence physical saliency matrix.

[0087] S5: Construct a two-stream convolutional neural network containing a backbone feature extraction stream and a physical prior modulation stream. Generate physically enhanced high-dimensional semantic feature data based on the physical saliency matrix. Perform global pooling and classification on the high-dimensional semantic feature data and output the quality evaluation results of the suction powder.

[0088] The elements in the physical saliency matrix reflect the probability that each pixel belongs to a structural / color defect particle. When traditional CNNs process images of suction powder, the pixel features of sparse structural / color defect particles are diluted by the massive pixel features of normal suction powder particles due to the large number of pixels and complex textures of normal suction powder particles.

[0089] Therefore, this step utilizes physical saliency as a strong prior and performs forced modulation before pooling. Since deep learning must be introduced for semantic analysis, this step uses the physical saliency of the pixel as strong prior knowledge and forces it into the neural network through a two-stream architecture, modulating and enhancing it before the feature is pooled.

[0090] Specifically, the image of the suction powder is input into the backbone feature extraction stream to extract intermediate layer feature data, and the physical saliency matrix is ​​input into the physical prior modulation stream. The convolutional layer of the physical prior modulation stream maps the physical saliency matrix into two parameters that constitute the attention weight matrix, one of which is the scale factor. The other is the bias factor matrix. Then, the spatial attention weight matrix is ​​used to perform affine transformation modulation on the intermediate layer feature data to generate physically enhanced high-dimensional semantic feature data. The high-dimensional semantic feature data is then globally pooled and classified to output the quality evaluation results of the suction powder.

[0091] S51: Affine transformation modulation.

[0092] Under weakly supervised training with only image labels (e.g., pass / fail), conventional CNNs tend to learn easily identifiable background statistical features while ignoring subtle local defects. Therefore, a physical prior modulation flow is used to map the saliency matrix to a scale factor. and bias factor The features are modulated using a formula:

[0093]

[0094] in, This is the high-dimensional semantic feature data output after modulation. The intermediate layer feature data is extracted from the main feature extraction stream. For element-wise multiplication, This is the scale scaling factor matrix contained in the spatial attention weight matrix. This is the bias factor matrix contained in the spatial attention weight matrix.

[0095] This operation allows us to determine the location of pixels representing structural defects and pixels representing color defects. A positive gain term forces the feature response to be amplified, while in large areas of normal suction powder particles, the feature response is suppressed. The bias term provides a physical offset to the feature baseline, which helps to recover the spatial information of subtle defects that may be lost during deep convolution and downsampling.

[0096] To ensure dimensional alignment during affine transformation modulation operations, the intermediate layer feature data extracted by the backbone feature extraction stream... The dimension is denoted as ,in, For the number of channels, and These represent the height and width of the feature map, respectively. Accordingly, the lightweight convolutional network in the physics-prior modulation stream is configured to map the single-channel physics saliency matrix to a matrix with the same number of channels. The output, i.e., the generated scale scaling factor matrix. and bias factor matrix All dimensions are In practice, the number of convolutional kernels in the terminal convolutional layers of the lightweight convolutional network is set to be equal to the feature data of the intermediate layers. Number of channels This makes the generated attention weight matrix equal to the channel dimension. This ensures consistency, thereby supporting subsequent element-wise multiplication and addition modulation operations on all channels.

[0097] This modulation mechanism no longer relies on the network's own random convergence, but instead forces the convolutional kernel to focus on physically significant defect regions when extracting features through strong mathematical constraints. This directly counteracts the dilution effect of global average pooling: even if the defect area is small, its feature value is greatly enhanced by physical priors, so it can occupy enough weight in the final pooling operation and be accurately identified by the classifier.

[0098] S52: Global Pooling and Classification.

[0099] To achieve spatial dimension compression and output the final classification result, global average pooling (GAP) and classification are performed on the modulated feature data. Specifically, a global average pooling layer is used to compress the spatial dimension of the high-dimensional semantic feature data, transforming the two-dimensional feature plane containing spatial information into a one-dimensional global feature vector. Subsequently, this global feature vector is input into a fully connected layer for linear mapping, mapping the high-dimensional feature space to the category space. Finally, the Softmax activation function is used to calculate the probability distribution of the air-absorbing powder belonging to each preset quality level (e.g., excellent, first-class, second-class, unqualified), and the quality level corresponding to the maximum value in the probability distribution is selected as the final quality evaluation result.

[0100] Since the feature map has already undergone physical prior denoising and enhancement before pooling, the GAP operation aggregates the semantic information of the pixels corresponding to high-purity structural defect particles and color defect particles, rather than the noise average of the pixels of normal suction powder particles, thereby improving the robustness and accuracy of the classifier in complex suction powder scenarios.

[0101] In summary, this step achieves accurate classification of defect types by constructing a collaborative processing mechanism that integrates spatial location information and category semantic representation. The physical prior modulation stream uses a single-channel physical saliency matrix as a spatial attention mask to accurately locate the spatial distribution region of defects in the powder suction image, while the backbone feature extraction stream is responsible for extracting high-dimensional feature vectors containing hundreds of channels from the powder suction image. These channels have encoded rich semantic information representing defect category attributes. For example, a specific channel responds to the chromaticity features of dark coking particles, while another channel responds to the texture features of dust agglomerates. Through an affine transformation formula, the weights... Instead of simple overlay, it acts as a multiplier to spatially weight the original high-dimensional features. This amplifies the spectral heterogeneity of coking particles or the geometrical anomalies of dust agglomerations, while effectively suppressing background noise from normal suction particles. Therefore, the features ultimately input into the classifier... It is high signal-to-noise ratio semantic information after physical prior denoising optimization. The classifier no longer processes data mixed with a large amount of background noise, but rather feature vectors that have been significantly enhanced and have clear category orientation. This is the advantage of the deep combination of the spatial attention mechanism guided by physical prior and the high-dimensional feature representation of deep neural networks, ensuring that the model can not only detect defects, but also identify the specific category of defects with extremely high accuracy.

[0102] This application also provides a deep learning-based image quality assessment system, including a processor and a memory. The processor is preferably a GPU that supports CUDA acceleration. The memory is used to cache high-resolution images and network model parameters. The system achieves fully automated assessment from optical imaging to physical feature decoupling and then to deep semantic cognition by executing the method flow from step S1 to step S5.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for evaluating the quality of suction powder images, characterized in that, include: Obtain images of the suction powder to be evaluated; For each pixel in the suction powder image, the multi-scale relative darkness and neighborhood gradient direction consistency of the pixel are obtained based on a multi-scale search strategy. The structural feature value of the pixel is determined based on the multi-scale relative darkness and neighborhood gradient direction consistency. The orthogonal spectral residual vector of the pixel in the RGB color space is calculated using the vector orthogonal projection method. The inverse weight is determined based on the brightness of the pixel and the average brightness of all pixels. The spectral heterogeneity of the pixel is obtained by multiplying the magnitude of the orthogonal spectral residual vector with the inverse weight. The physical saliency of each pixel is determined by fusing its structural feature values ​​and spectral heterogeneity to generate a physical saliency matrix. A two-stream convolutional neural network containing a backbone feature extraction stream and a physical prior modulation stream is constructed. The image of the suction powder is input into the backbone feature extraction stream to extract intermediate layer feature data, and the physical saliency matrix is ​​input into the physical prior modulation stream. The physical saliency matrix is ​​mapped to a spatial attention weight matrix through the convolutional layer of the physical prior modulation stream. The spatial attention weight matrix is ​​used to perform affine transformation modulation on the intermediate layer feature data to generate physically enhanced high-dimensional semantic feature data. Global pooling and classification are performed on the high-dimensional semantic feature data to output the quality evaluation result of the suction powder.

2. The method for evaluating the image quality of suction powder according to claim 1, characterized in that, Multi-scale relative darkness is determined based on the following method: For each preset scale, a neighborhood window is divided for each pixel at that scale. The average brightness of all pixels in the neighborhood window is calculated as the background brightness reference for the neighborhood window. The degree of downward deviation of the brightness of the pixel relative to the background brightness reference is calculated. The ratio of the degree of downward deviation to the background brightness reference is determined as the relative darkness at that scale. If the brightness of the pixel is not less than the background brightness reference, the relative darkness is directly set to zero. The maximum value of the relative darkness across all scales is selected as the multi-scale relative darkness of the pixel, and the neighborhood window corresponding to the maximum value is used as the adaptation neighborhood window of the pixel.

3. The method for evaluating the image quality of suction powder according to claim 2, characterized in that, Neighborhood gradient direction consistency is determined based on the following method: Using the edge detection operator, the gradient vector of each pixel is calculated based on the brightness of each pixel. The magnitude of the vector sum of the gradient vectors of all pixels within the adaptive neighborhood window of that pixel is taken as the synthetic gradient magnitude. The sum of the magnitudes of the gradient vectors of all pixels is taken as the absolute gradient sum. The ratio of the synthetic gradient magnitude to the absolute gradient sum is determined as the neighborhood gradient direction consistency of that pixel.

4. The method for evaluating the image quality of suction powder according to claim 2, characterized in that, Structural eigenvalues ​​are determined based on the following relationship: ; For the suction powder image located in The structural feature value of the pixel at that location, For located The gradient direction of the neighborhood of the pixel is consistent. The variance of the multi-scale relative darkness of all pixels within the adaptation neighborhood window for this pixel. For located The background brightness reference of the neighboring window for the pixel at that location. This represents the average brightness of all pixels in the image of the suction powder. Let V be the variance of the brightness of all pixels. This is a preset parameter to prevent the denominator from being zero. It is a natural exponential function.

5. The method for evaluating the image quality of suction powder according to claim 1, characterized in that, The orthogonal spectral residual vector of the pixel in the RGB color space is calculated using the orthogonal projection method. The inverse weight is determined based on the brightness of the pixel and the average brightness of all pixels. This process includes: pre-setting a reference color vector in the RGB color space to represent the pure background color of the air-absorbing powder; obtaining the color vector of each pixel based on its RGB color space value; projecting this color vector onto the reference color vector to obtain a projection vector; and calculating the difference between the color vector and the projection vector as the orthogonal spectral residual vector. The inverse weight satisfies the following relationship: ,in, For the suction powder image located in The brightness value of the pixel at that location. The average brightness of all pixels in the suction powder image. It is a natural exponential function.

6. The method for evaluating the image quality of suction powder according to claim 1, characterized in that, Physical saliency is determined based on the following relationship: ;in, For the suction powder image located in The physical salience of the pixel at that location. For the suction powder image located in The structural feature value of the pixel at that location, For the suction powder image located in Spectral heterogeneity of pixels at that location, It is a natural exponential function.

7. The method for evaluating the image quality of suction powder according to claim 2, characterized in that, After generating the physical saliency matrix, the following steps are also included: Calculate the mean of the physical saliency of all pixels within the adaptation neighborhood window of each pixel, and use it as a coherence factor. Multiply the physical saliency value of the pixel by the coherence factor to update the physical saliency of the pixel, and obtain the updated physical saliency matrix.

8. The method for evaluating the image quality of suction powder according to claim 1, characterized in that, Affine transformation modulation is applied to the intermediate layer feature data using a spatial attention weight matrix to generate physically enhanced high-dimensional semantic features, including: The input physical saliency matrix is ​​mapped to parameters used to construct the attention weight matrix through a lightweight convolutional network contained in the physical prior modulation stream: a scale factor matrix and a bias factor matrix. Affine transformation modulation is performed based on the following relationship: ;in, This is the high-dimensional semantic feature data output after modulation. The intermediate layer feature data is extracted from the main feature extraction stream. For element-wise multiplication, This is the scale scaling factor matrix contained in the spatial attention weight matrix. This is the bias factor matrix contained in the spatial attention weight matrix.

9. The method for evaluating the image quality of suction powder according to claim 1, characterized in that, Global pooling and classification are performed on high-dimensional semantic feature data to output the quality evaluation results of the suction powder, including: The high-dimensional semantic feature data is compressed spatially by using a global average pooling layer, which transforms the two-dimensional feature plane into a one-dimensional global feature vector. The global feature vector is then input into a fully connected layer for linear mapping. The Softmax activation function is used to calculate the probability distribution of the suction powder belonging to each preset quality level. The quality level corresponding to the maximum value in the probability distribution is selected as the quality evaluation result.

10. A deep learning-based image quality assessment system for suction powder, characterized in that, The suction powder image quality assessment system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the suction powder image quality assessment method as described in any one of claims 1-9.

Images