A method for identifying dust composition in electronic dismantling plants based on polarization scattering imaging

Abstract

This invention discloses a method for identifying dust components in electronic dismantling plants based on polarization scattering imaging. The method includes: acquiring multi-angle polarization scattering images of dust areas in the electronic dismantling plant using a UAV system equipped with a multi-angle polarization imaging module; calculating Stokes vector parameters based on polarization optics theory to extract the linear polarization degree matrix and polarization angle matrix of the dust scattered light; constructing a multi-dimensional polarization scattering fingerprint feature, fusing statistical features, distribution features, and contrast features to form a polarization scattering fingerprint feature vector; inputting the polarization scattering fingerprint feature vector into a pre-trained deep neural network for classification calculation, outputting the probability distribution of various dust components; and applying a decision function based on Bayesian posterior probability calculation to output the final identification result of the main dust components in the electronic dismantling plant. This invention can provide reliable technical support for environmental monitoring, resource recycling, and safety protection in electronic dismantling plants.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method for identifying the composition of dust in an electronics dismantling plant based on polarization scattering imaging. Background Technology

[0002] With the ever-accelerating pace of global electronic product upgrades, the treatment and recycling of electronic waste has become a crucial issue for environmental protection and resource recycling. Electronic dismantling plants, as a core link in electronic waste treatment, generate large amounts of dust containing various metal components and composite materials during the dismantling process. This dust not only poses a threat to the health of workers but is also a valuable recyclable resource. Accurately identifying dust composition is crucial for optimizing dismantling processes, improving resource recycling efficiency, and reducing environmental pollution risks. However, the composition of dust from electronic dismantling plants is complex and diverse, including pure metal dust such as copper and aluminum powder, non-metallic dust such as plastic dust, and various mixed metal and composite material dusts. This complexity presents a significant challenge to accurate identification.

[0003] Traditional dust composition identification methods mainly rely on chemical analysis and physical detection techniques, such as X-ray fluorescence spectroscopy and inductively coupled plasma mass spectrometry. While these methods offer high detection accuracy, they suffer from drawbacks such as long detection cycles, high costs, and the need for specialized personnel, making them unsuitable for real-time monitoring in industrial settings. Furthermore, these methods typically require offline sample collection and analysis, hindering continuous monitoring and spatial analysis of dust distribution and limiting their effectiveness in dynamic environments.

[0004] Optical detection methods developed in recent years, such as color feature-based image recognition technology and scattering spectrum analysis methods, have improved detection efficiency to some extent, but still have significant shortcomings. Color feature-based methods are easily affected by changes in lighting conditions and dust surface contamination, making it difficult to accurately distinguish between different metallic dusts with similar colors. While scattering spectrum-based methods can provide some compositional information, their accuracy in identifying complex dust mixtures is limited, especially when multiple dust components are present, as spectral signals often interfere with each other, leading to unreliable identification results.

[0005] Existing machine vision methods mostly employ traditional grayscale or color images for feature extraction. These methods primarily rely on superficial features such as shape, texture, and color, lacking a deep description of the inherent physical properties of dust materials. Since dust from different materials may exhibit similarities in appearance, especially oxidized or contaminated metal dust, traditional image features often fail to provide sufficient discriminative information. Furthermore, complex lighting conditions, variations in dust concentration, and background interference in industrial environments further reduce the reliability and stability of recognition methods based on traditional image features.

[0006] Polarization imaging, as an emerging optical detection method, can acquire the polarization information of target objects and reveal the intrinsic optical properties of materials, showing great potential in target detection and material identification. However, existing applications of polarization imaging are mainly concentrated in remote sensing, biomedicine, and military reconnaissance, while research and application in industrial dust identification are still in their infancy. In particular, methods for analyzing polarization scattering characteristics and identifying components in the complex dust environment of electronic dismantling plants have not yet been established, lacking a systematic theoretical foundation and practical technical solutions. Summary of the Invention

[0007] In view of this, the present invention provides a method for identifying dust components in electronic dismantling plants based on polarization scattering imaging. The purpose is to establish a polarization scattering imaging system, acquire multi-angle polarization images, extract the polarization scattering characteristics of dust based on polarization optics theory, construct a multi-dimensional polarization scattering fingerprint feature vector, and use a deep neural network to achieve accurate classification and identification of dust components, thereby providing reliable technical support for environmental monitoring, resource recycling, and safety protection in electronic dismantling plants.

[0008] To achieve the above objectives, the present invention provides a method for identifying the composition of dust in an electronics dismantling plant based on polarization scattering imaging, comprising the following steps: S1: Obtain multi-angle polarization scattering images of the dust area in an electronics dismantling plant using a drone system equipped with a multi-angle polarization imaging module; S2: Calculate the Stokes vector parameters based on polarization optics theory, and extract the linear polarization degree matrix and polarization angle matrix of dust scattered light; S3: Construct a multi-dimensional polarization scattering fingerprint feature, and fuse statistical features, distribution features and contrast features to form a polarization scattering fingerprint feature vector; S4: Input the polarization scattering fingerprint feature vector into a pre-trained deep neural network for classification calculation, and output the probability distribution of various dust components; S5: Based on Bayesian posterior probability calculation, the decision function is applied to output the final identification result of the main components of dust from the electronic dismantling plant.

[0009] As a further improvement of the present invention: Optionally, in step S1, a multi-angle polarization scattering image of the dust area in the electronics dismantling plant is acquired using a drone system equipped with a multi-angle polarization imaging module, including: A drone equipped with a multi-angle polarization imaging module is controlled to fly over the dusty area of ​​the target electronics dismantling plant and activate the multi-angle polarization imaging module. The module contains multiple linear polarizers aligned with the camera sensor and, through a synchronization mechanism, continuously captures images of the same dusty area in four linear polarization directions: 0°, 45°, 90°, and 135°. It captures the scattered light information of the dust particles on the incident light, which includes sunlight or an active light source. The images captured in the four polarization directions are then temporally and spatially registered to obtain a set of grayscale or color original polarization images corresponding to the 0°, 45°, 90°, and 135° polarization directions, denoted as [image 1]. , , , .

[0010] Optionally, in step S2, the Stokes vector parameters are calculated based on polarization optics theory to extract the linear polarization degree matrix and polarization angle matrix of the dust scattered light, including: Based on polarization optics theory, pixel-level calculations are performed on the four original input polarization images to extract the core polarization parameters reflecting the physical properties of dust; firstly, the first three components of the Stokes vector are calculated. , , The calculation formula is: ; ; ; in, This represents the first component of the Stokes vector, which represents the total light intensity. The second component of the Stokes vector represents the polarization difference between the 0° and 90° directions. The third component of the Stokes vector represents the polarization difference between the 45° and 135° directions; The degree of linear polarization and polarization angle are further calculated using the Stokes vector components. The calculation formula is as follows: ; ; in, Indicates the degree of linear polarization, used to quantify the degree of polarization of scattered light. Indicates the polarization angle, used to describe the principal polarization direction of the scattered light. This represents the arctangent function; the calculation process is performed pixel by pixel to obtain a linear polarization degree matrix and a polarization angle matrix with the same size as the original image.

[0011] This step establishes a complete Stokes vector parameter system. By calculating the first three components, the polarization state of the scattered light is comprehensively described. The total light intensity reflects the overall intensity characteristics of dust scattering, while the two polarization difference components capture polarization information in different directions. This multi-dimensional parameter extraction method ensures a comprehensive characterization of the dust polarization characteristics.

[0012] Optionally, in step S3, a multi-dimensional polarization scattering fingerprint feature is constructed, fusing statistical features, distribution features, and contrast features to form a polarization scattering fingerprint feature vector, including: Statistical features are extracted from the effective dust region in the linear polarization degree matrix and polarization angle matrix, and the first-order statistical moments are calculated as statistical features; the first-order statistical moments include the mean and variance, specifically: ; ; ; ; in, and These represent the mean values ​​of the degree of linear polarization and the polarization angle, respectively. and These represent the variances of the degree of linear polarization and the polarization angle, respectively. This represents the total number of pixels in the effective dust area. and They represent the first The linear polarization degree and polarization angle values ​​of each pixel; The distribution features of the polarization angle matrix are extracted, and a polarization angle distribution vector is generated through histogram analysis, dividing the polarization angle range into... Calculate the percentage of pixels in each equally spaced interval to form a... 3D distribution eigenvector The calculation formula is: ; in, The first vector representing the polarization angle distribution vector One portion, Indicates falling on the 1st The number of pixels within each angle range .

[0013] The contrast features are extracted by applying the gray-level co-occurrence matrix algorithm to the linear polarization degree matrix. Specifically, the contrast features are: ; in, Indicates texture contrast, used to characterize the spatial inhomogeneity of polarization scattering. Represents the number of gray levels. Represents the gray values ​​in the gray-level co-occurrence matrix. and The probability of coexistence; Statistical features, distribution features, and texture features are concatenated and fused to form a polarization scattering fingerprint feature vector. .

[0014] The statistical feature extraction method used in this step can accurately capture the central tendency and dispersion of linear polarization degree and polarization angle in spatial distribution. By calculating the two basic statistical moments of mean and variance, it not only reflects the average polarization level of dust scattering, but more importantly, it reveals the spatial variability of polarization parameters. This variability is directly related to the size distribution, shape complexity and material uniformity of dust particles, providing an important basis for distinguishing different types of dust.

[0015] The distribution feature extraction strategy introduced in this step effectively captures the angular distribution pattern of dust scattering through polarization angle histogram analysis. This interval-based statistical method can meticulously describe the probability distribution characteristics of polarization angles, reflecting the dominant orientation and dispersion degree of dust particles. Different dust components exhibit unique patterns in their polarization angle distribution due to differences in their physical structure and optical properties. This distribution characteristic provides strong feature support for dust type identification, especially for dust particles with regular shapes or specific orientations, whose polarization angle distribution often exhibits obvious peak characteristics.

[0016] Optionally, in step S4, the polarization scattering fingerprint feature vector is input into a pre-trained deep neural network for classification calculation, outputting the probability distribution of various dust components, including: Polarization scattering fingerprint feature vector The data is input into a pre-trained deep neural network; the deep neural network contains fully connected layers and nonlinear activation functions; the deep neural network establishes a nonlinear mapping relationship from fingerprint features to dust material by learning a large amount of labeled fingerprint feature data; The input feature vector undergoes forward propagation computation in the deep neural network, obtaining the final output through layer-by-layer linear transformations and non-linear activations. The output layer of the deep neural network uses the softmax activation function to transform the original output of the deep neural network into a probability distribution, calculated as follows: ; in, This indicates that the input sample belongs to the first... The probability of dust-like components. and These represent the output layers of the network, respectively. The and the first The original output value of each neuron. This indicates the total number of dust component categories. , , Represents the base of the natural logarithm; obtains an output vector containing the probability distributions of various dust components.

[0017] Optionally, in step S5, a decision function is applied based on Bayesian posterior probability calculation to output the final identification result of the main components of dust from the electronic dismantling plant, including: Based on Bayesian theory, the probability distribution of the deep neural network output is corrected using posterior probability, and prior knowledge is combined to improve recognition accuracy. The prior probability distribution of various dust types is determined based on historical statistical data of dust composition from electronic dismantling plants. Indicates the first The probability of dust-like components appearing in the environment of an electronic dismantling plant; Using Bayes' theorem to calculate the posterior probability, the conditional probability output by the deep neural network is taken as the likelihood probability. The calculation formula is as follows: ; in, This indicates that the observed polarization scattering fingerprint feature vector Under these conditions, dust belongs to the first category. Posterior probability of class Indicates the first Observed under dust-like conditions The likelihood probability is equal to the output probability of the deep neural network. , Indicates the first Prior probability of dust particles, express The marginal probability; The marginal probability Calculated using the law of total probability: ; in, Indicates the first Observation of eigenvectors under dust-like conditions The likelihood probability is equal to the output probability of the deep neural network in which the input sample belongs to the first... Probability of dust-like components , Indicates the first Prior probability of dust particles; Based on the modified posterior probability distribution, a Bayesian decision function is applied to select the category with the highest posterior probability as the final identification result, and a decision risk assessment mechanism is introduced. The formula for calculating the Bayesian decision function is as follows: ; in, This represents the final identification result based on Bayesian decision-making. This indicates that the value of c is minimized by the formula described below. This indicates that the observed polarization scattering fingerprint feature vector Under these conditions, dust belongs to the first category. Posterior probability of class Indicates the true category Misclassified as a category The loss function, when hour ,when hour Different loss values ​​are set according to the severity of the misjudgment; Set the posterior probability confidence threshold When the maximum posterior probability exceeds the posterior probability confidence threshold, a determined classification result and the corresponding posterior probability value are output. The determined classification result is the final identification result based on Bayesian decision-making. Otherwise, it is marked as uncertain and the probability distribution of each category is given to provide a reference for further manual judgment.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects: This invention acquires the polarization information of dust using polarization scattering imaging technology. Compared to traditional identification methods based on color and shape features, it can reflect the essential physical properties of dust materials at a deeper level. Polarization parameters can effectively distinguish dust types with similar appearances but different material properties. Especially for materials with different optical properties, such as metal dust, plastic dust, and composite material dust, polarization scattering features provide more stable and reliable discrimination information. By constructing a multi-dimensional polarization scattering fingerprint feature vector that includes statistical features, distribution features, and texture features, and combining it with the powerful learning capabilities of deep neural networks, this invention can achieve high-precision identification of five typical dust components, with an accuracy significantly superior to existing methods.

[0019] The polarization imaging system employed in this invention enables rapid imaging and real-time monitoring of dust distribution. The entire identification process, from image acquisition to result output, can be completed within seconds, meeting the rapid response requirements of industrial sites. Compared to traditional chemical analysis methods that require hourly detection cycles, the method of this invention offers a significant time advantage. Furthermore, this method eliminates the need for dust sampling and pretreatment, avoiding complex sample preparation processes and greatly improving detection efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a method for identifying dust components in an electronic dismantling plant based on polarization scattering imaging, according to an embodiment of the present invention. Figure 2 A schematic diagram of a multi-angle polarization image: (a) (b) (c) (d) ; Figure 3 Schematic diagram of the results of Stokes parameter calculation: (a) pseudo-color image of linear polarization degree matrix; (b) pseudo-color image of polarization angle matrix. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0022] Example 1: A method for identifying dust composition in an electronics dismantling plant based on polarization scattering imaging, such as... Figure 1 As shown, it includes the following steps: S1: Acquire multi-angle polarization scattering images of dust areas in an electronics dismantling plant using a drone system equipped with a multi-angle polarization imaging module: A drone equipped with a multi-angle polarization imaging module is controlled to fly over the dusty area of ​​the target electronics dismantling plant and activate the multi-angle polarization imaging module. The module contains multiple linear polarizers aligned with the camera sensor and, through a synchronization mechanism, continuously captures images of the same dusty area in four linear polarization directions: 0°, 45°, 90°, and 135°. It captures the scattered light information of the dust particles on the incident light, which includes sunlight or an active light source. The images captured in the four polarization directions are then temporally and spatially registered to obtain a set of grayscale or color original polarization images corresponding to the 0°, 45°, 90°, and 135° polarization directions, denoted as [image 1]. , , , ,like Figure 2 As shown. In this embodiment, the UAV flight altitude is set to 50-100 meters, and the synchronous shooting time interval of the multi-angle polarization imaging module is set to 10 milliseconds to ensure the temporal consistency of images in four polarization directions; the spatial registration accuracy is required to be within 1 pixel, the original polarization image resolution is 1920×1080 pixels, and the bit depth is 12 bits.

[0023] As an alternative implementation, the multi-angle polarization imaging module can use a liquid crystal variable polarizer instead of a fixed linear polarizer, and achieve dynamic switching of polarization direction by controlling the orientation of liquid crystal molecules through voltage. The polarization direction switching formula is as follows: ,in Indicates time polarization angle, Indicates the initial polarization angle. Indicates the magnitude of the change in polarization angle. Indicates the polarization switching frequency. It represents pi (π).

[0024] Optionally, in step S2, the Stokes vector parameters are calculated based on polarization optics theory to extract the linear polarization degree matrix and polarization angle matrix of the dust scattered light, including: Based on polarization optics theory, pixel-level calculations are performed on the four original input polarization images to extract the core polarization parameters reflecting the physical properties of dust; firstly, the first three components of the Stokes vector are calculated. , , The calculation formula is: ; ; ; in, This represents the first component of the Stokes vector, which represents the total light intensity. The second component of the Stokes vector represents the polarization difference between the 0° and 90° directions. The third component of the Stokes vector represents the polarization difference between the 45° and 135° directions; The degree of linear polarization and polarization angle are further calculated using the Stokes vector components, such as... Figure 3 As shown, the calculation formula is: ; ; in, Indicates the degree of linear polarization, used to quantify the degree of polarization of scattered light. Indicates the polarization angle, used to describe the principal polarization direction of the scattered light. This represents the arctangent function; the calculation process is performed pixel-by-pixel, yielding a linear polarization degree matrix and a polarization angle matrix with the same dimensions as the original image. In this embodiment, to avoid division by zero errors, when... When the value is less than the threshold of 0.01, the pixel's The value is set to 0; the range of values ​​in the linear polarization degree matrix is ​​limited to the interval [0,1], and the range of values ​​in the polarization angle matrix is ​​limited to... Within the range.

[0025] Optionally, in step S3, a multi-dimensional polarization scattering fingerprint feature is constructed, fusing statistical features, distribution features, and contrast features to form a polarization scattering fingerprint feature vector, including: Statistical features are extracted from the effective dust region in the linear polarization degree matrix and polarization angle matrix, and the first-order statistical moments are calculated as statistical features; the first-order statistical moments include the mean and variance, specifically: ; ; ; ; in, and These represent the mean values ​​of the degree of linear polarization and the polarization angle, respectively. and These represent the variances of the degree of linear polarization and the polarization angle, respectively. This represents the total number of pixels in the effective dust area. and They represent the first The linear polarization degree and polarization angle values ​​of each pixel; The distribution features of the polarization angle matrix are extracted, and a polarization angle distribution vector is generated through histogram analysis, dividing the polarization angle range into... Calculate the percentage of pixels in each equally spaced interval to form a... 3D distribution eigenvector The calculation formula is: ; in, The first vector representing the polarization angle distribution vector One portion, Indicates falling on the 1st The number of pixels within each angle range In this embodiment, it is set The contrast features are extracted by applying the gray-level co-occurrence matrix algorithm to the linear polarization degree matrix. Specifically, the contrast features are: ; in, Indicates texture contrast, used to characterize the spatial inhomogeneity of polarization scattering. Represents the number of gray levels. Represents the gray values ​​in the gray-level co-occurrence matrix. and The probability of coexistence; Statistical features, distribution features, and texture features are concatenated and fused. In this embodiment, all features are Z-score normalized to form a polarization scattering fingerprint feature vector. .

[0026] When encountering situations where mixed dust types lead to complex feature distributions, this embodiment employs a multimodal feature extraction method based on cluster analysis. A Gaussian mixture model is used to model the polarization feature distribution, with the number of mixed components set to 5. The model parameters are estimated using the EM algorithm, and the weights, mean, and variances of each mixed component are extracted as multimodal features.

[0027] Optionally, in step S4, the polarization scattering fingerprint feature vector is input into a pre-trained deep neural network for classification calculation, outputting the probability distribution of various dust components, including: Polarization scattering fingerprint feature vector The data is input into a pre-trained deep neural network. The deep neural network includes fully connected layers and non-linear activation functions. It establishes a non-linear mapping relationship between fingerprint features and dust material by learning from a large amount of labeled fingerprint feature data. In this embodiment, the labeled data includes five categories: pure copper dust, aluminum dust, plastic dust, mixed metal dust, and composite material dust. The deep neural network uses a 4-layer fully connected structure, with 44, 128, 64, 32, and 5 neurons in each layer, respectively. The hidden layer activation function uses the ReLU function. The deep neural network also employs Dropout regularization, with a Dropout ratio set to 0.3. The training process uses the Adam optimizer, with a learning rate set to 0.001, a batch size set to 32, and 200 training epochs. The loss function used is cross-entropy loss. The input feature vector undergoes forward propagation computation in the network, obtaining the final output through layer-by-layer linear transformations and non-linear activations. The network output layer uses the softmax activation function to transform the network's original output into a probability distribution, calculated as follows: ; in, This indicates that the input sample belongs to the first... The probability of dust-like components. and These represent the output layers of the network, respectively. The and the first The original output value of each neuron. This indicates the total number of dust component categories. , , Represents the base of the natural logarithm; obtains an output vector containing the probability distributions of various dust components.

[0028] Optionally, in step S5, a decision function is applied based on Bayesian posterior probability calculation to output the final identification result of the main components of dust from the electronic dismantling plant, including: Based on Bayesian theory, the probability distribution of the deep neural network output is corrected using posterior probability, and prior knowledge is combined to improve recognition accuracy. The prior probability distribution of various dust types is determined based on historical statistical data of dust composition from electronic dismantling plants. Indicates the first The probability of dust-like components appearing in the environment of an electronics dismantling plant; in this embodiment, based on field survey data from typical electronics dismantling plants, the prior probabilities of various dust components are set as follows: , , , , The prior probability can be adjusted according to the specific circumstances of different electronic dismantling plants. Using Bayes' theorem to calculate the posterior probability, the conditional probability output by the deep neural network is taken as the likelihood probability. The calculation formula is as follows: ; in, This indicates that the observed polarization scattering fingerprint feature vector Under these conditions, dust belongs to the first category. Posterior probability of class Indicates the first Observed under dust-like conditions The likelihood probability is equal to the output probability of the deep neural network. , Indicates the first Prior probability of dust particles, express The marginal probability; The marginal probability Calculated using the law of total probability: ; in, Indicates the first Observation of eigenvectors under dust-like conditions The likelihood probability is equal to the output probability of the deep neural network in which the input sample belongs to the first... Probability of dust-like components , Indicates the first Prior probability of dust particles; Based on the modified posterior probability distribution, a Bayesian decision function is applied to select the category with the highest posterior probability as the final identification result. A decision risk assessment mechanism is also introduced. The formula for calculating the Bayesian decision function is as follows: ; in, This represents the final identification result based on Bayesian decision-making. This indicates that the value of c is minimized by the formula described below. This indicates that the observed polarization scattering fingerprint feature vector Under these conditions, dust belongs to the first category. Posterior probability of class Indicates the true category Misclassified as a category The loss function, when hour ,when hour Different loss values ​​are set according to the severity of the misjudgment; in this embodiment, the loss function matrix... Set as a 5x5 matrix, with diagonal elements The non-diagonal elements are set according to the degree of environmental hazard of misjudgment: the loss value of misjudging toxic metal dust as harmless plastic dust is set to 10, the loss value of misjudging harmless plastic dust as toxic metal dust is set to 3, and the loss value of misjudgment within the same material is set to 1. Set the posterior probability confidence threshold When the maximum posterior probability exceeds the posterior probability confidence threshold, a determined classification result and the corresponding posterior probability value are output. This determined classification result is the final identification result based on Bayesian decision-making. Otherwise, it is marked as uncertain, and the probability distribution of each category is given as a reference for further manual judgment. In this embodiment, the posterior probability confidence threshold... Set to 0.6.

[0029] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0030] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0031] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for identifying dust components in an electronic dismantling plant based on polarization scattering imaging, characterized by, The method comprises the following steps: S1: obtaining multi-angle polarization scattering images of a dust area of an electronic disassembly plant by a multi-angle polarization imaging module mounted on a UAV system; S2: calculating Stokes vector parameters based on polarization optics theory, extracting linear polarization degree matrix and polarization angle matrix of dust scattering light; S3: constructing multi-dimensional polarization scattering fingerprint features, fusing statistical features, distribution features and contrast features to form a polarization scattering fingerprint feature vector; S4: inputting the polarization scattering fingerprint feature vector into a pre-trained deep neural network for classification calculation, and outputting probability distribution of each dust component; S5: applying a decision function based on Bayesian posterior probability calculation to output the final recognition result of the main components of the dust in the electronic disassembly plant.

2. The method of claim 1, wherein the method is a method of identifying components of dust in an electronic dismantling plant based on polarization scattering imaging, characterized by, The step S1 comprises: The unmanned aerial vehicle carrying the multi-angle polarization imaging module flies over the dust area of the target electronic disassembly plant, and the multi-angle polarization imaging module is started. The multi-angle polarization imaging module is internally provided with a plurality of linear polarizers aligned with the camera sensor. Through a synchronous mechanism, the same dust area is continuously photographed in four linear polarization directions of 0°, 45°, 90° and 135°. The scattering light information of the dust particles to the incident light is captured, and the incident light includes sunlight or an active light source. The images photographed in the four polarization directions are subjected to time and space registration processing to obtain a set of gray-scale or color original polarization images corresponding to 0°, 45°, 90° and 135° polarization directions respectively, denoted as 、 、 、 .

3. The method of claim 2, wherein the method further comprises: The step S2 comprises: Based on the theory of polarized optics, the core polarization parameters reflecting the physical properties of dust are extracted by pixel-level calculation on the input four original polarization images. Firstly, the first three components of Stokes vector are calculated , , , the calculation formula is: ； ； ； wherein S1denotes the first component of the Stokes vector, representing the total light intensity, S2denotes the second component of the Stokes vector, representing the polarization difference in the 0° and 90° directions, S3denotes the third component of the Stokes vector, representing the polarization difference in the 45° and 135° directions; The linear polarization degree and the polarization angle are further calculated by using the Stokes vector components, and the calculation formula is: ； ； wherein, denotes linear polarization degree, for quantifying the polarization degree of scattered light, denotes polarization angle, for describing the principal polarization direction of scattered light, denotes arctangent function; the calculation process is carried out pixel by pixel, obtaining a linear polarization degree matrix and a polarization angle matrix with the same size as the original image.

4. The method of claim 3, wherein the method further comprises: The step S3 comprises: The effective dust area in the linear polarization degree matrix and the polarization angle matrix is subjected to statistical feature extraction, and a first-order statistical moment is calculated as a statistical feature; the first-order statistical moment includes mean and variance, and specifically: ； ； ； ； wherein, and respectively represent the mean value of the linear polarization degree and the polarization angle, and respectively represent the variance of the linear polarization degree and the polarization angle, represents the total number of pixels of the effective dust region, and respectively represent the linear polarization degree value and the polarization angle value of the th pixel; The distribution feature of the polarization angle matrix is extracted; The contrast feature is extracted by applying a gray level co-occurrence matrix algorithm to the linear polarization degree matrix, and the contrast feature is specifically: ； wherein, denotes the texture contrast, used to characterize the spatial non-uniformity of the polarization scattering, denotes the number of gray levels, denotes the co-occurrence probability of the gray values and in the gray level co-occurrence matrix. The statistical features, distribution features and texture features are serially fused to form a polarized scattering fingerprint feature vector .

5. The polarized light scattering imaging based e-scrap plant dust composition identification method of claim 4, wherein, The distribution feature extraction of the polarization angle matrix comprises: The polarization angle distribution vector is generated by histogram analysis, the polarization angle range is divided into equal intervals, the pixel number proportion of each interval is calculated, and a dimensional distribution feature vector is formed , and the calculation formula is: ； wherein represents the i-th component of the polarization angle distribution vector, represents the number of pixels falling within the i-th angular interval, .​​ 6. The polarization scattering imaging based e-scrap plant dust composition identification method of claim 4, wherein, The step S4 comprises: polarization scattering fingerprint feature vector into a pre-trained deep neural network; the deep neural network comprises a fully connected layer and a nonlinear activation function; the deep neural network establishes a nonlinear mapping relationship from the fingerprint features to the dust material by learning a large number of labeled fingerprint feature data; The input feature vector is subjected to forward propagation calculation in the deep neural network, and the final output is obtained through layer-by-layer linear transformation and nonlinear activation; the softmax activation function is used in the output layer of the deep neural network, the original output of the deep neural network is converted into a probability distribution, and the calculation formula is: ； wherein, represents the probability that the input sample belongs to the class of dust components, and represent the raw output values of the and the neurons of the network output layer, respectively, represents the total number of classes of dust components, , , represents the base of the natural logarithm; An output vector containing probability distribution of each dust component is obtained.

7. The polarization scattering imaging based electronic dismantling plant dust component identification method according to claim 5, characterized by, The step S5 comprises: The probability distribution of the deep neural network output is corrected based on Bayesian theory, and prior knowledge is combined to improve the recognition accuracy; the prior probability distribution of each type of dust is determined according to historical statistical data of dust components in the electronic disassembly plant, and the prior probability represents the probability of the appearance of the dust component in the electronic disassembly plant environment; The posterior probability is calculated by using the Bayesian theorem, the conditional probability output by the deep neural network is taken as a likelihood probability, and the calculation formula is: ； wherein, represents the likelihood probability of observing the polarization scattering fingerprint vector under the condition that the dust belongs to the posterior probability of the dust belonging to the class, represents the likelihood probability of observing the polarization scattering fingerprint vector under the condition that the dust belongs to the class, represents the prior probability of the dust belonging to the class, represents the marginal probability of . The edge probability By the total probability formula: ； wherein, represents the likelihood probability of observing the feature vector under the assumption that the sample belongs to the class of dust components, equal to the probability that the input sample belongs to the class of dust components in the output probabilities of the deep neural network , represents the prior probability of the class of dust. A Bayesian decision function is applied based on the corrected posterior probability distribution, the class with the maximum posterior probability is selected as the final recognition result, and a decision risk evaluation mechanism is introduced, and the calculation formula of the Bayesian decision function is: ； wherein, represents the final recognition result based on the Bayesian decision, represents c that makes the following formula minimum, represents the posterior probability that the dust belongs to the class under the condition that the polarization scattering fingerprint feature vector is observed, represents the loss function when the real class is misjudged as the class , when , when different loss values are set according to the severity of the misjudgment. Setting a posterior probability confidence threshold When the maximum posterior probability exceeds the posterior probability confidence threshold, a determined classification result and a corresponding posterior probability value are output, the determined classification result being a final recognition result based on Bayesian decision, otherwise being marked as uncertain and giving a probability distribution of each category, providing a reference basis for further manual judgment.

Images