Non-ferrous metal intelligent impurity removal analysis method based on machine vision

Abstract

The invention relates to the technical field of machine vision and image processing, and discloses a non-ferrous metal intelligent impurity removal analysis method based on machine vision, and the method comprises the steps: collecting and correcting the hyperspectral time sequence data of the melt surface; constructing a spectrum prior weighted optical flow energy functional to calculate an accurate motion field capable of retaining a material boundary; combining information of two physical dimensions of spectrum and motion to identify a region of interest which is different from a background melt; extracting multi-dimensional features such as rheology, relative motion and spectral stability representing rigidity, motion independence and material stability of the region for the region; and carrying out collaborative analysis based on the characteristics to judge whether the artifacts are impurities or dynamic fluid artifacts. Through multi-dimensional feature collaborative analysis and physical judgment logic simulation, the problem that targets are difficult to distinguish only by means of single visual features is solved, and the accuracy and reliability of machine vision in intelligent impurity removal analysis application of non-ferrous metals are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of machine vision and image processing technology, specifically to a machine vision-based intelligent impurity removal analysis method for non-ferrous metals. Background Technology

[0002] In modern industrial production, non-ferrous metals such as aluminum and copper are key basic materials, and their purity directly determines the performance and quality of the final product. During the smelting process, solid or semi-solid impurities such as oxide inclusions and slag inevitably form in the melt. If these impurities are not effectively removed and enter subsequent processes, they will severely damage the mechanical properties and electrical conductivity of the material. Therefore, achieving online and accurate analysis and identification of impurities on the melt surface is a core element in improving product quality and optimizing production processes.

[0003] To replace traditional manual observation, machine vision technology has been applied to quality monitoring in non-ferrous metal smelting. Current applications primarily involve acquiring images of the melt surface using industrial cameras and then employing image processing algorithms for impurity removal analysis. These methods typically identify targets based on differences in two-dimensional visual features such as grayscale, color, or texture between impurities and the surrounding melt. When an area in the image differs in brightness or shape from the surrounding melt, the system identifies it as a potential impurity.

[0004] However, existing technologies have significant shortcomings in practical applications. First, the surface environment of non-ferrous metal melts is complex. Dynamic fluid phenomena such as eddies, bubbles, and temperature fluctuations can visually resemble certain impurities, making misjudgment highly likely when relying solely on simple image features. Second, the melt itself is in a flowing state, and traditional motion analysis techniques such as optical flow methods struggle to accurately distinguish between the rigid translational motion of solid impurities and the fluid deformation of the melt itself, leading to blurred impurity boundaries and failed motion trajectory tracking. Therefore, existing machine vision methods are insufficient to meet the demands of high-standard industrial production in terms of analytical accuracy and environmental adaptability.

[0005] Therefore, this invention proposes a machine vision-based intelligent impurity removal analysis method for non-ferrous metals to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a machine vision-based intelligent impurity removal analysis method for non-ferrous metals. This method solves the problem that traditional vision methods are unable to comprehensively analyze the kinematic characteristics and material composition stability of the target, thus making it difficult to accurately distinguish between solid impurities and dynamic fluid artifacts on the melt surface, resulting in low reliability of impurity removal analysis.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a machine vision-based intelligent impurity removal analysis method for non-ferrous metals, comprising the following steps:

[0008] S100: Acquire and correct the hyperspectral time series data sequence of the surface of non-ferrous metal melt;

[0009] S200. By utilizing the spectral information of the hyperspectral time series data sequence, a spectral prior weighted optical flow energy functional is constructed and solved to calculate the boundary-preserving motion field.

[0010] S300. Identify the region of interest on the surface of the melt based on the hyperspectral time-series data sequence and the motion field;

[0011] S400. For the region of interest, extract multi-dimensional features of the region of interest, including: rheological features, relative motion features, and spectral stability features;

[0012] S500. Based on the multi-dimensional features, perform collaborative analysis on the region of interest to determine whether the region of interest is an impurity or a dynamic fluid artifact.

[0013] Preferably, in step S200, the step of calculating the boundary-preserving motion field includes:

[0014] Construct the spectral prior weighted optical flow energy functional, which includes a data term and a smoothing term weighted by spectral prior weighting coefficients;

[0015] Calculate the prior weighting coefficients of the spectrum;

[0016] The boundary-preserving motion field is obtained by minimizing the spectral prior weighted optical flow energy functional.

[0017] Preferably, the step of calculating the spectral prior weighting coefficients includes:

[0018] Extract the local spectral gradient at any pixel location in the image. The local spectral gradient is used to quantify the degree of spectral difference between the pixel and its neighboring pixels.

[0019] The spectral prior weighting coefficients are constructed as a decreasing function of the local spectral gradient to apply a stronger smoothing constraint in spectrally homogeneous regions and a weaker smoothing constraint in regions where the spectrum changes.

[0020] Preferably, in step S300, the step of identifying the region of interest on the melt surface includes:

[0021] Based on the hyperspectral time-series data, the background melt region is determined by spectral clustering, and the average motion vector of the background melt region is calculated to obtain the background melt flow field.

[0022] Regions that differ from the background melt region in spectral characteristics or motion state are identified to generate the region of interest.

[0023] Preferably, the step of identifying regions whose spectral characteristics or motion states differ from those of the background melt region includes:

[0024] Calculate the spectral dissimilarity between the spectral vector of a pixel in the image and the spectral vector of the background, and identify pixels with spectral dissimilarity greater than the spectral threshold as spectral anomalous regions.

[0025] Calculate the magnitude of the difference between the motion vector of a pixel in the image and the vector difference between the background melt flow field, and identify pixels with a magnitude greater than the motion threshold as abnormal motion regions;

[0026] The spectral anomaly region and the motion anomaly region are merged, and connected component analysis is performed to generate the region of interest.

[0027] Preferably, in step S400, the step of extracting the rheological features includes:

[0028] Based on the motion field, calculate the curl and divergence of each pixel within the region of interest;

[0029] By averaging the curl and divergence values ​​of all pixels within the region of interest, average curl and average divergence are obtained to characterize the rigidity or fluidity of the region of interest.

[0030] Preferably, in step S400, the step of extracting the relative motion features includes:

[0031] The relative motion vector is obtained by calculating the difference between the motion vector of each pixel in the region of interest and the vector difference between the background melt flow field.

[0032] The average relative velocity, which characterizes the motion inertia of the region of interest, is obtained by calculating the average of the relative motion vector magnitudes of all pixels within the region of interest.

[0033] Preferably, in step S400, the step of extracting the spectral stability features includes:

[0034] Calculate spectral fidelity to measure the spectral consistency within the region of interest at a given time.

[0035] Spectral drift is calculated to measure the degree of change in the average spectral characteristics of the region of interest between consecutive time frames;

[0036] The spectral fidelity and the spectral drift are used together to characterize whether the material composition of the region of interest remains stable over time.

[0037] Preferably, in step S500, the step of collaboratively analyzing the region of interest based on the multi-dimensional features includes:

[0038] When the absolute values ​​of the mean curl and the mean divergence in the rheological features of a region of interest are both less than a preset rheological threshold, the spectral drift in the spectral stability features of the region of interest is less than a preset spectral drift threshold, and the average relative velocity in the relative motion features of the region of interest is greater than a preset relative motion threshold, the region of interest is determined to be an impurity.

[0039] Preferably, in step S500, the step of collaboratively analyzing the region of interest based on the multi-dimensional features includes:

[0040] When the absolute value of the mean curl or the absolute value of the mean divergence in the rheological features of a region of interest is greater than a preset rheological threshold, and the spectral drift in the spectral stability features of the region of interest is greater than a preset spectral drift threshold, the region of interest is determined to be a dynamic fluid artifact.

[0041] This invention provides a machine vision-based intelligent impurity removal analysis method for non-ferrous metals. It has the following beneficial effects:

[0042] 1. This invention calculates the motion field by constructing a spectral prior weighted optical flow energy functional, effectively solving the problem of decreased accuracy caused by motion blur in traditional machine vision methods when dealing with impurities and melt boundaries. This method adaptively applies smoothing constraints using spectral information, accurately solving for motion information while maintaining clear material boundaries. This provides a high-precision kinematic data foundation for subsequent identification and analysis, fundamentally improving the accuracy and reliability of the entire intelligent impurity removal analysis system.

[0043] 2. This invention successfully solves the technical challenge of distinguishing solid impurities from dynamic fluid artifacts such as eddies and bubbles by extracting multi-dimensional features including rheological characteristics, relative motion, and spectral stability, and by performing synergistic analysis of these features. This method simulates a comprehensive judgment logic at the physical level, achieving accurate classification of regions of interest by quantifying the rigidity, motion independence, and material composition stability of the target region. This significantly reduces the misjudgment rate of traditional analysis methods, making impurity identification results in non-ferrous metal production processes more intelligent and reliable.

[0044] 3. This invention identifies regions of interest by comprehensively utilizing two physical criteria: spectral anomalies and motion anomalies. Compared to detection methods that rely on a single feature, this significantly improves the detection rate of potential impurities and the overall robustness of the method. Whether the impurities are of special material composition but follow mainstream motion, or targets with independent motion but insignificant spectral characteristics, this invention can effectively capture them, ensuring comprehensive analysis and enabling it to better adapt to the complex and variable conditions in non-ferrous metal smelting processes. This provides a strong guarantee for achieving efficient and intelligent impurity removal analysis. Attached Figure Description

[0045] Figure 1 This is a flowchart of a machine vision-based intelligent impurity removal analysis method for non-ferrous metals according to the present invention.

[0046] Figure 2 This is a flowchart of the collaborative analysis and determination process of the present invention;

[0047] Figure 3 This is an architecture diagram of a machine vision-based intelligent impurity removal and analysis system for non-ferrous metals according to the present invention.

[0048] The module includes: 10. Data acquisition and preprocessing module; 20. Sports field calculation module; 30. Region of interest identification module; 40. Multidimensional feature extraction module; and 50. Collaborative analysis and decision-making module. Detailed Implementation

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] See attached document Figure 1 This invention provides a machine vision-based intelligent impurity removal analysis method for non-ferrous metals, comprising the following steps:

[0051] S100. Acquire and calibrate the hyperspectral time-series data sequence of the non-ferrous metal melt surface. This hyperspectral time-series data sequence is used as the data basis for subsequent motion field calculations and spectral feature analysis.

[0052] S200. By utilizing the spectral information of the hyperspectral time-series data sequence, a spectral prior-weighted optical flow energy functional is constructed and solved to calculate the boundary-preserving motion field. This motion field is used to characterize the motion state of each point on the melt surface within the time series.

[0053] S300. Based on the hyperspectral time-series data and motion field, identify the region of interest (ROI) on the melt surface. The ROI is a candidate region whose spectral characteristics or motion state differs from the background melt.

[0054] S400. Extract multi-dimensional features from the region of interest. These multi-dimensional features include: rheological features to characterize rigidity or fluidity, relative motion features to characterize inertia, and spectral stability features to characterize material stability.

[0055] S500 performs collaborative analysis on regions of interest based on multi-dimensional features (a combination of rheological features, relative motion features, and spectral stability features) to determine whether the region of interest is an impurity or a dynamic fluid artifact.

[0056] This invention discloses a machine vision-based intelligent impurity removal analysis method for non-ferrous metals, which can be executed by a machine vision-based intelligent impurity removal analysis system for non-ferrous metals. This machine vision-based intelligent impurity removal analysis system for non-ferrous metals can be deployed on edge computing devices, embedded industrial computers, or servers.

[0057] Physically, this machine vision-based intelligent impurity removal and analysis system for non-ferrous metals can be embodied as a data processing and control terminal. This data processing and control terminal can be an industrial computer or high-performance server equipped with a processor (CPU), a graphics processing unit (GPU), and memory, internally storing and running program instructions that implement the machine vision-based intelligent impurity removal and analysis method for non-ferrous metals of this invention.

[0058] See attached document Figure 3 This machine vision-based intelligent impurity removal and analysis system for non-ferrous metals can be functionally divided into multiple collaborative functional modules, which run as software programs on a data processing and control terminal. In a specific embodiment, the system may include: a data acquisition and preprocessing module 10, a motion field calculation module 20, a region of interest identification module 30, a multi-dimensional feature extraction module 40, and a collaborative analysis and decision-making module 50.

[0059] The data acquisition and preprocessing module 10 is used to perform step S100.

[0060] The sports field calculation module 20 is used to execute step S200.

[0061] Region of Interest (ROI) identification module 30 is used to perform step S300.

[0062] The multidimensional feature extraction module 40 is used to perform step S400.

[0063] Collaborative analysis and decision-making module 50 is used to execute step S500.

[0064] See attached document Figure 1 and Figure 2 To further clarify the technical details of the method of the present invention, the specific implementation methods, principles and technical contents of each of the above steps will be described in detail below.

[0065] In one specific embodiment, step S100 can be broken down into the following sub-steps.

[0066] S101, Data Acquisition. This step uses a hyperspectral imaging system to continuously and dynamically image the melt surface in a non-contact manner. In one embodiment, the hyperspectral imaging system can be a pushbroom or snapshot hyperspectral camera with a spectral range covering the visible to near-infrared bands to capture the characteristic spectra of different impurities. To ensure the capture of dynamic changes on the melt surface, the system's acquisition frame rate should meet the application scenario requirements. During deployment, the hyperspectral imaging system can be mounted above the smelting equipment with its lens optical axis perpendicular to the melt surface to reduce image geometric distortion caused by the tilt of the observation angle.

[0067] The data acquisition process generates a four-dimensional hyperspectral time-series data sequence, which can be represented as:

[0068] ;

[0069] in, and These are the spatial coordinates on the image plane; Spectral wavelength; This refers to the time or data frame number.

[0070] S102, Data Preprocessing. This step performs a series of corrections on the raw hyperspectral time-series data sequence acquired in S101 to eliminate errors introduced by the sensor itself and the external environment.

[0071] First, the raw hyperspectral time-series data is radiometrically calibrated. This process converts the dimensionless digital quantization (DN) values ​​recorded by the sensor into units of radiance or reflectance with explicit physical meaning. This calibration ensures the consistency and comparability of spectral data collected at different times and spatial locations, and forms the basis for subsequent spectral feature analysis and comparison.

[0072] Next, noise correction is performed on the radiometrically calibrated data. Since the high-temperature melting environment introduces thermal noise, and the sensor itself contains random noise (such as salt-and-pepper noise), this step applies median filtering to each single-band image in the data sequence. Median filtering is a non-linear filtering technique that effectively removes isolated impulse noise points while preserving edge details in the image, which is crucial for subsequent identification of impurity contours.

[0073] Furthermore, sensor artifact correction is performed, primarily targeting potential dead pixels and stripe noise in the hyperspectral sensor. For dead pixels, interpolation replacement can be performed using the spectral information of their neighboring valid pixels. For stripe noise correction, the methods involved are well-known in the field, such as statistical correction methods (e.g., histogram matching), and will not be elaborated upon here.

[0074] After the above data preprocessing sub-steps, a corrected hyperspectral time series data sequence that accurately reflects the physical and chemical state of the melt surface is finally obtained for subsequent processing steps.

[0075] Step S200 aims to calculate the motion field of the melt surface. Traditional motion field estimation algorithms suffer from motion ambiguity at material boundaries (e.g., the interface between solid impurities and liquid melt) due to their inherent smoothness assumptions, resulting in unclear boundaries. This step introduces spectral information from hyperspectral time-series data as prior knowledge, constructs and solves a spectral prior weighted optical flow energy functional, and finally calculates a boundary-preserving motion field that accurately characterizes the motion state of each independent region on the melt surface. In a specific embodiment, step S200 can be decomposed into the following sub-steps.

[0076] S201, Construct the spectral prior-weighted optical flow energy functional. This step is based on a variational optical flow model, which transforms the problem of solving the motion field into a minimization problem of an energy functional. Optical flow energy functional Typically composed of data items and smoothing terms constitute.

[0077] ;

[0078] in, The sports field to be solved. and In the image coordinate system and The velocity component in the direction; This is a regularization parameter used to balance the weight of data terms and smoothing terms in the functional.

[0079] Data Items This is based on the assumption of constant brightness, meaning that the brightness of the same physical point remains unchanged between two consecutive frames. In one embodiment, a characteristic band image that provides good differentiation between impurities and melt can be selected. To calculate this data item:

[0080] ;

[0081] in, , and Images exist Spatial direction Partial derivatives in the spatial and temporal directions.

[0082] Smoothing terms The purpose is to ensure that the solved motion field is spatially smooth. The key to this step is that a spectral prior weighting coefficient is introduced into the smoothing term. This achieves spatially adaptive anisotropic smoothing. The weighted smoothing term can be expressed as:

[0083] ;

[0084] in, ,right Similarly; These are the spectral prior weighting coefficients, whose values ​​are determined by the pixels. The spectral characteristics of the region and its neighborhood determine its function. Its role is to apply a strong smoothing constraint in spectrally homogeneous regions (such as the interior of a melt). Larger values ​​are applied to regions where the spectrum changes drastically (such as impurity boundaries), while weaker smoothing constraints are applied. (The value is relatively small), thus effectively preserving the material boundary when calculating the motion field.

[0085] S202, Calculate the spectral prior weighting coefficients First, for any pixel in the image... From hyperspectral time series data Extract it at time spectral vector Then, the local spectral gradient at that pixel location is calculated. The local spectral gradient is used to quantify the degree of spectral difference between the pixel and its neighboring pixels.

[0086] Spectral prior weighting coefficients Constructed as a local spectral gradient A decreasing function. In one embodiment, it can be defined in the form of a Gaussian function:

[0087] ;

[0088] in, It is a control parameter used to adjust the sensitivity of the weighting coefficients to changes in the spectral gradient; when the local spectral gradient... When it is large, The value approaches 0; when the local spectral gradient When smaller, The value approaches 1.

[0089] S203, Solve for the spectral prior weighted optical flow energy functional. Substituting the above data terms, weighted smoothing terms, and weighting coefficients, we obtain the final spectral prior weighted optical flow energy functional to be minimized:

[0090] ;

[0091] The minimization of this energy functional can be achieved by solving its corresponding Euler-Lagrange equations. Since the equations are typically nonlinear, in practical applications, numerical solutions can be obtained using numerical iterative methods (such as the Gauss-Seidel iterative method). The derivation and numerical solution of the Euler-Lagrange equations can be achieved by those skilled in the art based on publicly available variational optical flow theory, which is well-known in the field and will not be elaborated upon here.

[0092] By solving the above functional, the final sports field is obtained. This is a boundary-preserving motion field. This motion field can maintain abrupt velocity changes at the boundary between impurities and the melt, while remaining smooth within the impurities or melt, providing accurate kinematic information for subsequent region of interest identification and feature extraction.

[0093] Step S300 aims to identify all potential impurity candidate regions, i.e., regions of interest, from the melt surface. This step integrates the hyperspectral time-series data sequence obtained in step S100 and the boundary-preserving motion field calculated in step S200, and is achieved by identifying regions that differ significantly from the background melt in spectral characteristics or motion states. In a specific embodiment, step S300 can be decomposed into the following sub-steps.

[0094] S301, Estimate the background melt flow field. To establish a benchmark for comparison, the motion state of the main background melt needs to be defined and quantified first. This step can be done using spectral clustering methods, such as the K-Means clustering algorithm, to segment a frame of hyperspectral image in a hyperspectral time-series data sequence into multiple categories. Typically, the category occupying the largest area in the image corresponds to the background melt. This region is defined as the background melt region. .

[0095] Determining the background melt region Then, by calculating the average value of the motion vectors of all pixels within the background melt region, the background melt flow field is obtained. The background melt flow field represents the state at time t. The overall translational or flow trend of the melt body.

[0096] S302, Identify Spectral Anomalies. This step aims to identify regions where the melt material differs from the background melt material. First, the background melt region is calculated. The average value of the spectra of all pixels within the area yields a representative background spectral vector. Then iterate through each pixel in the image. Calculate its spectral vector With background spectral vector The spectral dissimilarity between them.

[0097] In one embodiment, spectral anisotropy can be measured by spectral angle. When the spectral angle of a pixel is greater than a preset spectral threshold... When the difference is 0.1 radians, the pixel is identified as a spectral anomaly. The set of all spectral anomalies constitutes the spectral anomaly region. .

[0098] S303, Identify regions of abnormal motion. This step aims to identify regions where the motion is independent of the background melt flow. This involves traversing every pixel in the image. Calculate its motion vector With background melt flow field Vector difference: .

[0099] This vector difference Length of the module Quantized pixels The degree of motion abnormality. When the magnitude exceeds the preset motion threshold. At a resolution of 1.5 pixels per frame (e.g., 1.5 pixels per frame), the pixel is identified as a motion anomaly. The set of all motion anomalies constitutes the motion anomaly region. .

[0100] S304, Generate the region of interest. This involves identifying the spectral anomalies in the preceding steps. With abnormal movement area The pixels are merged to obtain an initial candidate region mask. In one embodiment, the merging operation is a logical OR operation, meaning that any pixel belonging to a spectral anomaly region or a motion anomaly region is considered a candidate point.

[0101] Since the initial candidate region mask may contain noisy or discontinuous small regions, morphological processing can be performed on it for optimization. For example, morphological closing operations can be applied to fill small holes inside the candidate regions and connect adjacent region blocks.

[0102] Finally, connected component analysis is performed on the morphologically processed binary mask image, and each independent, connected region is marked as a region of interest. This step ultimately outputs a set of regions of interest, each representing a candidate impurity that requires further detailed analysis.

[0103] Step S400 extracts multi-dimensional features for each region of interest identified in step S300. These multi-dimensional features aim to quantify the attributes of the region of interest from different physical levels, providing a basis for the final determination in step S500. In a specific embodiment, step S400 can be decomposed into the following sub-steps.

[0104] S401, Extract rheological features. These rheological features characterize whether the region of interest tends towards rigid translation or fluid-like deformation and rotation. Based on the boundary-preserving motion field calculated in step S200. It can calculate two key rheological features: mean curl and mean divergence.

[0105] For any point on the sports field Its curl and divergence Defined as:

[0106] ;

[0107] ;

[0108] in, and The sports field is located in and The velocity component in the direction.

[0109] For a given region of interest Its mean curl and mean divergence It is obtained by averaging the curl and divergence values ​​of all pixels within the region. Curl Characterized by the local rotation intensity and divergence of the sports field This characterizes the intensity of local expansion or contraction of the motion field. Rigid objects have near-zero internal curl and divergence during translation, while fluids (such as melt eddies) exhibit significant curl or divergence.

[0110] S402, Extract relative motion features. These relative motion features characterize the region of interest as a whole, and its motion is independent of the surrounding background melt, thus reflecting its motion inertia. This is based on the background melt flow field estimated in step S301. It can calculate the average relative velocity of the region of interest.

[0111] For the region of interest any cell within Calculate its motion vector With background melt flow field The vector difference is used to obtain the relative motion vector of the pixel. ,Right now:

[0112] The average relative velocity in this region The following is obtained by calculating the average of the relative motion vector magnitudes of all pixels within the region:

[0113] ;

[0114] in, Region of Interest The total number of pixels within the area. A solid impurity with a large mass and inertia will move significantly independently of the background melt flow field, thus exhibiting a large average relative velocity. Conversely, scum or purely fluid artifacts with a density close to that of the melt tend to follow the background melt, with a smaller average relative velocity.

[0115] S403, Extract spectral stability features. These features characterize whether the composition of the material in the region of interest remains stable over time. For stable solid impurities, the spectral features should exhibit high temporal consistency, while the spectral features of dynamic fluid artifacts (such as bubble bursts and localized temperature fluctuations) change rapidly. This step extracts two features—spectral fidelity and spectral drift—to quantify this stability.

[0116] Spectral fidelity Used to measure at a certain moment Area of ​​Interest Internal spectral consistency. First, calculate the region of interest at time [time value missing]. average spectral vector Then, the spectral vector of each pixel within the region of interest is calculated. With this average spectral vector The spectral similarity is the average of all similarities, which is the spectral fidelity.

[0117] Spectral drift Used to measure the average spectral characteristics of a region of interest across consecutive frames (e.g., from time 1 to 2). arrive The degree of change between ) . The sports field will be used to represent the time intervals. Region of interest Mapping to time The position is obtained Calculate the average spectral vectors for these two regions respectively. and The spectral dissimilarity between these two average spectral vectors is called spectral drift. Spectral similarity and dissimilarity can be measured using methods such as spectral angle or spectral correlation coefficient.

[0118] By executing the above sub-steps, a multi-dimensional feature, encompassing rheological characteristics, relative motion characteristics, and spectral stability characteristics, is generated for each region of interest. This multi-dimensional feature comprehensively describes the physical and chemical properties of the candidate region, providing a quantitative basis for subsequent collaborative analysis and decision-making.

[0119] Step S500 performs collaborative analysis based on the multi-dimensional features extracted for each region of interest in step S400 to ultimately determine whether the region of interest is an impurity with a physical entity or a dynamic fluid artifact generated by the melt's own flow. This analysis process utilizes the inherent physical correlation between rheological features, relative motion features, and spectral stability features. In a specific embodiment, step S500 is implemented through a rule-based decision logic.

[0120] S501, Collaborative Analysis and Judgment. For each region of interest... This step classifies the data by evaluating a set of logical conditions, along with its corresponding multi-dimensional features.

[0121] A region of interest is identified as a dynamic fluid artifact if its characteristics satisfy the following combination of conditions: the region of interest exhibits significant fluid motion characteristics and its material composition is unstable. Specifically, this includes its rheological characteristics (mean curl). or mean divergence The absolute value of ) is greater than the preset rheological threshold ( or For example, 0.05 frames -1 ), and its spectral stability characteristics include spectral drift. Greater than the preset spectral drift threshold (e.g., 0.08 radians). Dynamic fluid artifacts, such as eddies or bubbles on the surface of a melt, involve internal rotation or deformation, and their spectral characteristics change rapidly due to temperature fluctuations or morphological changes.

[0122] A region of interest is identified as an impurity if its characteristics satisfy a combination of conditions opposite to those of dynamic fluid artifacts: the region of interest exhibits rigid body motion characteristics, its material composition is stable, and its motion state is independent of the background melt. Specifically, this refers to its rheological characteristics (mean curl). and mean divergence The absolute values ​​of all values ​​are less than the preset rheological threshold (0.05frame). -1 ), its spectral stability characteristics include spectral drift Less than the preset spectral drift threshold (e.g., 0.08 radians), and its relative motion characteristics (average relative velocity) The relative motion threshold is greater than the preset threshold. (e.g., 1.5 pixels / frame). Solid or semi-solid impurities move as a whole, with minimal internal deformation and rotation. Their chemical composition remains stable for a short period, and due to their own inertia, their trajectory differs significantly from that of the surrounding liquid melt.

[0123] in, , , and The judgment threshold is determined through experimental calibration or empirical setting based on the melt physical properties, impurity types, and imaging system parameters of the specific application scenario.

[0124] By performing the aforementioned collaborative analysis, step S500 completes the classification of all regions of interest. The output is a classification map, which clearly marks the location and outline of regions identified as impurities, as well as identified and excluded dynamic fluid artifacts. This result can be directly used to guide subsequent physical impurity removal operations or to quantitatively assess the purity of the smelting process.

[0125] To illustrate the specific application of the machine vision-based intelligent impurity removal analysis method for non-ferrous metals of the present invention, an example in an aluminum alloy smelting and casting production line is given below.

[0126] The objective of this embodiment is to perform online monitoring of the molten surface within a holding furnace for 6061 aluminum alloy, and to automatically identify and locate slag that needs to be removed before casting. The temperature of the molten aluminum within the holding furnace is maintained at 720℃±10℃.

[0127] A pushbroom hyperspectral camera was mounted approximately 2.5 meters directly above the opening of the holding furnace, with its lens pointing vertically towards the molten surface. The camera was housed within a high-temperature resistant protective enclosure with forced air cooling. Its spectral range was 400 nm to 1000 nm, and the acquisition frame rate was set to 50 Hz to capture the dynamic processes on the molten surface.

[0128] A hyperspectral camera is connected via industrial Ethernet to an edge computing device deployed next to the furnace. This edge computing device is an industrial computer equipped with a high-performance graphics processing unit (GPU) for real-time execution of the analysis method of this invention. The calculation results from this edge computing device are communicated via the Profibus bus protocol to the programmable logic controller (PLC) of a six-axis industrial robot responsible for performing physical slag removal operations. The end effector of this industrial robot is equipped with a high-temperature resistant slag removal plate.

[0129] In a typical operating cycle, this machine vision-based intelligent impurity removal and analysis system for non-ferrous metals performs the following steps:

[0130] Data acquisition and preprocessing (corresponding to S100)

[0131] The data acquisition and preprocessing module 10 controls the hyperspectral camera to continuously image the surface of the melt, generating... Data stream (hyperspectral time-series data sequence). The acquired raw DN value data is radiometrically calibrated and converted to radiance units. Subsequently, the system performs a 3x3 window median filter on the image of each band to suppress random thermal noise.

[0132] Sports field calculation (corresponding to S200)

[0133] The motion field calculation module 20 receives the preprocessed hyperspectral time-series data sequence. On the surface of the molten aluminum alloy, there exists a slow background flow caused by the holding current or residual agitation. Simultaneously, a semi-solid alumina slag, approximately 5cm x 8cm in size, may float on the surface. The motion field calculation module 20 utilizes the significant spectral difference between the slag and the molten aluminum in the 700-900 nm wavelength range to calculate the spectral prior weighting coefficients. At the boundary between the slag and the molten aluminum, The value is close to 0, thus preserving the velocity discontinuity on both sides of the boundary when solving the optical flow energy functional. The final output boundary-preserved motion field accurately depicts the flow field of the background molten aluminum and the overall translational vector of the slag.

[0134] Region of Interest (ROI) identification (corresponding to S300)

[0135] The region of interest (ROI) identification module 30 first segments the image based on the K-Means clustering algorithm and identifies the background melt region with the largest area. The average motion vector was calculated as the background melt flow field. At this point, in addition to the aforementioned scum, there is also a liquid vortex on the surface of the melt caused by localized temperature unevenness.

[0136] Spectral anomaly identification: spectral vector of scum and background spectral vector The spectral angle is greater than the spectral threshold. Therefore, the scum area was marked as a spectral anomaly region. The liquid vortex, being made of the same material as the background melt, had a spectral angle less than the threshold and was therefore not labeled.

[0137] Anomaly detection: Due to its inertia, the motion vector of scum interacts with the background melt flow field. The difference is greater than the motion threshold. Therefore, it was marked as a region of abnormal motion. The motion of the liquid eddies also differs from the background flow field and is similarly labeled.

[0138] Merging and Generating: The Region of Interest (ROI) identification module 30 merges the scum and eddy regions and generates two independent ROIs through connected component analysis. (Scum) and (vortex).

[0139] Multi-dimensional feature extraction (corresponding to S400)

[0140] The multidimensional feature extraction module 40 respectively... and Feature extraction:

[0141] for (scum):

[0142] Rheological characteristics: Its internal motion field exhibits overall translational motion, and the calculated mean curl... and mean divergence The value is close to zero.

[0143] Relative motion characteristics: its average relative velocity The calculated value is significantly greater than zero, indicating that its motion is independent of the background flow.

[0144] Spectral stability characteristics: As a physical entity, its average spectral vector changes very little between consecutive frames, and the calculated spectral drift... The value is very low.

[0145] for (vortex):

[0146] Rheological characteristics: There is obvious rotation in its internal motion field, and the calculated mean curl The absolute value is very large.

[0147] Relative motion characteristics: Its motion basically follows the background flow, with an average relative velocity. Smaller.

[0148] Spectral stability characteristics: The temperature and morphology of the eddy region are constantly changing, causing spectral drift. The value is relatively high.

[0149] Collaborative analysis and decision-making (corresponding to S500)

[0150] The collaborative analysis and decision-making module 50 makes judgments based on preset decision rules:

[0151] for Its rheological characteristic value is small, its relative motion characteristic value is large, and its spectral stability characteristic value is high (i.e., small spectral drift). This combination of characteristics satisfies the conditions for determining it to be an impurity.

[0152] for Its rheological characteristic value is large. This characteristic satisfies the condition for determining it as a dynamic fluid artifact.

[0153] The final output of the collaborative analysis and decision-making module 50 is the judgment result: Position, circumscribed rectangle size, and current motion vector The information is packaged and sent to the PLC of the industrial robot via the Profibus bus. It is then judged as an artifact and ignored.

[0154] System linkage and execution

[0155] After receiving the impurity information, the robot's PLC immediately starts the preset slag removal program. Based on the received coordinate and speed information, the robot calculates the lead time and controls the end-effector to precisely move in front of the slag and push it away from the casting area, completing a precise and automated impurity removal operation.

[0156] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine vision-based intelligent impurity removal analysis method for non-ferrous metals, characterized in that, Includes the following steps: S100: Acquire and correct the hyperspectral time series data sequence of the surface of non-ferrous metal melt; S200. By utilizing the spectral information of the hyperspectral time series data sequence, a spectral prior weighted optical flow energy functional is constructed and solved to calculate the boundary-preserving motion field. S300. Identify the region of interest on the surface of the melt based on the hyperspectral time-series data sequence and the motion field; S400. For the region of interest, extract multi-dimensional features of the region of interest, including: rheological features, relative motion features, and spectral stability features; S500. Based on the multi-dimensional features, perform collaborative analysis on the region of interest to determine whether the region of interest is an impurity or a dynamic fluid artifact.

2. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 1, characterized in that, In step S200, the step of calculating the boundary-preserving motion field includes: Construct the spectral prior weighted optical flow energy functional, which includes a data term and a smoothing term weighted by spectral prior weighting coefficients; Calculate the prior weighting coefficients of the spectrum; The boundary-preserving motion field is obtained by minimizing the spectral prior weighted optical flow energy functional.

3. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 2, characterized in that, The steps for calculating the spectral prior weighting coefficients include: Extract the local spectral gradient at any pixel location in the image. The local spectral gradient is used to quantify the degree of spectral difference between the pixel and its neighboring pixels. The spectral prior weighting coefficients are constructed as a decreasing function of the local spectral gradient to apply a stronger smoothing constraint in spectrally homogeneous regions and a weaker smoothing constraint in regions where the spectrum changes.

4. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 1, characterized in that, In step S300, the step of identifying the region of interest on the surface of the melt includes: Based on the hyperspectral time-series data, the background melt region is determined by spectral clustering, and the average motion vector of the background melt region is calculated to obtain the background melt flow field. Regions that differ from the background melt region in spectral characteristics or motion state are identified to generate the region of interest.

5. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 4, characterized in that, The step of identifying regions whose spectral characteristics or motion states differ from the background melt region includes: Calculate the spectral dissimilarity between the spectral vector of a pixel in the image and the spectral vector of the background, and identify pixels with spectral dissimilarity greater than the spectral threshold as spectral anomalous regions. Calculate the magnitude of the difference between the motion vector of a pixel in the image and the vector difference between the background melt flow field, and identify pixels with a magnitude greater than the motion threshold as abnormal motion regions; The spectral anomaly region and the motion anomaly region are merged, and connected component analysis is performed to generate the region of interest.

6. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 1, characterized in that, In step S400, the step of extracting the rheological features includes: Based on the motion field, calculate the curl and divergence of each pixel within the region of interest; By averaging the curl and divergence values ​​of all pixels within the region of interest, average curl and average divergence are obtained to characterize the rigidity or fluidity of the region of interest.

7. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 1, characterized in that, In step S400, the step of extracting the relative motion features includes: The relative motion vector is obtained by calculating the difference between the motion vector of each pixel in the region of interest and the vector difference between the background melt flow field. The average relative velocity, which characterizes the motion inertia of the region of interest, is obtained by calculating the average of the relative motion vector magnitudes of all pixels within the region of interest.

8. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 1, characterized in that, In step S400, the step of extracting the spectral stability features includes: Calculate spectral fidelity to measure the spectral consistency within the region of interest at a given time. Spectral drift is calculated to measure the degree of change in the average spectral characteristics of the region of interest between consecutive time frames; The spectral fidelity and the spectral drift are used together to characterize whether the material composition of the region of interest remains stable over time.

9. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 1, characterized in that, In step S500, the step of collaboratively analyzing the region of interest based on the multi-dimensional features includes: When the absolute values ​​of the mean curl and the mean divergence in the rheological features of a region of interest are both less than a preset rheological threshold, the spectral drift in the spectral stability features of the region of interest is less than a preset spectral drift threshold, and the average relative velocity in the relative motion features of the region of interest is greater than a preset relative motion threshold, the region of interest is determined to be an impurity.

10. The intelligent impurity removal analysis method for non-ferrous metals based on machine vision according to claim 1, characterized in that, In step S500, the step of collaboratively analyzing the region of interest based on the multi-dimensional features includes: When the absolute value of the mean curl or the absolute value of the mean divergence in the rheological features of a region of interest is greater than a preset rheological threshold, and the spectral drift in the spectral stability features of the region of interest is greater than a preset spectral drift threshold, the region of interest is determined to be a dynamic fluid artifact.

Images