A machine vision-based quality detection method for basic cobalt carbonate powder material

Abstract

This invention discloses a machine vision-based method for quality inspection of basic cobalt carbonate powder, belonging to the field of online powder quality inspection technology. The method acquires cross-polarized images and multi-exposure image sequences of basic cobalt carbonate powder, combines a reference color chart with qualified batch samples to restore the true color, and performs powder layer morphology correction and thickness interference compensation on the true color images, effectively suppressing false anomalies such as residual wet reflection and uneven thickness. Then, through anomaly feature extraction, time-series tracking, and clustering, stable defect clusters are obtained. After batch consistency judgment, a minimum re-inspection sampling window and re-inspection guidelines are generated. This invention achieves online, real-time, and multi-dimensional quality inspection of basic cobalt carbonate powder, significantly improving defect detection accuracy, reducing false alarm and false negative rates, and greatly narrowing the scope of re-inspection.

Description

Technical Field

[0001] This invention belongs to the field of online powder quality detection technology, specifically relating to a machine vision-based method for detecting the quality of basic cobalt carbonate powder materials. Background Technology

[0002] Basic cobalt carbonate is a key precursor for cathode materials in power batteries, and its quality stability directly affects the performance of downstream batteries, making it crucial in the power battery industry chain. However, batch-to-batch variations in raw material batches, mother liquor ratios, and powder moisture content during the preparation of cobalt powder and cobalt chloride raw materials result in complex visual phenomena in the powder, making online detection difficult. Existing technologies are insufficient in this area, becoming a bottleneck for industry development.

[0003] Currently, the industry generally adopts a model that primarily uses offline sampling inspection, supplemented by online visual inspection. Offline inspection requires manual sampling followed by testing various powder indicators using physicochemical methods such as laser particle size analyzers, ICP, and XRD. Online inspection often uses ordinary industrial cameras for simple threshold identification, with manual verification when anomalies are suspected. Some production lines supplement this with laser particle size analyzers to monitor particle size, but neither can achieve multi-dimensional online integrated inspection.

[0004] The existing detection technologies mentioned above have significant shortcomings and cannot meet the high-quality detection requirements of battery-grade basic cobalt carbonate powder in the synthesis and preparation system of cobalt powder and cobalt chloride raw materials. On the one hand, most existing online visual inspections are based on a single visible light image and simple threshold rules, which are easily affected by factors such as residual moisture on the powder surface, specular reflection of light, fluctuations in powder layer thickness, and changes in fabric texture. They are difficult to effectively remove false anomaly signals and often misjudge non-quality defect areas that appear to be agglomerates or color deviations as real defects, resulting in insufficient detection accuracy, high false alarm rate, and inability to accurately identify quality problems introduced by real agglomeration, uneven crystal growth, and impurities. On the other hand, even if existing detection solutions can identify local coarse particles or color deviations, they only remain at the level of single-frame, single-region anomaly indication. They cannot link local anomalies with batch-level quality uniformity, anomaly causes, and re-inspection range, and cannot simultaneously output complete detection results including defect type, batch quality level, anomaly cause, and minimum re-inspection sampling range. Furthermore, offline sampling inspection has a significant lag and cannot promptly report quality anomalies in the production process, which can easily lead to the generation of batches of unqualified batches, increasing production costs and quality control risks. Therefore, this invention proposes a machine vision-based method for quality inspection of basic cobalt carbonate powder materials. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a machine vision-based method for quality inspection of basic cobalt carbonate powder materials. The aim is to at least partially solve one or more problems existing in the prior art, such as low accuracy of online detection of basic cobalt carbonate powder, high false alarm rate, inability to achieve multi-dimensional integrated detection, and delayed detection results.

[0007] To at least solve the above problems, the present invention provides the following technical solution: A machine vision-based method for quality inspection of basic cobalt carbonate powder materials, comprising: Cross-polarization images and multi-exposure image sequences of basic cobalt carbonate powder were acquired to obtain the original detection image set; Based on the reference color chart and qualified batch samples, the original detection image set is restored to true color to obtain true color image results; The true color image result is subjected to powder layer morphology correction and thickness interference compensation to obtain a thickness-corrected image result; Extract the abnormal features of the thickness correction image results, perform time-series tracking on the abnormal features, and obtain stable defect cluster results; The consistency of the stable defect cluster results is determined to obtain the batch quality inspection results; Based on the batch quality inspection results, a minimum re-inspection sampling window is generated, and a re-inspection area is planned for the minimum re-inspection sampling window to obtain the re-inspection guidance results.

[0008] As a preferred embodiment, true color restoration is performed on the original detection image set based on a reference color chart and qualified batch samples to obtain true color image results, including: Collect images of white, gray, and black reference plates and images of qualified batch samples, perform reference calibration, and obtain reference data; Based on the aforementioned reference data, reflection suppression and brightness compensation processing are performed on the cross-polarized image and the multi-exposure image to obtain calibration parameters; Based on the calibration parameters, the original detection image set is corrected to obtain a true color image result.

[0009] As a preferred embodiment, based on the reference data, reflection suppression and brightness compensation processing are performed on the cross-polarized image and the multi-exposure image to obtain calibration parameters, including: Reflection suppression processing is performed on the cross-polarized image to remove interference from residual moisture and surface reflection, thus obtaining reflection suppression features; Brightness equalization processing is performed on multi-exposure images to compensate for the difference between highlights and shadows, resulting in brightness and shadow compensation features; The calibration parameters are obtained by normalizing and fitting the reflection suppression feature and brightness compensation feature together with the reference data.

[0010] As a preferred implementation, the reflection suppression feature and brightness compensation feature are normalized and fitted using the reference data to obtain calibration parameters, including: The reflection suppression feature and the brightness compensation feature are respectively subjected to numerical normalization to obtain normalized feature data; The normalized feature data and the benchmark reference data are used to establish a feature mapping relationship, and the feature weights are obtained by fitting. The preset correction threshold and the feature weights are integrated to generate calibration parameters.

[0011] As a preferred implementation, establishing a feature mapping relationship between the normalized feature data and the benchmark reference data, and fitting the feature weights, includes: Using the color and brightness standards in the benchmark reference data as a reference, feature matching is performed on the normalized feature data to obtain the feature correspondence; Based on the aforementioned feature correspondence, linear fitting calculations are performed to obtain feature weights.

[0012] In a preferred embodiment, the true color image result is subjected to powder layer morphology correction and thickness interference compensation to obtain a thickness-corrected image result, including: The true color image results are subjected to powder layer boundary contour extraction and morphological optimization to obtain an accurate powder layer boundary image; Based on the standard powder layer thickness parameters of qualified batch samples, thickness difference analysis is performed on the precise powder layer boundary image to obtain thickness deviation data. Based on the thickness deviation data, thickness compensation correction is performed on local areas of the powder layer to obtain thickness correction image results.

[0013] As a preferred embodiment, abnormal features of the thickness correction image result are extracted, and time-series tracking of these abnormal features is performed to obtain stable defect cluster results, including: Anomaly detection is performed on the thickness-corrected image results, and abnormal features are extracted; Multi-frame temporal position matching and persistence determination are performed on the abnormal features to obtain temporally persistent abnormal features; Clustering and integrating the time-series persistent anomaly features yields stable defect clusters.

[0014] As a preferred implementation, multi-frame temporal position matching and persistence determination are performed on the abnormal features to obtain temporally persistent abnormal features, including: The extracted abnormal features are then calibrated for center coordinates and region range to complete cross-frame feature location matching. Based on the continuous occurrence and positional overlap of the aforementioned abnormal features, a temporal persistence determination is performed to obtain a persistence determination result. The survival determination results are filtered based on preset conditions to obtain time-series continuous abnormal characteristics.

[0015] As a preferred embodiment, consistency determination is performed on the stable defect cluster results to obtain batch quality inspection results, including: The stable defect cluster results are subjected to parameter extraction to obtain a defect feature parameter set; The defect feature parameter set is compared with the standard defect parameters of a preset qualified batch to obtain parameter deviation data; Based on the consistency determination of the parameter deviation data and the preset qualified threshold, the batch quality qualifiedness determination result is obtained. The batch quality conformity judgment results are integrated to obtain the batch quality inspection results.

[0016] As a preferred implementation, a minimum re-inspection sampling window is generated based on the batch quality inspection results. Re-inspection area planning is then performed on the minimum re-inspection sampling window to obtain re-inspection guidance results, including: Defect location marking is performed on the batch quality inspection results to obtain defect location information; The defect core area is delineated based on the defect location information, and the minimum re-inspection boundary is determined by combining the preset sampling density to obtain the minimum re-inspection sampling window; Based on the minimum re-inspection sampling window, the re-inspection sampling range is determined, the re-inspection area is planned, and the re-inspection guidance result is obtained.

[0017] The beneficial effects of this invention are: 1. By acquiring cross-polarized images and multi-exposure image sequences of basic cobalt carbonate powder, and combining them with a reference color chart and qualified batch samples, true color restoration is performed. At the same time, powder layer morphology correction and thickness interference compensation are performed on the true color image results. This solves the problem that online visual inspection is easily affected by residual moisture, light reflection, and powder layer thickness fluctuations, resulting in low detection accuracy and high false alarm rate. It effectively removes false anomaly signals and avoids misjudging non-quality defect areas as real defects. While improving detection accuracy and reducing false alarm rate, it provides a precise image basis for subsequent anomaly feature extraction, ensuring accurate identification of quality problems introduced by real agglomeration, uneven crystal growth, and impurities.

[0018] 2. By extracting three types of abnormal features—agglomeration, crystal texture, and color shift—from the thickness-corrected image results, and performing multi-frame temporal position matching and persistence determination on the abnormal features, and then clustering and integrating the temporally persistent abnormal features to obtain stable defect clusters, this method has the ability to overcome the limitations of existing detection methods that only provide single-frame and single-region abnormality indications, and achieve multi-dimensional integrated detection. It enables linkage between abnormality identification and defect confirmation, avoids interference from single-frame pseudo-abnormalities, ensures the stability of defect identification, and provides a reliable defect basis for batch quality judgment.

[0019] 3. By extracting parameters from stable defect clusters, comparing them with standard defect parameters, and determining consistency, batch quality inspection results are obtained. Based on these results, a minimum re-inspection sampling window is generated to achieve online real-time detection of basic cobalt carbonate powder. This avoids detection lag, allows for timely feedback on quality anomalies during production, reduces the scope of manual re-inspection, and improves re-inspection efficiency. Without increasing the frequency of offline chemical testing, it enhances the ability of online judgment to indicate the risk of uniformity in subsequent batches, thereby reducing production costs and quality control risks. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0021] Figure 1 A flowchart of a machine vision-based method for quality inspection of basic cobalt carbonate powder materials provided in an embodiment of the present invention.

[0022] Figure 2 This is a bar chart comparing the key performance indicators of the present invention with those of existing technologies. Detailed Implementation

[0023] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0024] Combination Figure 1 The following is an exemplary flowchart of a machine vision-based method for quality inspection of basic cobalt carbonate powder materials. The specific implementation steps are as follows: Cross-polarization images and multi-exposure image sequences of basic cobalt carbonate powder were acquired to obtain the original detection image set; In the continuous filtration, low-temperature drying, and graded discharge stages of the cobalt powder and cobalt chloride raw material synthesis production line for preparing basic cobalt carbonate, a planar imaging device and a switchable polarization light source are arranged 30cm above the powder conveying and distributing device to establish an online detection station, ensuring that the imaging area completely covers the width of the conveyor belt with no blind spots. The planar imaging device uses a resolution of 2048. The industrial area scan camera has 1536 pixels and a lens focal length of 25mm. The switchable polarization light source uses an LED polarization light source with a wavelength of 550nm and a polarization degree of ≥99%, ensuring the stability of polarization imaging.

[0025] Before testing, the imaging device is calibrated. The lens focal length is adjusted to clearly capture the details of the powder particles. Using a standard white board with a reflectivity of 99% as a reference, the polarizer angle is adjusted to 90°. The minimum reflected light intensity of the standard white board is used as the calibration target to complete the polarization imaging calibration. At the same time, two exposure levels are set. The high exposure parameter is used to capture the weak texture area of ​​the powder, and the low exposure parameter is used to capture the high reflectivity area of ​​the powder, avoiding overexposure of highlights or underexposure of shadows.

[0026] In the same detection area, three types of images are simultaneously acquired: one set of ordinary illumination images (without polarizer, exposure time 80μs); one set of cross-polarized images (polarizer angle 90°, exposure time 80μs); and one set of high and low exposure images (one frame each for high exposure 120μs and low exposure 30μs). Images are acquired every 0.5s, for a total of 100 frames, forming a raw detection image set containing multi-source visual information. The imaging resolution is determined based on the D50 of a qualified batch of basic cobalt carbonate. In this embodiment, the D50 of a qualified batch of basic cobalt carbonate is 15μm. One median particle size is set to correspond to 6 pixels, ensuring that each agglomerated particle contains at least 36 pixels, clearly distinguishing particle boundaries and internal textures, avoiding missed detection of small agglomerates due to insufficient resolution, and providing basic data for subsequent false anomaly removal and true color restoration. The area array imaging device and switchable polarization light source are both connected to the industrial controller signal. The industrial controller simultaneously acquires the conveyor belt encoder speed signal, achieving synchronization between image acquisition and material movement, ensuring accurate matching of adjacent frame image positions.

[0027] This step addresses the unique visual defects of battery-grade basic cobalt carbonate powder prepared from raw materials, such as strong surface reflection, large brightness-dark differences, and fine particles. The multi-source imaging scheme is designed to capture diffuse reflection information of the powder through cross-polarized images, compensate for brightness-dark differences through multi-exposure images, and use ordinary illumination images as a reference. The three work together to provide irreplaceable multi-source input for subsequent pseudo-anomaly removal.

[0028] Based on a reference color chart and qualified batch samples, the original detection image set is restored to its true color to obtain true color image results. Through reference calibration and multi-source image fusion, false anomalies caused by illumination fluctuations, residual moisture reflections, and exposure differences are eliminated. Specifically, the following steps are included: Collect images of white, gray, and black reference plates and images of qualified batch samples, perform reference calibration, and obtain reference data; Specifically, a baseline calibration operation is performed before each daily start-up test and during each batch of powder switching: white, gray, and black reference plates are placed in the conveyor belt detection area, and three sets of images of the three reference plates are acquired, with five frames in each set. After removing abnormal frames with grayscale value deviations exceeding ±5%, the average grayscale value and chromaticity value of each set of images are calculated as the brightness and color reference in the baseline reference data. Furthermore, the white plate has a reflectivity of 99%, the gray plate has a reflectivity of 50%, and the black plate has a reflectivity of 1%. At the same time, basic cobalt carbonate powder from the most recent 30 batches that have been confirmed by laser particle size analyzer, ICP-MS, and XRD to be qualified in terms of particle size, composition, and crystal phase are selected as qualified batch samples. Their images are acquired, and the color distribution range and average brightness value of the qualified batch samples are extracted to establish a qualified batch baseline color gamut, which is added to the baseline reference data. The baseline reference data is stored in the detection system in real time for subsequent calibration parameter calculations.

[0029] Based on the aforementioned reference data, reflection suppression and brightness compensation processing are performed on the cross-polarized image and the multi-exposure image to obtain calibration parameters, including: The cross-polarized image is subjected to reflection suppression processing to remove residual moisture and surface reflection interference, thereby obtaining reflection suppression features; Specifically, a polarization component analysis algorithm is used to process the cross-polarized image, separating the polarized reflective component in the image from the inherent color component of the powder itself. A reflective threshold is set, and pixels with brightness values ​​exceeding the reflective threshold are removed, while retaining the inherent color and texture information of the powder to obtain a reflective suppression feature. This feature is stored in the form of a gray-level matrix, with the matrix elements being the processed pixel gray-level values. The specific implementation steps of the polarization component analysis algorithm are as follows: First, perform a Fourier transform on the cross-polarized image to separate the linear polarization component and the circular polarization component; second, based on Malus's law, calculate the polarization extinction ratio when the polarizer angle is 90°, set the extinction ratio threshold to 100:1, and remove reflective pixels with an extinction ratio lower than this threshold; finally, perform Gaussian filtering on the remaining pixels to smooth the noise and obtain reflection suppression features. This algorithm is specifically adapted to the residual wet reflective characteristics of basic cobalt carbonate powder and can effectively remove surface specular reflection interference.

[0030] The multi-exposure image is subjected to brightness equalization processing to compensate for the difference between highlights and shadows, thus obtaining brightness compensation features; Furthermore, the Retinex image enhancement algorithm is used to fuse the high-exposure and low-exposure images: the high-exposure image is subjected to highlight compression to reduce the brightness of the highlight areas with a compression coefficient of 0.6 to avoid loss of details in overexposed areas; the low-exposure image is subjected to shadow brightening to increase the brightness of the shadow areas with a brightening coefficient of 1.4 to compensate for insufficient details in the shadow areas; the processed high-exposure and low-exposure images are then fused at the pixel level, and the average brightness of each pixel is calculated to obtain a uniform brightness compensation feature, which is also stored in the form of a grayscale matrix. The Retinex image enhancement algorithm is implemented using a single-scale Retinex algorithm, with the following parameters: standard deviation of the Gaussian wrap function σ = 80, gain coefficient of 1.0, and offset of 0; highlight compression uses a nonlinear compression function. Dark area brightening uses a linear stretching function. , The original grayscale values ​​are used to ensure uniform image brightness after brightness compensation, while preserving powder texture details. The reflection suppression feature and brightness compensation feature are normalized and fitted using the aforementioned benchmark reference data to obtain calibration parameters, including: The reflection suppression feature and the brightness compensation feature are respectively subjected to numerical normalization to obtain normalized feature data; Specifically, a normalization algorithm is used to map the gray values ​​(0-255) of the reflection suppression feature and the brightness compensation feature to the [0, 1] interval. The normalization formula is as follows: In the formula: These are the normalized eigenvalues; The original grayscale value; This is the minimum grayscale value of the feature; This is the maximum grayscale value of this feature.

[0031] After processing, normalized feature data is obtained, eliminating the dimensional differences between different features, which facilitates subsequent fitting calculations. Establishing a feature mapping relationship between the normalized feature data and the benchmark reference data, and fitting the feature weights, includes: Using the color and brightness standards in the benchmark reference data as a reference, feature matching is performed on the normalized feature data to obtain the feature correspondence; Furthermore, using the color and brightness standards in the benchmark reference data as a reference, feature matching is performed on the normalized feature data: the normalized reflection suppression features are matched with the color features of the qualified batch samples, and the normalized brightness compensation features are matched with the brightness benchmark of the benchmark reference data to establish a one-to-one correspondence between the two. Based on the aforementioned feature correspondence, linear fitting calculations are performed to obtain feature weights; Based on this feature correspondence, the least squares method is used for linear fitting calculation, and the fitting equation is: In the formula: The standard value is the reference value in the benchmark data; Normalized reflectivity suppression characteristics; This is a normalized brightness and darkness compensation feature; and The feature weights obtained from the fitting.

[0032] In this embodiment, the fitting result is... =0.65, =0.35, the feature weight is used for subsequent calibration parameter integration.

[0033] The preset correction threshold and the feature weights are integrated to generate calibration parameters; Specifically, a preset correction threshold is used based on the median ±3 of the color distribution of the qualified batch samples. MAD setting, wherein the calculation method of MAD is as follows: first calculate the median of the color distribution data of the qualified batch of samples, then calculate the absolute deviation of each data point from the median, and finally take the median of the absolute deviations. In this embodiment, the correction threshold for the red channel is [120, 150], the correction threshold for the green channel is [110, 140], and the correction threshold for the blue channel is [100, 130]. ​​The preset correction thresholds are combined with the fitted feature weights: =0.65, =0.35 is integrated to generate a calibration parameter set that includes color correction coefficient, brightness correction coefficient, reflection suppression threshold, and brightness compensation coefficient. The calibration parameter set is updated to the detection system in real time for the correction of the original detection image set.

[0034] Based on the calibration parameters, the original detection image set is corrected to obtain a true color image result; Furthermore, the color correction coefficients and brightness correction coefficients in the calibration parameter set are invoked to perform global color and brightness correction on each frame of the original detection image set: the RGB three color channels of the image are corrected separately to eliminate color deviations caused by uneven lighting; combined with the reflection suppression threshold and brightness compensation coefficient, residual wet reflections and exposure difference interference are further eliminated to restore the true color and texture of the powder itself, resulting in a true color image. In this embodiment, the color deviation of the true color image is controlled within ±3%, and the brightness uniformity error is ≤5%, ensuring the accuracy of subsequent feature extraction.

[0035] The true color image result is subjected to powder layer morphology correction and thickness interference compensation to obtain a thickness-corrected image result. The purpose is to separate the powder layer thickness difference from the material's inherent abnormalities, and to avoid powder layer thickness fluctuations being misjudged as quality anomalies. Specifically, it includes the following sub-steps: The true color image results are subjected to powder layer boundary contour extraction and morphological optimization to obtain an accurate powder layer boundary image; Furthermore, the Canny edge detection algorithm is used to extract the boundary contours of the powder layer in the true color image: Edge detection thresholds are set: low threshold 50, high threshold 150, to extract the boundary contours between the powder layer and the conveyor belt background; morphological optimization operations are employed: dilation followed by erosion, with a dilation kernel of 3... 3, Corrosion Core 3 3. Remove edge noise and discrete interference points, fill in the tiny gaps in the boundary contour, and obtain an accurate powder layer boundary image. The powder layer boundary extraction threshold is calibrated based on the grayscale difference between an empty conveyor belt and a fully loaded qualified sample. In this embodiment, the grayscale value of the empty conveyor belt is 200, the average grayscale value of the fully loaded qualified sample is 120, and the grayscale difference is 80. The edge detection threshold is set accordingly.

[0036] Based on the standard powder layer thickness parameters of qualified batch samples, thickness difference analysis is performed on the precise powder layer boundary image to obtain thickness deviation data. Furthermore, a laser thickness gauge is used to measure the standard powder layer thickness of qualified batch samples to obtain standard powder layer thickness parameters. Based on the accurate powder layer boundary image, the gray-level gradient analysis method is used to calculate the gray-level gradient value of local areas of the powder layer, and a mapping relationship between gray-level gradient and powder layer thickness is established: the larger the gray-level gradient, the thinner the powder layer thickness; the smaller the gray-level gradient, the thicker the powder layer thickness. Combined with the standard powder layer thickness parameters, the thickness deviation value of each local area is calculated: actual thickness - standard thickness, to obtain thickness deviation data. The thickness deviation data is stored in matrix form, and the thickness deviation of each area is marked.

[0037] In this embodiment, the local window size L is 8. The pixel width corresponding to D50 is 48 A local window of 48 pixels is used to perform partitioned analysis on the precise powder layer boundary image; when the thickness gradient of three consecutive windows changes in the same direction, it is determined to be a thickness fluctuation area rather than a direct defect area, thus avoiding misjudging continuous thickness fluctuations as quality anomalies.

[0038] Based on the thickness deviation data, thickness compensation and correction are performed on local areas of the powder layer to obtain thickness correction image results; Furthermore, for positive deviations in the thickness deviation data (areas where the actual thickness is greater than the standard thickness), grayscale suppression is applied: suppression coefficient = standard thickness / actual thickness, reducing the grayscale value in this area to simulate the grayscale performance under the standard thickness. For negative deviations (areas where the actual thickness is less than the standard thickness), grayscale enhancement is applied: enhancement coefficient = standard thickness / actual thickness, increasing the grayscale value in this area to eliminate the grayscale deviation caused by the thickness difference. For areas with thickness deviations within ±0.05mm, no compensation processing is performed, and the original grayscale information is retained. After thickness compensation correction, the thickness-corrected image result is obtained. This result eliminates the pseudo-anomalies caused by powder layer thickness fluctuations and only retains the quality anomaly information of the powder itself.

[0039] The abnormal features of the thickness-corrected image results are extracted, and the abnormal features are time-series tracked to obtain stable defect cluster results. Drawing on the concept of video continuous event tracking, instead of using single-frame anomalies as the final defect, stable anomaly clusters existing in consecutive frames are used as effective defect objects to avoid interference from instantaneous pseudo-anomalies. Specifically, multi-frame position matching and persistence determination based on the conveyor belt motion law are used to eliminate instantaneous dust, reflections, and other pseudo-anomalies. This step further includes the following steps: Anomaly detection is performed on the thickness-corrected image results, and abnormal features are extracted; Furthermore, threshold segmentation and feature extraction algorithms are used to detect abnormal regions in the thickness-corrected image results: Abnormality thresholds are set: based on the feature distribution of qualified batch samples, the clustering feature threshold is a connected region area > 50 pixels, the crystal texture feature threshold is a texture entropy > 0.8, and the color shift feature threshold is a true color deviation value > 5. Three types of abnormal features are extracted respectively. Agglomeration characteristics include connected region area, equivalent circle diameter, and boundary roughness, which are used to characterize powder agglomeration anomalies. The connected region area is calculated using the pixel counting method, the equivalent circle diameter is obtained by converting the connected region area, and the boundary roughness is calculated by the ratio of the number of boundary pixels to the circumference of the equivalent circle. Crystal texture features include local texture entropy, orientation consistency, and micro-edge density, which are used to characterize crystal growth non-uniformity anomalies. Local texture entropy is calculated using the gray-level co-occurrence matrix, orientation consistency is calculated using the edge orientation histogram, and micro-edge density is calculated using the ratio of the number of micro-edge pixels to the total area of ​​the region. Specifically, the formula for calculating local texture entropy is: In the formula: The number of gray levels is 256. grayscale value The probability of occurrence is based on 5 5. Gray-level co-occurrence matrix calculation (distance) =1, angle =0°, 45°, 90°, 135°, and take the average of the four angles as the final texture entropy). Formula for calculating directional consistency: In the formula: This refers to the number of pixels at the micro-edge. The direction angle of the i-th micro-edge (0°~180°); Micro-edge density calculation formula: In the formula: This refers to the number of pixels at the micro-edge. This represents the total area of ​​the local region (unit: pixels).

[0040] Color deviation characteristics: including true color deviation value and local dispersion, used to characterize color anomalies introduced by impurities. The true color deviation value is the Euclidean distance between the color of this area and the reference color gamut of the qualified batch, and the local dispersion is the standard deviation of the RGB three color channels of this area. The true color deviation value is calculated using the Euclidean distance in the RGB space, and the formula is as follows: In the formula: ( , , () is the center value of the reference color gamut for qualified batches. () represents the pixel value of the area to be tested.

[0041] Feature similarity is calculated using cosine similarity, and the formula is: In the formula: , These are the anomaly feature vectors of the preceding and following frames, respectively. After extracting the above three types of abnormal features, regional abnormal feature results are generated. Each abnormal region is numbered, and its feature parameters and location information are recorded. Simultaneously, a comprehensive index formula can be used to calculate the regional quality anomaly index, as follows: In the formula: This refers to the regional quality anomaly index. This refers to clustering anomalies (deviations of the area of ​​connected components from the standard value). This refers to the degree of grain disorder (the deviation of the texture entropy from the standard value). This is the true color deviation amount (true color deviation value). This is a timing persistence quantity (initially set to 0, subsequently updated based on the number of sustained frames); timing persistence quantity The update rule is: for every consecutive frame that appears in the abnormal region, Increase by 0.25 when the number of consecutive frames reaches (4 frames) =1.0, and for each additional consecutive frame thereafter, Increase by 0.1, with a maximum of 1.5, to ensure a reasonable contribution of time series persistence to the abnormal index; The weighting coefficients are determined through grid search. Specifically, the training set is based on historical qualified and unqualified batches, with the detection rate and false alarm rate as optimization objectives. The combination is traversed in the range of 0 to 1 with a step size of 0.05, and the group with the highest F1 score is selected as the final weight to ensure that the weights are adapted to the powder working conditions.

[0042] In this embodiment, the initial weight is set to... =0.35、 =0.25、 =0.20、 =0.20, when the production line changes the drying temperature range or the fabric speed, the weight is fine-tuned again; The threshold for judging regional quality anomalies is: connected component area > 50 pixels, texture entropy > 0.8, and true color deviation > 5. Based on the feature distribution statistics of 1500 frames from the most recent 30 batches of qualified samples, the maximum feature value of the qualified samples is taken + 3. The standard deviation is used as a threshold to ensure that 99% of qualified samples are not misjudged; Multi-frame temporal position matching and persistence determination are performed on the aforementioned abnormal features to obtain temporally persistent abnormal features. This step is used to eliminate transient pseudo-anomalies through cross-frame matching and persistence determination; including: The extracted abnormal features are then calibrated for center coordinates and region range to complete cross-frame feature location matching. Specifically, for each abnormal region, its center coordinates (x, y) and region range (width, height) are determined. Based on the conveyor belt speed and frame rate, the theoretical movement distance of the abnormal region between adjacent frames is calculated using the following formula: In the formula: The theoretical distance traveled by the abnormal region in adjacent frames, in mm / frame; The conveyor belt speed is expressed in mm / s. The frame rate of the imaging device is expressed in frames per second (fps).

[0043] In this embodiment, =800mm / s, =30fps, calculated ≈26.7mm / frame. The Hungarian algorithm is used to perform cross-frame feature location matching on abnormal regions in consecutive frames. The criterion is: the distance between the center coordinates of abnormal regions in two frames is < If the feature similarity is ±5mm and the cosine similarity of features such as aggregation, crystal texture, and color shift is >0.8, a successful match is determined to be a continuous frame performance of the same abnormal region. Based on the continuous occurrence and positional overlap of the aforementioned abnormal features, a temporal persistence determination is performed to obtain a persistence determination result. Furthermore, set a continuous frame rate threshold. =4 frames (if the transmission speed increases, then) The calculation is based on the principle of constant unit time, for example, if the speed increases to 1.0 m / s, (Adjusted to 5 frames), count the number of consecutive frames for each abnormal region; simultaneously calculate the positional overlap of abnormal regions in consecutive frames, using the following formula: In the formula: The degree of overlap of abnormal regions in consecutive frames; The overlapping area of ​​abnormal regions in two frames, in pixels; The total area of ​​the abnormal region, in pixels.

[0044] Set a position overlap threshold ≥ 0.7. If the number of consecutive frames in the abnormal area is ≥ If the position overlap is ≥0.7, it is determined to be a continuous anomaly feature in time series; otherwise, it is determined to be a transient pseudo-anomaly, and the survival determination result is obtained. The survival determination results are filtered based on preset conditions to obtain time-series continuous abnormal characteristics; Furthermore, preset filtering criteria: abnormal areas Value > 1.2 (based on qualified batch samples) Value distribution setting, qualified batch samples The value is ≤1.0), and the abnormal type (agglomeration, crystal texture, color deviation) is clear and there is no mixed interference; combined with the survival judgment result, the abnormal features that simultaneously meet the time-series continuous abnormal features and the preset screening conditions are selected to obtain the time-series continuous abnormal features, and the transient pseudo-abnormalities are eliminated.

[0045] Clustering and integrating the aforementioned time-series persistent anomaly features yields stable defect cluster results; Furthermore, the K-means clustering algorithm is used to spatially cluster and integrate the time-series persistent anomaly features: Based on the center coordinates and feature parameters of the anomaly region, a clustering distance threshold of 50 pixels is set. Time-series persistent anomaly features with a distance < 50 pixels and a feature similarity > 0.85 are clustered into a defect cluster, with each defect cluster corresponding to a continuous real quality defect region. Parameter statistics are performed on the clustered defect clusters, recording the location range, anomaly type, feature parameters, and duration of each defect cluster to obtain stable defect cluster results. In this embodiment, the stable defect clusters after clustering are those that appear consecutively for ≥ 4 frames and have a location overlap ≥ 0.7. The true defect area with a value > 1.2 is free from transient pseudo-anomaly interference.

[0046] The specific implementation steps of the K-means clustering algorithm are as follows: Initial cluster center selection: Using the elbow rule, based on the number of time-series continuous abnormal features, determine the number of clusters K=3~5 and dynamically adjust it in combination with the number of defects; Cluster distance calculation: Euclidean distance is used to calculate the distance between each outlier feature and the cluster center; Iteration termination condition: Stop iteration when the movement distance of the cluster center is less than 2 pixels, or when the number of iterations reaches 50. Post-clustering processing: Small clusters with fewer than 3 samples are removed, and these clusters are considered spurious anomalies. Valid defect clusters are retained to obtain stable defect cluster results.

[0047] The consistency of the stable defect cluster results is assessed to obtain batch quality inspection results. This step involves a two-tiered consistency assessment, encompassing both local and batch levels, simultaneously determining local anomalies and batch-level release criteria. Specifically, it involves comparing local defect parameters with the qualified batch standard library. The steps include: The stable defect cluster results are subjected to parameter extraction to obtain a defect feature parameter set; Specifically, for each stable defect cluster, its core parameters are extracted to form a set of defect feature parameters, including: defect type: agglomeration type, grain type, color deviation type, or mixed type; defect size: maximum diameter, area; defect density: number of defect clusters per unit area; defect distribution uniformity: standard deviation of the center coordinates of the defect cluster; and anomaly degree. The deviations of values, characteristic parameters, and standard values ​​are quantified and recorded for each parameter. For example, the maximum diameter of agglomerated defects is expressed in μm, and the defect density is expressed in units per cm³. 2 This ensures that the parameters are quantifiable and comparable. The defect feature parameter set is compared with the standard defect parameters of a preset qualified batch to obtain parameter deviation data; Specifically, a standard defect parameter library for pre-defined qualified batches is established. This library is based on the test data of the most recent 50 qualified samples and includes the allowable ranges for various types of defects: maximum diameter of agglomerated defects ≤ 50 μm, defect density ≤ 2 defects / cm². 2 For crystalline defects, the texture entropy should be ≤0.8 and the micro-edge density ≤0.1; for color-shifting defects, the true color deviation should be ≤5 and the local dispersion ≤3; for mixed defects, the allowable ranges of all types of defects must be met simultaneously. Each parameter in the defect feature parameter set is compared with the corresponding parameter in the standard defect parameter library to calculate the parameter deviation value: Parameter deviation value = Actual parameter - Standard parameter. The degree of deviation for each parameter is then labeled: Slight deviation: Deviation value within 10% of the allowable range; Moderate deviation: Deviation value between 10% and 30% of the allowable range; Severe deviation: Deviation value exceeding 30% of the allowable range. Based on the consistency determination of the parameter deviation data and the preset qualified threshold, the batch quality qualifiedness determination result is obtained. Specifically, the preset pass threshold is determined by selecting points on the ROC curve based on the results of the most recent 50 batches of test comparisons. In this embodiment, the pass threshold is set as follows: the number of slightly deviated parameters is ≤3, and there are no moderate or severe deviation parameters; if there are moderate deviation parameters and the number is ≤1, the batch is judged to be qualified but needs to be closely monitored; if there are severe deviation parameters or the number of moderate deviation parameters is ≥2, the batch is judged to be unqualified.

[0048] Based on parameter deviation data and compared with preset qualification thresholds, the qualification of local areas corresponding to each stable defect cluster is judged, and the overall qualification of the entire inspection batch is judged to obtain the batch quality qualification result, including the qualification of local areas and the overall batch qualification.

[0049] It should be noted that when both agglomeration-type and color deviation-type anomalies exceed the threshold in the same detection window, the batch is prioritized as a high-risk batch and must be immediately suspended from discharge for further re-inspection to avoid the generation of unqualified batches. The batch quality conformity judgment results are integrated to obtain the batch quality inspection results; Furthermore, the batch quality compliance judgment results are summarized and integrated to generate complete batch quality inspection results. These results include at least: Quality grade results: qualified, qualified and under key monitoring, unqualified, and high risk; Anomaly cause results: clarifying the cause of each defect cluster, such as agglomeration anomalies caused by impurities in cobalt powder / cobalt chloride raw materials, and color deviation anomalies caused by fluctuations in the synthesis mother liquor ratio; Batch uniformity results: a score for the uniformity of defect distribution within the batch, ranging from 1 to 10, with 10 being the most uniform; Local defect distribution information: the location coordinates, size, and type of defect clusters. The batch quality inspection results are output in report form for easy viewing and subsequent processing by staff.

[0050] Based on the batch quality inspection results, a minimum re-inspection sampling window is generated. Re-inspection area planning is then performed on this minimum re-inspection sampling window to obtain re-inspection guidance results. Specifically, this involves dynamically defining the minimum sampling range based on the actual location of the defect, rather than using a fixed proportion of sampling or sampling the entire batch. This step links local defects with the re-inspection range, reducing the scope of manual re-inspection and improving re-inspection efficiency. The specific steps include: Defect location marking is performed on the batch quality inspection results to obtain defect location information; Specifically, based on the local defect distribution information in the batch quality inspection results, the location of each stable defect cluster is precisely marked: A two-dimensional coordinate system is established with the starting end of the conveyor belt as the origin: the x-axis represents the length of the conveyor belt, and the y-axis represents the width of the conveyor belt. The center coordinates (x, y) and the area range (length, width) of each defect cluster are marked. Simultaneously, combining the conveyor belt speed and the duration of the defect cluster, the length of the material segment corresponding to the defect cluster is deduced. The calculation formula is as follows: In the formula: The length of the material segment corresponding to the defect cluster, in mm; The conveyor belt speed is expressed in mm / s. The duration of the defect cluster, in seconds.

[0051] Complete defect location information is obtained, accurate to ±10mm, ensuring precise location during re-inspection.

[0052] The defect core area is delineated based on the defect location information, and the minimum re-inspection boundary is determined by combining the preset sampling density to obtain the minimum re-inspection sampling window; Furthermore, using the defect cluster area in the defect location information as the core, a defect core area is defined: core area range = defect cluster area + 10mm extension; the preset sampling density is 5 samples / cm², based on the sampling standard of qualified batch samples to ensure sampling representativeness. Combining this with the area of ​​the defect core area, the required number of sampling points is calculated: number of sampling points = core area. Sampling density; Simultaneously, a sampling window expansion coefficient is set, preferably between 1.2 and 1.8. The determination of the expansion coefficient of 1.2 to 1.8 is based on experimental verification of 10 batches of different anomaly types: for 5 batches of clustered anomalies and 5 batches of color-shifted anomalies, the re-inspection range is defined using an expansion coefficient of 1.2 to 1.8 respectively. The re-inspection results are compared with the offline full inspection results. When the expansion coefficient of clustered anomalies is ≥1.6, the re-inspection coverage rate reaches more than 98%; when the expansion coefficient of color-shifted anomalies is ≥1.2, the re-inspection coverage rate reaches more than 97%. The above expansion coefficient range is determined while taking into account the re-inspection efficiency. When the anomaly is caused by aggregation, the expansion coefficient is high (1.6~1.8). In this embodiment, the expansion coefficient for aggregation-type defects is 1.7. When the anomaly is caused by local color reversal or slight texture anomalies, the expansion coefficient is low (1.2~1.4). In this embodiment, the expansion coefficient for color-shifting defects is 1.3. Taking the core area of ​​the defect as the center, the expansion coefficient is used to determine the minimum re-inspection boundary. The boundary range must cover all core areas of the defects and the surrounding areas that may be affected, forming the minimum re-inspection sampling window. The coordinate range, number of sampling points, and sampling position of the sampling window are marked.

[0053] Based on the minimum re-inspection sampling window, the re-inspection sampling range is determined, the re-inspection area is planned, and the re-inspection guidance result is obtained. Furthermore, based on the minimum re-inspection sampling window, the re-inspection sampling range is defined, i.e., the coordinate range of the sampling window, and the re-inspection area is planned: the sampling window is divided into several re-inspection sub-regions, each with a size of 50mm. For a 50mm defect, prioritize areas with severe deviations, followed by areas with moderate deviations, and then areas with slight deviations. Determine the re-inspection priority for these areas. Specify the sampling points and quantity for each re-inspection sub-area (≥3 sampling points per sub-area), the sampling method (using a sterile sampling spoon, ≥5g / point), and the re-inspection requirements (e.g., ICP testing or laser particle size analysis after sampling, with testing indicators including impurity content and particle size distribution). Integrate this information to generate re-inspection guidance results, output in text and graph format. Staff can accurately complete the re-inspection operation based on the guidance, eliminating the need for comprehensive sampling of the entire batch, significantly improving re-inspection efficiency and reducing costs.

[0054] To verify the technical advantages of the method of the present invention compared with existing conventional detection methods, 100 batches of parallel comparative tests were conducted on the same cobalt powder / cobalt chloride synthesis production line for preparing basic cobalt carbonate.

[0055] The method of this invention employs a complete process of cross-polarization, multi-exposure imaging, true color restoration, thickness compensation, time-series tracking, and minimum re-inspection window; Existing technologies use ordinary industrial cameras with single-threshold visual recognition, but lack polarization, exposure fusion, temporal filtering, thickness correction, and dynamic re-inspection planning.

[0056] In the comparative experiment, the laser particle size analyzer used was a Malvern Mastersizer 3000 model, with a detection range of 0.01μm~3000μm; the ICP-MS used was an Agilent 7900 model, with a detection accuracy of ≤0.01ppb; and the XRD used was a Bruker D8 Advance model, with a scanning range of 2θ=10°~80° and a scanning speed of 5° / min. All the above equipment was calibrated according to industry standards to ensure the accuracy of the test results. The comparison results are as follows: Table 1. Comparison of Detection Effects between the Invention and Existing Technologies Based on Table 1 above, it can be seen that the present invention effectively solves the technical problems of numerous false anomalies, high false alarms, numerous missed detections, large re-inspection range, and delayed results in online detection of basic cobalt carbonate powder by combining technologies such as cross-polarization de-reflection, multi-exposure brightness and darkness equalization, powder layer thickness compensation, time-series multi-frame tracking, and dynamic minimum re-inspection window. It has outstanding substantive features and significant progress.

[0057] Combination Figure 2 The bar chart comparing the key performance indicators of the present invention with those of the prior art shows that: Defect detection rate: The defect detection rate of the method of the present invention is 96.5%, while that of the existing conventional detection method is 68.2%. The present invention improves the detection rate by 28.3 percentage points compared with the prior art, which significantly improves the ability to identify real defects. False alarm rate: The false alarm rate of the method of this invention is only 5.3%, while the false alarm rate of the prior art is as high as 42.7%. This invention reduces the false alarm rate by about 37.4 percentage points through pseudo-anomaly stripping, a reduction of 88%. Re-inspection sampling ratio: The method of the present invention only requires re-inspection sampling of 12.8% of the material area, while the prior art requires sampling of 100% of the material in the whole batch. The re-inspection range of the present invention is reduced by 87.2%, which greatly improves the re-inspection efficiency and reduces the testing cost.

[0058] The above comparison results clearly demonstrate that the method of the present invention is significantly superior to the prior art in terms of detection accuracy, false alarm suppression, and re-inspection efficiency, possessing outstanding substantive features and significant progress.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A machine vision-based method for detecting the quality of basic cobalt carbonate powder material, characterized in that, include: Cross-polarization images and multi-exposure image sequences of basic cobalt carbonate powder were acquired to obtain the original detection image set; Based on the reference color chart and qualified batch samples, the original detection image set is restored to true color to obtain true color image results; The true color image result is subjected to powder layer morphology correction and thickness interference compensation to obtain a thickness-corrected image result; Extract the abnormal features of the thickness correction image results, perform time-series tracking on the abnormal features, and obtain stable defect cluster results; The consistency of the stable defect cluster results is determined to obtain the batch quality inspection results; Based on the batch quality inspection results, a minimum re-inspection sampling window is generated, and the re-inspection area is planned within the minimum re-inspection sampling window to obtain the re-inspection guidance results. Based on the reference color chart and qualified batch samples, the original detection image set is restored to its true color to obtain true color image results, including: Collect images of white, gray, and black reference plates and images of qualified batch samples, perform reference calibration, and obtain reference data; Based on the aforementioned reference data, reflection suppression and brightness compensation processing are performed on the cross-polarized image and the multi-exposure image to obtain calibration parameters; Based on the calibration parameters, the original detection image set is corrected to obtain a true color image result; Based on the aforementioned reference data, reflection suppression and brightness compensation processing are performed on the cross-polarized image and the multi-exposure image to obtain calibration parameters, including: The cross-polarized image is subjected to reflection suppression processing to remove residual moisture and surface reflection interference, thereby obtaining reflection suppression features; The multi-exposure image is subjected to brightness equalization processing to compensate for the difference between highlights and shadows, thus obtaining brightness compensation features; The reflection suppression feature and brightness compensation feature are normalized and fitted using the reference data to obtain calibration parameters. The true color image result is subjected to powder layer morphology correction and thickness interference compensation to obtain a thickness-corrected image result, including: The true color image results are subjected to powder layer boundary contour extraction and morphological optimization to obtain an accurate powder layer boundary image; Based on the standard powder layer thickness parameters of qualified batch samples, thickness difference analysis is performed on the precise powder layer boundary image to obtain thickness deviation data. Based on the thickness deviation data, thickness compensation and correction are performed on local areas of the powder layer to obtain thickness correction image results; Extracting anomalous features from the thickness-corrected image results, and performing time-series tracking on these anomalous features to obtain stable defect cluster results, including: Anomaly detection is performed on the thickness-corrected image results, and abnormal features are extracted; Multi-frame temporal position matching and persistence determination are performed on the abnormal features to obtain temporally persistent abnormal features; Clustering and integrating the time-series persistent anomaly features yields stable defect clusters.

2. The method for detecting the quality of the basic cobalt carbonate powder material according to claim 1, characterized in that, The reflection suppression feature and brightness compensation feature are normalized and fitted using the aforementioned benchmark reference data to obtain calibration parameters, including: The reflection suppression feature and the brightness compensation feature are respectively subjected to numerical normalization to obtain normalized feature data; The normalized feature data and the benchmark reference data are used to establish a feature mapping relationship, and the feature weights are obtained by fitting. The preset correction threshold and the feature weights are integrated to generate calibration parameters.

3. The method for detecting the quality of the basic cobalt carbonate powder material according to claim 2, characterized in that, Establishing a feature mapping relationship between the normalized feature data and the benchmark reference data, and fitting the feature weights, includes: Using the color and brightness standards in the benchmark reference data as a reference, feature matching is performed on the normalized feature data to obtain the feature correspondence; Based on the aforementioned feature correspondence, linear fitting calculations are performed to obtain feature weights.

4. The method for detecting the quality of the basic cobalt carbonate powder material according to claim 3, characterized in that, Multi-frame temporal position matching and persistence determination are performed on the aforementioned abnormal features to obtain temporally persistent abnormal features, including: The extracted abnormal features are then calibrated for center coordinates and region range to complete cross-frame feature location matching. Based on the continuous occurrence and positional overlap of the aforementioned abnormal features, a temporal persistence determination is performed to obtain a persistence determination result. The survival determination results are filtered based on preset conditions to obtain time-series continuous abnormal characteristics.

5. The method for detecting the quality of basic cobalt carbonate powder material according to claim 1, characterized in that, The consistency of the stable defect cluster results is determined to obtain batch quality inspection results, including: The stable defect cluster results are subjected to parameter extraction to obtain a defect feature parameter set; The defect feature parameter set is compared with the standard defect parameters of a preset qualified batch to obtain parameter deviation data; Based on the consistency between the parameter deviation data and the preset qualified threshold, the batch quality qualifiedness judgment result is obtained. The batch quality conformity judgment results are integrated to obtain the batch quality inspection results.

6. The method for detecting the quality of the basic cobalt carbonate powder material according to claim 1, characterized in that, Based on the batch quality inspection results, a minimum re-inspection sampling window is generated. Re-inspection area planning is then performed on the minimum re-inspection sampling window to obtain re-inspection guidance results, including: Defect location information is obtained by marking the defect location in the batch quality inspection results. The defect core area is delineated based on the defect location information, and the minimum re-inspection boundary is determined by combining the preset sampling density to obtain the minimum re-inspection sampling window; Based on the minimum re-inspection sampling window, the re-inspection sampling range is determined, the re-inspection area is planned, and the re-inspection guidance result is obtained.

Images