Method and System for Detecting the Operating Condition of Automatic Discharge Valve in Drying Kettle Based on Multi-view Image Fusion

Abstract

A method and system for detecting the operating condition of an automatic discharge valve in a drying autoclave using multi-view image fusion is disclosed. This method relates to data processing systems or methods specifically applicable to administrative, commercial, financial, management, supervisory, or predictive purposes. In this method, after the discharge valve of the drying autoclave is opened, multi-view time-series images of the valve body surface are acquired. Geometric transformation parameters are calculated based on rigid texture features in a reference frame, and the image sequence is spatially registered. A thermal real-world image is generated through temporal statistical filtering, and thermal expansion compensation is performed on the cold-state baseline image based on real-time temperature data. The thermal real-world image and the thermal expansion compensation image are then subjected to difference operations and connected component analysis. When the connected area exceeds a threshold, foreign matter residue is determined to be present. This application aims to improve the accuracy and stability of identifying minute residues, thereby improving the overall efficiency of production batch turnover.

Description

Technical Field

[0001] This application belongs to the field of data processing systems or methods specifically applicable to administrative, commercial, financial, management, supervision or forecasting purposes, and in particular relates to a method and system for detecting the operating conditions of an automatic discharge valve of a drying kettle using multi-view image fusion. Background Technology

[0002] In the pharmaceutical raw material drying and packaging process, the vertical vacuum drying autoclave is the core equipment, which achieves material drying through jacket heating and an internal spiral agitator. Unlike rotary equipment, the discharge valve (usually a hemispherical valve or bottom valve) of the drying autoclave is located at the bottom of the fixed cylinder, serving as the key channel for material collection and discharge. After batch production is completed, the cleanliness of the sealing surface of the discharge valve (the contact surface between the valve core and the valve seat) needs to be tested to ensure that there is no residual powder or cleaning droplets from the previous batch, thereby ensuring the purity of the drug and the reliability of the seal.

[0003] To meet the high standards of aseptic production, related technologies typically employ machine vision systems to assist manual inspection. Specifically, an industrial camera is deployed below the discharge valve, using a light source at a specific angle to image the valve's sealing surface after it is opened. Image processing algorithms are then used to identify residues or surface defects. Under ideal conditions where the equipment is at room temperature, this technology can effectively identify minute foreign objects, contributing to improved standardization and objectivity in the inspection process.

[0004] However, in actual continuous production cycles, the massive metal body of the drying autoclave often retains residual heat during the discharge stage. Since the discharge valve is located at the bottom of the equipment, when the valve is opened for detection, the hot airflow inside the autoclave and the external cold air create intense convection at the valve opening. When the imaging optical path passes through the turbulent layer with uneven density, the refractive index of the light dynamically changes, causing slight jitter or blurring in the acquired image of the sealing surface. This thermal disturbance increases the difficulty of image feature extraction, reduces the accuracy and stability of identifying minute residues, and may consequently reduce the overall efficiency of production batch turnover. Summary of the Invention

[0005] This application provides a method and system for detecting the working condition of an automatic discharge valve in a drying kettle using multi-view image fusion, which improves the accuracy and stability of identifying minute residues, thereby enhancing the overall efficiency of production batch turnover.

[0006] In the first aspect, this application provides a multi-view image fusion method for detecting the working condition of the automatic discharge valve of the drying kettle. Under the high temperature state after the discharge valve of the drying kettle is opened and the discharge action is completed, multi-frame time-series images of the valve body surface of the discharge port from multiple perspectives are continuously acquired.

[0007] Identify the rigid texture features of the valve body surface within a reference frame in a multi-frame temporal image;

[0008] Using the rigid texture features of the valve body surface in the reference frame as a benchmark, calculate the geometric transformation parameters of the other time-series images relative to the reference frame;

[0009] Spatial registration is performed on multiple frames of temporal images based on geometric transformation parameters to generate a spatially aligned corrected image sequence;

[0010] Temporal statistical filtering is performed on the grayscale value of the same pixel coordinate point in the time axis direction in the corrected image sequence to obtain the thermal real scene image;

[0011] Acquire real-time temperature data at the outlet of the drying kettle and pre-stored cold-state reference diagram of the valve body;

[0012] Based on real-time temperature data and preset valve body material thermal expansion coefficient, the cold state reference image of the valve body is scaled and deformed to generate a thermal expansion compensation image that matches the size of the hot state real scene image.

[0013] A grayscale difference image is obtained by performing a difference operation between the thermal real-world image and the thermal expansion compensation image;

[0014] Perform connectivity analysis on the grayscale difference graph and calculate the connectivity area of ​​non-zero regions;

[0015] If the connected area exceeds a preset threshold, it is determined that there is foreign object residue.

[0016] By employing the above technical solution, multi-view, multi-frame time-series images are continuously acquired under high-temperature conditions after the discharge valve of the drying kettle is opened. Spatial registration is then performed based on the rigid texture features of the valve body surface, effectively reducing interference from viewpoint differences and jitter during image acquisition. Temporal statistical filtering is used to generate a thermal real-world image, which, combined with real-time temperature data, performs thermal expansion compensation on the cold-state baseline image, reducing the impact of valve body deformation caused by temperature changes on the detection results. Differential operations and connected component analysis between the thermal real-world image and the thermal expansion compensation image improve the detection accuracy of foreign matter residues on the valve body surface and reduce the probability of false detections caused by temperature changes, viewpoint differences, and other factors. This detection method based on multi-view image fusion and thermal compensation enhances the reliability and adaptability of foreign matter residue detection in the automatic discharge valve of the drying kettle, thereby improving the overall efficiency of production batch turnover.

[0017] In conjunction with some implementations of the first aspect, in some implementations, identifying the rigid texture features of the valve body surface within a reference frame in a multi-frame temporal image specifically includes:

[0018] Calculate the grayscale gradient magnitude of each pixel in the reference frame in the horizontal and vertical directions;

[0019] Regions with grayscale gradient magnitudes less than a preset gradient filtering threshold are marked as flat regions; regions with grayscale gradient magnitudes not less than a preset gradient filtering threshold are marked as high-frequency texture regions.

[0020] Isolated noise regions with fewer than a preset noise area threshold connected pixels in high-frequency texture regions are removed to obtain the rigid texture features of the valve body surface.

[0021] By employing the aforementioned technical solution, the horizontal and vertical grayscale gradient magnitudes of pixels in the reference frame are calculated. Combined with a preset gradient filtering threshold, flat regions and high-frequency texture regions are divided, while small, isolated noise areas are eliminated, thus improving the extraction quality of rigid texture features on the valve body surface. This feature extraction method based on gradient analysis and region filtering reduces the interference of image noise on feature recognition and enhances the representativeness and stability of texture features. Through a reasonable feature filtering mechanism, this method improves the accuracy of subsequent image registration, thereby increasing the accuracy of foreign object detection. The feature extraction method exhibits strong robustness, enabling the detection system to better adapt to the imaging characteristics of the valve body surface under different operating conditions.

[0022] In conjunction with some implementations of the first aspect, in some implementations, identifying the rigid texture features of the valve body surface within a reference frame in a multi-frame temporal image specifically includes:

[0023] Several non-overlapping rectangular detection sub-windows are divided on the reference frame;

[0024] Calculate the variance statistic of pixel grayscale values ​​within each rectangular detection sub-window;

[0025] Rectangular detection sub-windows whose variance statistics fall within a preset texture saliency range are selected as candidate feature blocks;

[0026] Calculate the spatial distance between each candidate feature block, and remove redundant candidate feature blocks whose spatial distance is less than a preset overlap suppression threshold to obtain alternative feature blocks;

[0027] The candidate feature blocks distributed at the edge contour of the valve body are identified as the rigid texture features of the valve body surface.

[0028] By employing the above technical solution, the rigid texture features of the valve body surface are identified through rectangular detection sub-window division and variance statistics analysis. Combined with a redundant feature suppression mechanism based on preset texture saliency intervals and spatial distance constraints, the reliability of feature extraction is improved. By selecting feature blocks distributed along the valve body edge contour as the final rigid texture features, the discriminativeness and stability of feature points are enhanced. This feature extraction method reduces the impact of local illumination variations and noise on feature recognition, improving the accuracy of feature matching. Features extracted using this method exhibit good repeatability, improving the accuracy and reliability of image registration and thus enhancing the system's detection performance.

[0029] In conjunction with some embodiments of the first aspect, in some embodiments, after determining that foreign matter residue exists, the method further includes:

[0030] Based on the grayscale difference map, connected regions with grayscale values ​​greater than a preset difference threshold are extracted as suspected foreign object regions;

[0031] Using edge detection operators, the first gradient direction angle of each pixel in the suspected foreign object region of the thermal real-world image and the second gradient direction angle of the corresponding pixel in the suspected foreign object region of the thermal expansion compensation image are calculated respectively.

[0032] Calculate the cosine average of the angle differences between the first gradient direction angle and the second gradient direction angle of all pixels in the suspected foreign object region;

[0033] The cosine mean value is used as the texture structure similarity of the suspected foreign object region;

[0034] If the texture structure similarity is greater than the preset consistency threshold, the suspected foreign object area is determined to be an artifact area caused by the nonlinear deformation of the valve body.

[0035] Remove artifact regions from foreign object residue.

[0036] By employing the aforementioned technical solution, and analyzing the gradient direction angles of pixels within the suspected foreign object region, the textural similarity between the thermal real-world image and the thermal expansion compensation image in that region can be calculated. This allows for the differentiation between artifact regions caused by nonlinear deformation of the valve body and actual foreign object residues. This analysis method based on local texture structure similarity reduces the interference of valve body nonlinear deformation on the detection results and improves the accuracy of foreign object detection. By introducing comparative analysis of gradient direction angles, this method enhances the detection algorithm's ability to distinguish between valve body deformation and foreign object residues, reducing false alarms caused by valve body thermal deformation. This artifact region identification and removal mechanism improves the reliability of the detection results.

[0037] In conjunction with some implementations of the first aspect, in some implementations, the cosine average of the angle difference between the first gradient direction angle and the second gradient direction angle of all pixels within the suspected foreign object region is calculated, specifically including:

[0038] Using edge detection operators, the gradient magnitude of each pixel in the suspected foreign object area of ​​the thermal real-world image is calculated;

[0039] Calculate the cosine of the difference between the first gradient direction angle and the second gradient direction angle for each pixel;

[0040] The weighted similarity is obtained by summing the products of the gradient magnitude and cosine value of all pixels in the suspected foreign object region.

[0041] Calculate the sum of gradient magnitudes of all pixels within the suspected foreign object region to obtain the total weight;

[0042] Calculate the ratio of the weighted sum of similarities to the sum of weights, and determine the ratio as the cosine average.

[0043] By employing the aforementioned technical solution, a gradient magnitude-weighted texture structure similarity calculation method is constructed by calculating the gradient magnitude of each pixel within the suspected foreign object region and multiplying it by the cosine of the gradient direction angle difference as a weight. This calculation method ensures that edge pixels with larger gradient magnitudes receive higher weights in the similarity calculation, while the influence of pixels in flat areas with smaller gradient magnitudes is appropriately reduced. Since edge regions often contain more texture detail features, this weighted averaging strategy improves the sensitivity of the texture structure similarity index to local deformation features. By using the ratio of the weighted similarity sum to the weight sum as the final cosine average, a normalized similarity metric is established, reducing the influence of the size of the suspected foreign object region on the similarity calculation results and improving the stability of the texture structure similarity index.

[0044] In some embodiments, in conjunction with the first aspect, after removing the artifact region from the foreign matter residue, the method further includes:

[0045] When the texture structure similarity is determined to be no greater than the preset consistency threshold, low-pass filtering is performed on the image data of the suspected foreign object region and the image data of the thermal expansion compensation map at the corresponding positions to obtain the first low-frequency feature map and the second low-frequency feature map respectively.

[0046] Calculate the structural trend similarity between the first low-frequency feature map and the second low-frequency feature map;

[0047] If the structural trend similarity is greater than the preset trend tolerance threshold, the suspected foreign object area is determined to be an artifact caused by thermal deformation of the valve body's three-dimensional structure, and the foreign object attribute of the suspected foreign object area is excluded.

[0048] By employing the aforementioned technical solution, low-pass filtering is applied to the suspected foreign object area and the corresponding area of ​​the thermal expansion compensation map, suppressing high-frequency detail fluctuations in the image while preserving the overall grayscale distribution trend characteristics. By comparing the structural trend similarity between low-frequency feature maps, a detection mechanism for large-scale grayscale change patterns caused by thermal deformation is established. When the texture structure similarity is low but the structural trend similarity is high, it indicates that although there are local detail differences in the area, the overall grayscale change trend is consistent, consistent with the characteristic behavior of non-uniform deformation of the valve body's three-dimensional structure under heating. This discrimination method based on multi-scale feature analysis improves the system's accuracy in distinguishing between real foreign object residues and thermal deformation artifacts, and reduces the false alarm rate of the detection results.

[0049] In conjunction with some implementations of the first aspect, in some implementations, the structural trend similarity between the first low-frequency feature map and the second low-frequency feature map is calculated, specifically including:

[0050] Obtain the gray values ​​of all pixels in the first low-frequency feature map to construct a first gray-level matrix, and obtain the gray values ​​of all pixels in the second low-frequency feature map to construct a second gray-level matrix;

[0051] The correlation coefficient between the first gray matrix and the second gray matrix is ​​calculated using the normalized cross-correlation algorithm, and the correlation coefficient is determined as the structural trend similarity.

[0052] By employing the aforementioned technical solution and using a normalized cross-correlation algorithm to calculate the correlation coefficient of gray-level matrices between low-frequency feature maps, a quantitative evaluation method for structural trend similarity is established. The normalized cross-correlation algorithm is insensitive to overall brightness and contrast changes in the image, primarily reflecting the similarity of gray-level spatial distribution patterns. By constructing the pixel gray-level values ​​of the first and second low-frequency feature maps into matrix forms, the two-dimensional spatial structural information of the image is preserved. This correlation analysis method based on global gray-level distribution improves the robustness of the structural trend similarity index and reduces the interference of local noise and illumination inhomogeneity on similarity calculation.

[0053] Secondly, embodiments of this application provide a multi-view image fusion automatic discharge valve condition detection system for a drying kettle. This multi-view image fusion automatic discharge valve condition detection system for a drying kettle includes: one or more processors and a memory; the memory is coupled to one or more processors, and the memory is used to store computer program code, which includes computer instructions. One or more processors call the computer instructions to cause the system to execute a multi-view image fusion automatic discharge valve condition detection method for a drying kettle as described in the first aspect and any possible implementation thereof.

[0054] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a system, cause the system to perform a multi-view image fusion method for detecting the operating condition of an automatic discharge valve in a drying kettle, as described in the first aspect and any possible implementation thereof.

[0055] Fourthly, embodiments of this application provide a computer program product that, when running on a system, causes the system to execute a multi-view image fusion method for detecting the operating condition of an automatic discharge valve in a drying kettle, as described in any possible implementation of the first aspect.

[0056] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0057] 1. This application provides a multi-view image fusion method for detecting the operating condition of an automatic discharge valve in a drying kettle. By continuously acquiring multi-view, multi-frame time-series images under high-temperature conditions after the discharge valve of the drying kettle is opened, and performing spatial registration based on the rigid texture features of the valve body surface, interference caused by viewpoint differences and jitter during image acquisition can be effectively reduced. Temporal statistical filtering is used to generate a thermal real-world image, and thermal expansion compensation is performed on the cold-state reference image using real-time temperature data, which can reduce the impact of valve body deformation caused by temperature changes on the detection results. Through differential operations and connected component analysis between the thermal real-world image and the thermal expansion compensation image, the detection accuracy of foreign matter residue on the valve body surface is improved, and the probability of false detection caused by factors such as temperature changes and viewpoint differences is reduced. This detection method based on multi-view image fusion and thermal compensation improves the reliability and adaptability of foreign matter residue detection in the automatic discharge valve of the drying kettle, thereby improving the overall efficiency of production batch turnover.

[0058] 2. This application provides a multi-view image fusion method for detecting the operating condition of an automatic discharge valve in a drying reactor. By analyzing the gradient direction angle of pixels within a suspected foreign object region and calculating the texture structure similarity between the thermal real-world image and the thermal expansion compensation image in that region, it can distinguish between artifact regions caused by nonlinear deformation of the valve body and actual foreign object residues. This analysis method based on local texture structure similarity reduces the interference of valve body nonlinear deformation on the detection results and improves the accuracy of foreign object detection. By introducing comparative analysis of gradient direction angles, this method enhances the detection algorithm's ability to distinguish between valve body deformation and foreign object residues, reducing false alarms caused by valve body thermal deformation. This artifact region identification and removal mechanism improves the reliability of the detection results.

[0059] 3. This application provides a multi-view image fusion method for detecting the operating conditions of an automatic discharge valve in a drying vessel. Low-pass filtering is applied to the suspected foreign object area and the corresponding area of ​​the thermal expansion compensation map to suppress high-frequency detail fluctuations in the image while preserving the overall grayscale distribution trend. By comparing the structural trend similarity between low-frequency feature maps, a detection mechanism for large-scale grayscale change patterns caused by thermal deformation is established. When the texture structure similarity is low but the structural trend similarity is high, it indicates that although there are local detail differences in the area, the overall grayscale change trend is consistent, which is consistent with the characteristic behavior of non-uniform deformation of the valve body's three-dimensional structure under heating. This discrimination method based on multi-scale feature analysis improves the accuracy of the system in distinguishing between real foreign object residues and thermal deformation artifacts, and reduces the false alarm rate of the detection results. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a method for detecting the operating conditions of an automatic discharge valve in a drying kettle using multi-view image fusion, as described in an embodiment of this application.

[0061] Figure 2 This is another flowchart illustrating a multi-view image fusion method for detecting the operating conditions of an automatic discharge valve in a drying kettle, as described in an embodiment of this application.

[0062] Figure 3 This is a schematic diagram of the physical device structure of a multi-view image fusion automatic discharge valve detection system for a drying kettle, provided in an embodiment of this application. Detailed Implementation

[0063] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0064] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0065] The following example is used in conjunction with Figure 1 This application describes a method for detecting the operating condition of an automatic discharge valve in a drying kettle using multi-view image fusion in an embodiment of the present application:

[0066] Please see Figure 1 This is a flowchart illustrating a method for detecting the operating conditions of an automatic discharge valve in a drying kettle using multi-view image fusion, as described in an embodiment of this application.

[0067] S101. Under high temperature conditions after the discharge valve of the drying kettle is opened and the discharge action is completed, continuously acquire multi-frame time-series images of the valve body surface of the discharge port from multiple perspectives.

[0068] The discharge valve of the drying kettle is a mechanical device used to control the discharge of materials from the drying equipment. Its high-temperature state after opening and completing the discharge action refers to the physical state where the valve surface has not yet cooled down and remains close to the drying process temperature at the moment of material discharge or for a very short time afterward. Multi-view refers to observing the same target through imaging devices at different spatial positions or angles; multi-frame time-series images refer to a sequence of images taken sequentially in chronological order. In this step, the acquisition equipment needs to operate in a high-temperature environment, possibly accompanied by residual steam or dust. The specific type of imaging equipment is not limited; industrial-grade CCD cameras, CMOS cameras, or infrared thermal imagers can be used. For thresholds involving multiple types, such as exposure time threshold and gain threshold, dynamic adjustments are needed based on the ambient light intensity and the reflective characteristics of the valve surface to ensure the image brightness is within the linear response range. Specifically, this step aims to capture the surface morphology of the valve body at its closest operating limit, where the valve body is at or near the peak of thermal expansion, and any attached foreign matter is also under heat. The continuous acquisition process needs to ensure sufficient temporal resolution to capture any airflow disturbances or minute vibrations that may exist on the valve body surface, providing a data foundation for subsequent image registration and denoising. The acquired image data should include complete texture information and edge contour information of the valve body sealing surface to avoid blind spots.

[0069] For image acquisition in this step, one specific implementation method can employ multi-camera array synchronous triggering technology. Multiple industrial cameras are mounted on a pre-set annular support around the outlet of the drying vessel, with the optical axes of these cameras converging at the center of the valve body. Through trigger signal lines connected to the PLC control system, when a signal indicating the outlet valve has opened to the correct position is detected and the outlet flow meter reading returns to zero, all cameras are synchronously triggered to perform high-speed continuous shooting within a set delay period. Image data is transmitted to the industrial computer's memory via a gigabit Ethernet port or a Camera Link interface.

[0070] The second specific implementation method can use a single camera in conjunction with a high-speed robotic arm or a rotating gimbal. A high-resolution industrial camera is mounted at the end of a robotic arm capable of multi-degree-of-freedom motion, and a scanning trajectory around the discharge port is pre-planned. After the discharge action is completed, the robotic arm drives the camera to move rapidly along the predetermined trajectory, during which the camera takes pictures at a fixed high frame rate, thereby acquiring an image sequence covering different perspectives of the valve body surface.

[0071] S102. Identify the rigid texture features of the valve body surface within the reference frame in a multi-frame temporal image;

[0072] To identify the rigid texture features of the valve body surface within a reference frame in a multi-frame temporal image, the first approach is to: calculate the gray-level gradient magnitude of each pixel in the reference frame in the horizontal and vertical directions; mark regions with gray-level gradient magnitudes less than a preset gradient filtering threshold as flat regions; mark regions with gray-level gradient magnitudes not less than the preset gradient filtering threshold as high-frequency texture regions; and remove isolated noise regions in the high-frequency texture regions where the number of connected pixels is less than a preset noise area threshold to obtain the rigid texture features of the valve body surface.

[0073] The second implementation method can be as follows: divide the reference frame into several non-overlapping rectangular detection sub-windows; calculate the variance statistics of pixel grayscale values ​​in each rectangular detection sub-window; select rectangular detection sub-windows whose variance statistics are located within a preset texture saliency range as candidate feature blocks; calculate the spatial distance between each candidate feature block and remove redundant candidate feature blocks whose spatial distance is less than a preset overlap suppression threshold to obtain alternative feature blocks; determine the alternative feature blocks distributed at the valve body edge contour position as the rigid texture features of the valve body surface.

[0074] A reference frame is the frame selected as the baseline in a multi-frame time-series image series, typically the image with the highest clarity or located in the middle of the sequence. Rigid texture features on the valve body surface refer to the inherent texture structure of the valve body's metal surface that does not undergo non-linear deformation over time, such as metal machining marks, screw hole edges, casting defects, or specially engraved alignment marks. The gray-level gradient magnitude reflects the drastic change in image pixel gray-level. A preset gradient filtering threshold is used to distinguish between flat and textured regions; its value should be higher than the average gradient level of the image background noise. Flat regions are areas with gentle gray-level changes, while high-frequency textured regions are areas with drastic gray-level changes. A preset noise area threshold is used to filter out noise points with excessively small areas. A rectangular detection sub-window is an image patch used for local feature analysis; variance statistics are used to measure the dispersion of gray-level distribution within the window. A preset texture saliency interval defines the variance range of the effective texture, and a preset overlap suppression threshold is used to control the sparsity of the feature patch distribution. The core of this step is to extract stable and reliable feature points or feature blocks from the complex valve body surface image. These features must be insensitive to changes in lighting and be able to represent the rigid motion state of the valve body, eliminating interference caused by foreign objects blocking or local reflections.

[0075] For the first implementation method (gradient-based feature extraction), the Sobel operator or Prewitt operator can be used to calculate the gradient components of the reference frame image in the horizontal (x-direction) and vertical (y-direction) directions, respectively. The gradient magnitude of each pixel is obtained by calculating the square root of the sum of the squares of the gradient components in the two directions. Then, the gradient filtering threshold is determined using the Otsu method or the fixed threshold method to generate a binarized texture mask. For high-frequency texture regions in the mask, morphological opening operations (erosion followed by dilation) are applied to remove tiny burrs and isolated points. Then, a connected component labeling algorithm is used to count the number of pixels in each connected component. Regions with a number lower than a preset noise area threshold are set to zero. The finally retained regions are the rigid texture features.

[0076] For the second implementation method (variance-based block selection), a window of size N×N can be slid across the reference frame image with a step size of S, or the image can be directly segmented into a grid. The variance of the pixel grayscale values ​​within each window is calculated; a larger variance indicates richer texture. A lower and upper limit for the variance are set to form a saliency interval, and windows with excessively small variances (flat areas) or excessively large variances (potentially strong reflective points) are removed. For the retained candidate feature blocks, the Euclidean distance between their center coordinates is calculated. If the distance between two feature blocks is less than a preset overlap suppression threshold, the block whose variance is closer to the median of the interval is retained, or the block closer to the valve body edge model is retained. The final set of selected feature blocks constitutes the rigid texture features used for subsequent registration.

[0077] S103. Using the rigid texture features of the valve body surface in the reference frame as a benchmark, calculate the geometric transformation parameters of the other time-series images relative to the reference frame.

[0078] Geometric transformation parameters typically refer to mathematical parameters describing the spatial relationship between two images, including translation, rotation angle, scaling, and coefficients of possible affine or perspective transformation matrices. The reference frame refers to the reference frame and its features determined in step S102. The remaining multi-frame temporal images refer to all images in the acquisition sequence other than the reference frame. These images have positional deviations from the reference frame due to minor jitter or viewpoint differences during acquisition. The calculation process aims to find a mathematical transformation that allows the texture features in the image to be registered to coincide as closely as possible with the corresponding features in the reference frame in terms of coordinates after the transformation. This step is crucial for achieving multi-view image fusion and directly determines the effectiveness of subsequent denoising and difference operations. This process requires handling computational tasks such as feature matching, mismatch removal, and transformation model estimation, demanding high computational efficiency and robustness against interference.

[0079] The first specific approach to this step is based on feature point matching. First, using algorithms such as SIFT (Scale Invariant Feature Transform), SURF (Speeded Robust Feature Transform), or ORB (Oriented Fast Rotation Briefing), feature descriptors are extracted from the rigid texture feature regions of the reference frame, and simultaneously, the same feature descriptors are extracted from the global or local regions of the image to be registered. The Hamming or Euclidean distance between the two sets of descriptors is calculated for initial matching. Then, the RANSAC (Random Sample Consensus) algorithm is used iteratively to filter inlier points, eliminating incorrect matching pairs. Finally, using the filtered correct matching point pairs, the affine transformation matrix or homography matrix is ​​solved using the least squares method; this matrix represents the required geometric transformation parameters.

[0080] The second specific approach to this step is based on optical flow tracking. This assumes extremely short time intervals between adjacent frames in the image sequence and constant brightness. Several key points are selected at the rigid texture feature locations determined in the reference frame. Using the Lucas-Kanade optical flow algorithm or the Farneback dense optical flow algorithm, the motion vectors of these key points in the next frame are calculated. By tracing the trajectories of these points throughout the image sequence, the displacement vector field of each frame relative to the reference frame can be obtained. Statistical analysis of the displacement vector field is then performed to calculate global motion parameters (such as average translation and average rotation angle), or to fit global motion model parameters, which are then used as geometric transformation parameters.

[0081] S104. Spatial registration of multiple time-series images is performed based on geometric transformation parameters to generate a spatially aligned corrected image sequence.

[0082] Spatial registration refers to mapping the pixel coordinates of the image to be registered to the coordinate system of the reference frame using calculated geometric transformation parameters, ensuring that the coordinate position of the same physical point is consistent across different images. Correcting the image sequence refers to the set of images that, after geometric transformation, appear visually static. This step involves image resampling and interpolation. Since the coordinates after geometric transformation are usually not integers, interpolation algorithms are needed to calculate the pixel grayscale values ​​at the new coordinates. The accuracy of spatial alignment directly affects the effect of subsequent temporal filtering; inaccurate alignment can lead to image blurring or ghosting. This step also requires handling image boundary issues, as the effective area of ​​the transformed image may rotate or translate, causing blanking or out-of-bounds areas at the original rectangular boundaries.

[0083] The first specific method to implement this step is to use inverse mapping combined with bilinear interpolation. For each integer coordinate pixel (x, y) in the reference frame coordinate system, its corresponding floating-point coordinates (u, v) in the original image to be registered are calculated using the inverse of the geometric transformation matrix. Four neighboring integer pixels around (u, v) are found, and using the gray values ​​of these four points, bilinear interpolation is performed according to the distance weights between (u, v) and them to calculate the gray value of the target pixel. This process is repeated for each frame in the sequence, ultimately generating a corrected image sequence with the same resolution as the reference frame.

[0084] The second specific method to achieve this step is to use perspective transformation combined with bicubic interpolation. If step S103 calculates the homography matrix, then the perspective transformation formula is applied to correct the image. In the interpolation stage, 16 neighboring pixels (4×4 window) around the corresponding floating-point coordinates in the original image are selected, and a gray-level surface is fitted using cubic polynomial basis functions to calculate the gray value of the target point. Compared to bilinear interpolation, bicubic interpolation can better preserve the high-frequency details and edge sharpness of the image. Although it requires more computation, it can generate a higher-quality corrected image sequence, which is beneficial for the subsequent detection of small foreign objects.

[0085] S105. Perform time-domain statistical filtering on the grayscale value of the same pixel coordinate point in the time axis direction in the corrected image sequence to obtain the thermal real scene image;

[0086] The time axis direction refers to the temporal sequence of image acquisition. For each fixed two-dimensional spatial coordinate (x, y), its grayscale value is extracted from frame 1 to frame N, forming a one-dimensional time series. Temporal statistical filtering involves statistical analysis of this one-dimensional sequence to remove random noise and transient interference (such as flying dust, flickering light spots, and flowing hot air textures). The thermal real-world image is a single-frame image generated after filtering. It represents the most realistic and stable surface appearance of the valve body at high temperatures, removing dynamic interference from the acquisition process. This step utilizes the information redundancy characteristics of multiple frames, significantly improving the signal-to-noise ratio of the image.

[0087] The first specific way to implement this step is median filtering. For each pixel position (x, y) in the corrected image sequence, its grayscale values ​​in all frames are collected, forming an array G = [g_1, g_2, ..., g_n]. The elements in array G are sorted by value, and the value in the middle position after sorting is taken as the grayscale value of that pixel in the thermal real-world image. Median filtering is very effective in removing salt-and-pepper noise and fast-moving foreign objects (such as flying insects or rapidly drifting dust clumps) because these transient disturbances only appear for a very short time on the time axis and do not occupy the middle position of the sequence.

[0088] The second specific method to implement this step is the Gaussian weighted average method. Considering that the reference frame is usually located in the middle of the sequence or has the best image quality, different weights can be assigned to different frames based on temporal distance. A one-dimensional Gaussian kernel function is constructed, centered on the reference frame, with frames farther away having smaller weights. The time-series grayscale values ​​of each pixel are weighted and summed, and the weighted average is calculated as the pixel value of the thermal real-world image. This method suppresses Gaussian white noise while smoothing out minor fluctuations in illumination, resulting in a visually softer and more natural image that retains more background details.

[0089] S106. Obtain the real-time temperature data of the current outlet of the drying kettle and the pre-stored valve body cold state reference diagram;

[0090] Real-time temperature data refers to the temperature value of the valve body surface or outlet environment measured by a temperature sensor simultaneously with image acquisition. This data must be highly real-time and accurate, as it directly determines the parameters for subsequent thermal expansion compensation. The valve body cold-state reference image refers to a standard image of the valve body taken when the equipment is shut down and cooled to room temperature or standard temperature (e.g., 20°C), with no material residue or foreign matter adhering to it. This reference image should be consistent with the viewing angle during hot-state acquisition and stored in non-volatile memory. Pre-storage means that the image is data that has been acquired and verified as qualified before the start of the production cycle or during maintenance. This step bridges the gap between physical environmental parameters and image processing algorithms, ensuring that the detection algorithm can perceive environmental changes.

[0091] The first specific method to achieve this step is through acquiring temperature data and retrieving images from a database via an industrial fieldbus. Temperature sensors (such as PT100 RTDs or thermocouples) are installed at key locations on the valve body, and the signals are connected to a PLC or DCS system. The image processing unit communicates with the control system via Modbus TCP, Profinet, or OPC UA protocols to read the values ​​from the real-time temperature register. Simultaneously, the image processing unit connects to a local or cloud-based SQL database to query and download the corresponding cold-state reference based on the valve's device ID and model. Figure 2 Large Array Object (BLOB) data.

[0092] The second specific method to achieve this step is to utilize an infrared temperature measurement module and a local file system. A non-contact infrared temperature probe is integrated alongside the vision inspection system, directly measuring the temperature of the valve body surface. The data is transmitted to the processor via RS485 or an analog interface. The valve body cold-state baseline image is saved in a specific format (such as BMP or TIFF) in a designated directory on the industrial computer's solid-state drive, with the filename including the device number. The program directly reads the baseline image data according to the path in the configuration file through file I / O operations and performs hash verification to ensure the file has not been tampered with.

[0093] S107. Based on real-time temperature data and preset thermal expansion coefficient of valve body material, scale transformation and deformation compensation are performed on the cold state reference image of valve body to generate a thermal expansion compensation image that matches the size of the hot state real scene image.

[0094] The coefficient of thermal expansion is a physical parameter describing the dimensional changes of a material with temperature variations, typically referring to the coefficient of linear expansion. Scale transformation refers to overall magnification or reduction of the image based on the calculated expansion ratio. Deformation compensation refers to local adjustments made to address the non-uniform expansion of different parts of the valve body caused by variations in structural thickness and material. The thermal expansion compensation map is a cold-state baseline image after the above transformations; theoretically, it should perfectly match the geometric dimensions of the clean valve body at the current high temperature. The purpose of this step is to eliminate the influence of the "temperature" variable on image contrast, transforming the problem into simple "foreign object" detection. The preset coefficient of thermal expansion for the valve body material needs to be obtained by consulting tables or through actual measurement based on the specific material of the valve body.

[0095] The first specific method to achieve this step is based on an affine transformation of an isotropic uniform expansion model. Assuming the valve body material is uniform and heated uniformly, the expansion rate k = 1 + \alpha \times(T_{real} - T_{base}), where \alpha is the linear expansion coefficient, T_{real} is the real-time temperature, and T_{base} is the cold-state reference temperature. A scaling matrix is ​​constructed, with k as the main diagonal element. Using the image center or the geometric center of the valve body as anchor points, this scaling matrix is ​​applied to resample the cold-state reference image, generating an enlarged thermal expansion compensation map. This method is computationally simple and suitable for scenarios with symmetrical structures and small temperature differences.

[0096] The second specific method to achieve this step is a lookup table method based on finite element analysis (FEM) pre-calculation. For valve bodies with complex structures, thermal-structural coupling analysis is performed in advance using simulation software such as ANSYS to simulate the displacement field of the valve body surface at different temperature nodes. A lookup table is established to map the temperature to the image pixel displacement field. During detection, the current pixel displacement field is obtained based on real-time temperature interpolation. A remapping algorithm is then used to nonlinearly shift the position of each pixel in the cold-state reference image, thereby generating a high-precision thermal expansion compensation map. This method can accurately simulate non-uniform thermal deformation.

[0097] S108. Perform a difference operation between the thermal real-world image and the thermal expansion compensation image to obtain a grayscale difference image;

[0098] Difference operations involve subtracting the grayscale values ​​of corresponding pixels in two images to highlight the differences between them. A grayscale difference image is a grayscale image reflecting the degree of difference, where areas with pixel values ​​close to zero indicate similarity between the two images (i.e., background), while areas with larger pixel values ​​indicate significant differences (i.e., potential foreign objects). In this step, the thermal reality image represents the "current state," and the thermal expansion compensation image represents the "ideal clean state." Through difference operations, the textured background of the valve body itself is eliminated, retaining only the information of newly added foreign objects. Difference operations can be simple algebraic subtraction or more complex difference measurement algorithms.

[0099] The first specific method to implement this step is the absolute difference method. For each pixel (x, y) of the thermal real-world image I_{hot} and the thermal expansion compensation image I_{comp}, calculate the absolute difference in grayscale values ​​D(x, y) = |I_{hot}(x, y) - I_{comp}(x, y)|. To enhance contrast, the result can be linearly stretched, mapping the difference range to 0-255. This method is extremely fast and intuitively reflects the brightness difference.

[0100] The second specific method to achieve this step is the Structural Similarity-Induced Simulation (SSIM) difference map method. Instead of direct subtraction, it calculates the brightness, contrast, and structural similarity indices of the two images within a local window. This generates an SSIM map where values ​​closer to 1 indicate greater similarity, and values ​​closer to 0 indicate greater dissimilarity. The SSIM map is then inverted (i.e., 1-SSIM) and normalized to obtain a grayscale difference map. This method has better tolerance for variations in lighting and small alignment errors, and can more accurately reflect structural differences (such as the shape of foreign objects) rather than just brightness differences.

[0101] S109. Perform connectivity analysis on the grayscale difference map and calculate the connectivity area of ​​non-zero regions;

[0102] Connected component analysis (CBI) is an image processing technique used to group adjacent pixels with the same pixel value in a binary or grayscale image into a single, connected region. A non-zero region refers to the set of pixels representing differences retained after thresholding in a grayscale difference image. The connected area refers to the total number of pixels contained within each connected region, corresponding to the projected area of ​​an object in the physical world on the image. The purpose of this step is to transform pixel-level differences into object-level descriptions, aggregating discrete difference points into physically meaningful "patches" and quantifying their size. Before performing CBI, the grayscale difference image is typically binarized, with a brightness threshold set. Points above this threshold are considered "foreground" (potential foreign objects), and points below the threshold are considered "background."

[0103] The first specific way to implement this step is the two-pass algorithm. First, the binarized difference graph is scanned, visiting pixels row by row. Temporary labels are assigned based on the labels of the pixels' left and top neighbors, and the equivalence relationships between labels (i.e., different labels actually belong to the same connected component) are recorded. After the first scan, the equivalence table is parsed, and equivalent labels are merged. A second scan is performed, replacing the temporary label of each pixel with the final normalized label. Finally, all labels are traversed, and the number of pixels corresponding to each label is counted, which is the connected area.

[0104] The second specific approach to this step is based on contour-tracking algorithms (such as the Suzuki algorithm). This method does not directly scan all pixels, but instead searches for the boundary contours of the foreground region in the image. Tracking is performed along the boundaries to determine the hierarchical relationship of the contours (outer contours and holes). The area is calculated using Green's theorem or by directly counting the number of pixels inside the contour. This method not only yields the area but also simultaneously obtains the shape features of the foreign object (such as perimeter, centroid, and bounding box), facilitating subsequent advanced analysis.

[0105] S110. If the connected area exceeds the preset judgment threshold, it is determined that there is foreign object residue.

[0106] The preset judgment threshold is a numerical standard set according to actual production process requirements and error tolerance. The unit is usually the number of pixels or a converted physical area (e.g., square millimeters). If the connected area exceeds this threshold, it means the detected difference area is large enough to be unlikely to be caused by image noise, lighting flicker, or fine dust, but rather by substantial material residue or foreign matter. Determining the presence of foreign matter residue is the final logical judgment result, which will trigger corresponding control actions, such as alarms, production line shutdown, or triggering automatic cleaning programs. This step is the decision-making link in the entire detection process, requiring a balance between the recall rate and the precision rate.

[0107] A fixed area threshold T_{area} can be set. The areas A_i of all connected regions obtained in S109 are iterated over. If any A_i > T_{area}, the system outputs a high-level signal indicating the presence of a foreign object; if all A_i are less than or equal to T_{area}, a low-level signal indicating the absence of a foreign object is output. This method is logically simple, has a fast response time, and is suitable for scenarios where the size of the foreign object has a clear limit.

[0108] In the above embodiments, by continuously acquiring multi-view, multi-frame time-series images under high-temperature conditions after the drying kettle discharge valve is opened, and performing spatial registration based on the rigid texture features of the valve body surface, interference caused by viewpoint differences and jitter during image acquisition can be effectively reduced. Temporal statistical filtering is used to generate a thermal real-world image, and thermal expansion compensation is performed on the cold-state reference image using real-time temperature data, which reduces the impact of valve body deformation caused by temperature changes on the detection results. Differential operations and connected component analysis between the thermal real-world image and the thermal expansion compensation image improve the detection accuracy of foreign matter residues on the valve body surface and reduce the probability of false detections caused by temperature changes, viewpoint differences, and other factors. This detection method based on multi-view image fusion and thermal compensation improves the reliability and adaptability of foreign matter residue detection in the automatic discharge valve of the drying kettle, thereby improving the overall efficiency of production batch turnover.

[0109] The above embodiments achieve preliminary detection of foreign matter residues on the valve body surface through multi-view image fusion and temperature compensation. However, in practical applications, due to the non-uniform deformation of the valve body material at high temperatures, artifacts similar to foreign matter residue characteristics may appear in some areas. To further improve the accuracy of detection, it is necessary to conduct in-depth analysis and screening of the suspected foreign matter areas obtained from the preliminary detection. The following section combines... Figure 2 Another method for detecting the operating condition of an automatic discharge valve in a drying kettle using multi-view image fusion, as described in this application, is as follows:

[0110] Please see Figure 2 This is another flowchart illustrating a multi-view image fusion method for detecting the operating conditions of an automatic discharge valve in a drying kettle, as described in this application.

[0111] S201. Based on the grayscale difference map, extract the connected regions with grayscale values ​​greater than the preset difference threshold as suspected foreign object regions;

[0112] A grayscale difference image is an image generated by subtracting the grayscale values ​​of corresponding pixels and taking the absolute value after pixel alignment of a thermal reality image and a thermal expansion compensation image. This image can intuitively reflect the difference in brightness or thermal radiation intensity between the two images. A connected region refers to a set of pixels with the same or similar pixel values ​​and adjacent positions in a binary or grayscale image. These pixels are spatially connected, forming an independent block structure. A preset difference threshold is a pre-defined grayscale judgment standard. This threshold can be a fixed empirical value or an adaptive value based on the statistical characteristics of the image histogram. Its function is to filter out insignificant differences caused by sensor noise or minor environmental changes. A suspected foreign object region refers to a region in the grayscale difference image where the grayscale intensity exceeds the preset difference threshold and forms a spatially connected shape. These regions are initially identified as locations where foreign object residues may exist, but artifacts caused by valve body deformation have not yet been ruled out. In this step, setting the preset difference threshold is crucial, as it directly determines the detection sensitivity. If the threshold is set too high, it may lead to missed detection of small foreign objects; if the threshold is set too low, it will introduce excessive noise interference. The extraction process based on the grayscale difference image is actually a rough screening of abnormal thermal radiation areas on the valve body surface, aiming to quickly locate all potential anomalies and provide regions of interest (ROIs) for subsequent refined texture analysis and structural trend analysis.

[0113] The specific implementation of this step mainly involves image thresholding and morphological processing techniques. The first implementation method combines a global fixed thresholding method with a connected component labeling algorithm. First, the system traverses each pixel in the grayscale difference image, comparing its grayscale value with a preset fixed difference threshold. If the pixel's grayscale value is greater than the threshold, it is marked as a foreground pixel (e.g., set to 255); otherwise, it is marked as a background pixel (e.g., set to 0), thus generating a binarized mask image. Next, a two-pass scanning method or a seed-filling algorithm is used to perform connectivity analysis on the foreground pixels in the binarized image (usually using 8-neighborhood connectivity). Interconnected pixels are grouped into independent connected components and assigned unique identifiers. Finally, all marked connected regions are extracted as suspected foreign object regions.

[0114] The second approach combines adaptive thresholding with morphological filtering. Considering the potentially uneven temperature distribution on the valve body surface, a local adaptive thresholding algorithm (such as the Sauvola algorithm) is used to calculate the statistical features within the neighborhood of each pixel as the threshold for that point, generating a binary image. Subsequently, a morphological opening operation (erosion followed by dilation) is performed on the binary image to remove isolated noise and small burrs. Then, a morphological closing operation (dilation followed by erosion) is performed to fill in holes and broken edges within connected regions. Finally, a contour-finding algorithm is used to extract the morphologically optimized connected regions as suspected foreign object regions.

[0115] S202. Using the edge detection operator, calculate the first gradient direction angle of each pixel in the suspected foreign object area of ​​the thermal real scene image, and the second gradient direction angle of the corresponding pixel in the suspected foreign object area of ​​the thermal expansion compensation image.

[0116] Edge detection operators are mathematical tools used to identify points with significant brightness changes in digital images. They are commonly used to extract structural features from images, such as the Sobel operator, Prewitt operator, or Scharr operator. The thermal real-world image refers to a pre-processed image of the valve body surface acquired in real-time by an imaging device while the automatic discharge valve of the drying vessel is operating at high temperatures. The thermal expansion compensation image is a reference image generated by an algorithm based on the physical model and temperature field distribution of the valve body, simulating an ideal state free of foreign matter residue and compensated for thermal deformation. The first gradient direction angle refers to the angle between the direction of the fastest grayscale change of a pixel within the suspected foreign matter area in the thermal real-world image and the horizontal axis; similarly, the second gradient direction angle refers to the grayscale change direction angle of the corresponding pixel in the thermal expansion compensation image. The gradient direction angle reflects the direction of local texture and edge normals in the image, exhibiting rotation invariance and illumination insensitivity. In this step, the purpose of calculating the gradient direction angle is to capture the structural features of the valve body surface at the microscopic texture level. When a valve body undergoes nonlinear deformation at high temperatures, the texture direction (i.e., gradient direction) on its surface often maintains a certain continuity. However, the presence of foreign objects disrupts this continuity, leading to significant differences in the gradient direction at the same location between the real-world image and the compensated image. Therefore, accurately calculating and comparing the gradient direction angles of the two images within the suspected region is a crucial prerequisite for distinguishing deformation artifacts from real foreign objects.

[0117] Implementing this step typically involves image convolution operations and trigonometric function transformations. One specific implementation uses the Sobel operator for convolution calculation. First, Sobel convolution kernels are constructed in the horizontal and vertical directions respectively. These two kernels are then convolved with the pixel matrix within the suspected foreign object region of the thermal real-world image and the thermal expansion compensation image, respectively, to obtain the horizontal and vertical gradient components of each pixel. Next, the arctangent function (arctan²) is used to convert the ratio of the vertical to horizontal gradient components of each pixel into an angle value, which is the gradient direction angle. The angle obtained from the thermal real-world image is denoted as the first gradient direction angle, and the angle obtained from the thermal expansion compensation image is denoted as the second gradient direction angle.

[0118] The second implementation utilizes the Scharr operator, which is more sensitive to high-frequency noise and provides more accurate gradient direction estimation. The process is similar to the Sobel operator, but the central neighboring pixels are given higher weights in the convolution kernel to enhance the ability to capture edge details. After calculating the horizontal and vertical gradients, the gradient direction angle in radians is calculated using inverse trigonometric functions and mapped to a range of 0 to 360 degrees for subsequent angle difference calculations.

[0119] S203. Calculate the cosine average of the angle difference between the first gradient direction angle and the second gradient direction angle of all pixels in the suspected foreign object region;

[0120] The calculation of the cosine average of the angle difference between the first gradient direction angle and the second gradient direction angle of all pixels within the suspected foreign object region includes: using an edge detection operator to calculate the gradient magnitude of each pixel in the thermal real-world image within the suspected foreign object region; calculating the cosine value of the difference between the first gradient direction angle and the second gradient direction angle of each pixel; summing the products of the gradient magnitude and cosine value of all pixels within the suspected foreign object region to obtain the weighted similarity sum; calculating the sum of the gradient magnitudes of all pixels within the suspected foreign object region to obtain the weighted sum; and calculating the ratio of the weighted similarity sum to the weighted sum, and determining the ratio as the cosine average.

[0121] The cosine mean is a statistical measure of the directional consistency between two vector fields. Its value typically ranges from -1 to 1, with values ​​closer to 1 indicating greater directional consistency. Gradient magnitude refers to the rate of grayscale change of an image pixel along the gradient direction, reflecting edge strength or texture saliency. The weighted similarity sum is the result of multiplying the directional similarity (cosine value) of each pixel by its edge strength (gradient magnitude) and then summing the results. Physically, this gives greater weight to regions with clear edges and significant textures in the similarity assessment, while reducing the weight of flat or noisy regions. The weighted sum is the sum of the gradient magnitudes of all participating pixels, used to normalize the weighted similarity sum. The core idea of ​​this step is that simply comparing gradient direction angles may be affected by noise in flat regions (small gradient magnitudes, unstable direction angles). Therefore, gradient magnitude is introduced as a weight to construct a weighted texture direction similarity metric. By calculating the cosine value of the angle difference, the angle difference can be converted into numerical similarity (0-degree difference corresponds to cosine value 1, 90-degree difference corresponds to cosine value 0, and 180-degree difference corresponds to cosine value -1). Combined with amplitude weighting, the texture matching degree between the thermal real scene map and the thermal expansion compensation map in the suspected area can be accurately quantified.

[0122] According to the refined technical solution, the specific process for implementing this step is as follows: First, using an edge detection operator (such as the Sobel operator), calculate the horizontal and vertical gradient components of each pixel in the suspected foreign object region of the thermal real-world image, and calculate the gradient magnitude of each pixel using the Pythagorean theorem (the square root of the sum of squares). Second, for each pixel in the region, obtain the first gradient direction angle and the second gradient direction angle calculated in step S202, calculate the absolute value of the difference between these two angles, and obtain the cosine function value of this difference. This cosine value represents the single-point directional similarity. Third, multiply the gradient magnitude of each pixel by the corresponding cosine value to obtain the weighted similarity component of that point, and sum the weighted similarity components of all pixels in the region to obtain the weighted similarity sum; at the same time, sum the gradient magnitudes of all pixels in the region to obtain the weighted sum. Finally, divide the weighted similarity sum by the weighted sum, and the quotient is the cosine average of the angle differences between the first and second gradient direction angles of all pixels in the suspected foreign object region.

[0123] The second implementation method can adopt the idea of ​​vector dot product, treating the gradient of each pixel as a vector, directly calculating the dot product of the gradient vector of the real image and the gradient vector of the compensation image, then dividing by the product of the magnitudes of the two vectors to obtain the cosine value, and then multiplying by the gradient magnitude of the real image as the weight for weighted averaging. This method is mathematically equivalent, but matrix operations can be used to speed up the calculation process.

[0124] S204. Use the cosine average value as the texture structure similarity of the suspected foreign object region;

[0125] Texture structure similarity is a quantitative indicator used to characterize the degree of agreement between the texture features of a thermally rendered image and a thermal expansion compensation image within a suspected foreign object region. In this application, this similarity is directly defined by the cosine average value calculated in step S203. A higher cosine average value indicates that the gradient directions of the two images are more consistent in this region, and the texture orientations are closer; conversely, a lower cosine average value indicates a significant difference in the texture structures of the two images. This step is a crucial step in transforming complex image feature data into a single numerical indicator, providing a direct quantitative basis for subsequent logical judgments. Defining the cosine average value directly as texture structure similarity utilizes the excellent properties of the cosine function in measuring directional consistency, while combining it with gradient magnitude weighting, so that this indicator includes both directional and intensity information, and can robustly reflect whether the deformation characteristics of the valve body surface at high temperatures are consistent with the model prediction.

[0126] The specific implementation of this step mainly involves numerical mapping and data storage. The first method is direct assignment, where a floating-point variable is allocated in memory, and the cosine average calculated in step S203 is directly assigned to a variable named "texture structure similarity". This variable's value range is between -1 and 1, and subsequent threshold judgments will be based on this range. The second method is normalized mapping. For easier intuitive understanding and parameter configuration, the calculated cosine average can be linearly mapped to the interval between 0 and 100. Specifically, the cosine average is added by 1, divided by 2, and then multiplied by 100 to obtain a percentage-based texture structure similarity score. This mapping method maps cases with completely opposite directions to 0 and cases with completely consistent directions to 100, making the physical meaning of similarity more consistent with human cognitive habits and facilitating the setting and adjustment of subsequent consistency thresholds by operators.

[0127] S205. If the texture structure similarity is greater than the preset consistency threshold, the suspected foreign object area is determined to be an artifact area caused by the nonlinear deformation of the valve body.

[0128] The preset consistency threshold is a critical value used to distinguish between normal deformation textures and foreign object textures. This threshold is set based on statistical analysis of a large amount of valve body high-temperature deformation data, and is usually set at a relatively high level (e.g., 0.8 or 0.85), meaning that only when the texture direction of the real-world image and the compensation image is highly consistent will it be judged as a non-foreign object. Artifact regions caused by nonlinear deformation of the valve body refer to the brightness differences that occur during image differencing due to uneven expansion or distortion of the valve body material under high temperature and pressure. However, the micro-texture of its surface (such as metal processing marks, surface roughness direction) does not undergo abrupt changes and still maintains a high degree of similarity to the baseline model (thermal expansion compensation image). The logic behind this step is that if it is a real foreign object (such as material residue), its texture features are usually completely different from the texture of the valve body surface, leading to chaotic gradient directions and thus low texture structure similarity; while if it is the deformation of the valve body itself, although the position has shifted, its inherent surface texture structure is continuous, so the texture structure similarity will remain at a high level. This determination step can effectively identify and exclude false positive test results that are merely due to deformation.

[0129] The specific implementation of this step mainly involves conditional logic judgment. The first method uses a simple single-threshold comparison logic. The system reads the texture structure similarity value determined in step S204 and compares it with a preset consistency threshold stored in the configuration file. If the similarity value is strictly greater than the threshold, the judgment logic is triggered, marking the currently processed suspected foreign object area as an "artifact"; otherwise, it maintains its "suspected foreign object" state and proceeds to subsequent processing. The second method uses fuzzy judgment logic with confidence intervals. Two thresholds are set: a high threshold and a low threshold. If the similarity is greater than the high threshold, it is directly judged as an artifact; if the similarity is between the high and low thresholds, a secondary confirmation is performed based on the average brightness or edge intensity of the area. If the edge intensity is weak and the similarity is acceptable, it is also inclined to be judged as an artifact. This method is more flexible than hard threshold judgment and can handle fuzzy cases in a critical state.

[0130] S206. Remove artifact areas from foreign object residue;

[0131] Artifact areas, identified in step S205 as image anomalies caused by nonlinear deformation of the valve body, are not actually foreign objects, despite showing differences in the grayscale difference image. Foreign object residue refers to material residue or impurities confirmed as genuinely present in the final output to the user. The removal operation involves canceling the alarm marker for this area at the data or image mask level, classifying it as background or normal area to prevent false alarms. This step is a crucial "cleaning" step in the entire detection process, directly affecting the system's false alarm rate. By removing artifacts, the system can focus on areas where the texture structure truly undergoes abrupt changes, thus providing operators with accurate cleanliness feedback.

[0132] The specific implementation of this step mainly involves data structure operations and image mask updates. The first method is list-based elimination. Internally, a "suspected target list" is maintained, storing the coordinates, area, similarity, and other attributes of all candidate regions. When a region is identified as an artifact, the corresponding node or object is directly deleted from the list, or its attribute flag field is updated to "Ignored". In the final output of the detection results, only targets marked "Valid" in the list are traversed. The second method is image manipulation based on a binary mask. A binary result image of the same size as the original image is maintained. Initially, all suspected regions have a pixel value of 1 (white), and the background is 0 (black). When a region is identified as an artifact, a drawing function is used to draw a filled polygon with a pixel value of 0 at the corresponding position in the binary result image, thus "erasing" the region at the image level. Subsequent steps only process the regions that remain white in the binary result image.

[0133] S207. When the texture structure similarity is determined to be no greater than the preset consistency threshold, low-pass filtering is performed on the image data of the suspected foreign object region and the image data of the thermal expansion compensation map at the corresponding positions to obtain the first low-frequency feature map and the second low-frequency feature map respectively.

[0134] When the texture structure similarity is not greater than the preset consistency threshold, it means that the micro-texture of the area is inconsistent with the model. However, this does not necessarily mean it is a foreign object; it could also be that the three-dimensional structure of the valve body has undergone significant low-frequency deformation after thermal expansion (such as overall bulging or depression). This deformation may not match in texture details, but it is still correlated in the overall trend. Low-pass filtering is a signal processing technique that aims to remove high-frequency components (such as sharp edges, noise, and fine textures) from the image while retaining low-frequency components (such as overall brightness distribution, slowly changing contours, and large structural trends). The first and second low-frequency feature maps are the results of low-pass filtering on the thermal real-world image and the thermal expansion compensation image, respectively. These two images no longer focus on surface roughness or processing texture, but rather on the trend of light and shadow changes and the macroscopic three-dimensional structure within the area. The purpose of this step is to introduce macroscopic structural trend analysis after micro-texture analysis to further distinguish between real three-dimensional foreign objects (such as blocky materials) and the three-dimensional thermal deformation of the valve body itself.

[0135] The specific implementation of this step mainly involves convolutional filtering algorithms. The first method is to use Gaussian low-pass filtering. A relatively large Gaussian convolution kernel (e.g., 5x5 or 7x7) with a suitable standard deviation is constructed, and convolution operations are performed on the ROI (Region of Interest) in the suspected foreign object area in both the thermal reality image and the thermal expansion compensation image. The weights of the Gaussian filter have a bell-shaped distribution, which can smoothly remove details and generate a soft low-frequency feature map. The second method is to use mean filtering or box filtering. A convolution kernel with all 1s is constructed, and the average value of pixels within the image area is calculated. This method is computationally fast, significantly effective in removing high-frequency textures, and the resulting low-frequency feature map can well reflect the average brightness distribution trend within the area. When processing boundaries, edge replication or reflection can be used to avoid edge effects after filtering.

[0136] S208. Calculate the structural trend similarity between the first low-frequency feature map and the second low-frequency feature map;

[0137] Calculating the structural trend similarity between the first low-frequency feature map and the second low-frequency feature map specifically includes: obtaining the gray values ​​of all pixels in the first low-frequency feature map to construct a first gray-level matrix, and obtaining the gray values ​​of all pixels in the second low-frequency feature map to construct a second gray-level matrix; calculating the correlation coefficient between the first gray-level matrix and the second gray-level matrix using a normalized cross-correlation algorithm, and determining the correlation coefficient as the structural trend similarity.

[0138] Structural trend similarity is an indicator used to measure the consistency of the variation trends of two images in low-frequency components. It focuses not on the specific pixel values, but on the variation patterns and relative distribution of pixel values. For example, if two images both exhibit a "bright center, dark edges" distribution in the same area, their structural trends are considered highly similar. The first and second grayscale matrices are numerical representations of the first and second low-frequency feature maps, respectively. Normalized Cross-Correlation (NCC) is a classic image matching algorithm that measures the degree of linear correlation between two sequences by calculating the correlation coefficient. NCC has brightness invariance; that is, if an image becomes brighter or darker overall, its NCC value remains unchanged, which is particularly important for processing valve images affected by lighting. The core of this step is to determine whether the suspected area is merely a "bulge" or "depression" on the valve surface (i.e., three-dimensional structural thermal deformation) or a completely foreign object by comparing the statistical correlation of the low-frequency feature maps. If it is a three-dimensional deformation of the valve itself, its light and shadow trends should remain highly consistent in the real-world image and the compensated image.

[0139] According to the refined technical solution, the specific process for implementing this step is as follows: First, expand the grayscale values ​​of all pixels in the suspected foreign object region of the first low-frequency feature map by row or column to construct a one-dimensional vector or a two-dimensional first grayscale matrix; similarly, obtain the pixel values ​​of the second low-frequency feature map to construct a second grayscale matrix. Second, calculate the mean and standard deviation of all elements in the first grayscale matrix, and the mean and standard deviation of all elements in the second grayscale matrix. Then, calculate using the normalized cross-correlation formula: subtract the mean from each element of the first grayscale matrix, and subtract the mean from each element of the second grayscale matrix, and calculate the sum of the products of the corresponding elements in these two mean-free matrices (i.e., the numerator of the covariance). Finally, divide the sum of this product by (total number of elements in the first grayscale matrix × standard deviation of the first matrix × standard deviation of the second matrix), and the result is the correlation coefficient, which is determined as the structural trend similarity. The second implementation method can use the Pearson Correlation Coefficient, which is mathematically similar to the NCC (Neutral Coefficient of Covariance) and is achieved by calculating the ratio of the product of the covariance and standard deviation of the two variable sequences. In programming, existing mathematical library functions can be used to directly perform correlation analysis on the two matrices and output a scalar value between -1 and 1.

[0140] S209. If the structural trend similarity is greater than the preset trend tolerance threshold, the suspected foreign object area is determined to be an artifact caused by thermal deformation of the valve body's three-dimensional structure, and the foreign object attribute of the suspected foreign object area is excluded.

[0141] The preset trend tolerance threshold is a boundary used to distinguish between three-dimensional structural deformation and real foreign objects. This threshold is usually set relatively high (e.g., 0.75 or 0.8), meaning that only when the low-frequency light and shadow trend in the real-world image is very close to the trend predicted by the model is it considered a three-dimensional deformation of the valve body itself. Three-dimensional structural thermal deformation of the valve body refers to the macroscopic geometric changes that occur in the valve body at high temperatures, such as warping of the sealing surface and bulging around bolt holes. These deformations will form specific light and shadow distributions in the image. Excluding foreign object attributes means confirming that although the area differs in grayscale difference and texture details, it is part of the valve body in terms of macroscopic structure, and therefore its classification as a "foreign object" is revoked. This step is the last line of defense in the multi-view image fusion detection method. It specifically targets non-foreign object areas whose texture may have changed due to wear or carbon buildup (resulting in S205 not being removed), but whose overall shape remains unchanged, further reducing the false alarm rate and ensuring that the final output foreign object alarm has extremely high reliability.

[0142] The specific implementation of this step mainly involves comparison logic and state updates. The first method is conditional branching. The structural trend similarity calculated in step S208 is numerically compared with a preset trend tolerance threshold. If the similarity is greater than the threshold, the system internally issues an instruction to change the classification label of the area from "Suspected" to "Deformation_Artifact" and remove it from the final alarm output queue. If the similarity is less than or equal to the threshold, the area is confirmed as a real foreign object residue, its foreign object attribute is retained, and an alarm is triggered. The second method is a comprehensive scoring judgment combined with weights. In addition to directly comparing trend similarity, the spatial location weight of the area can also be combined (for example, the edge area of ​​the valve body is more prone to three-dimensional deformation, so the weight can be set low; the central area has a high weight). A weighted trend score is calculated; if the score exceeds the threshold, it is determined to be three-dimensional structural thermal deformation and excluded.

[0143] In the above embodiments, by analyzing the gradient direction angles of pixels within the suspected foreign object region and calculating the texture structure similarity between the thermal real-world image and the thermal expansion compensation image in that region, artifact regions caused by nonlinear deformation of the valve body and actual foreign object residues can be distinguished. This analysis method based on local texture structure similarity reduces the interference of valve body nonlinear deformation on the detection results and improves the accuracy of foreign object detection. By introducing comparative analysis of gradient direction angles, this method enhances the detection algorithm's ability to distinguish between valve body deformation and foreign object residues, reducing false alarms caused by valve body thermal deformation. This artifact region identification and removal mechanism improves the reliability of the detection results.

[0144] The system in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 3This is a schematic diagram of the physical device structure of an automatic discharge valve detection system for a drying kettle with multi-view image fusion provided in an embodiment of this application.

[0145] It should be noted that, Figure 3 The structure of the system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0146] like Figure 3 As shown, the system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on a program stored in Read-Only Memory (ROM) 302 or a program loaded from storage portion 308 into Random Access Memory (RAM) 303, such as executing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0147] The following components are connected to I / O interface 305: input section 306 including a camera, infrared sensor, etc.; output section 307 including a liquid crystal display (LCD) and speakers, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card and a modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0148] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0149] It should be noted that the computer-readable medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein a computer-readable computer program is carried. The transmitted data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0150] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0151] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the system described in the above embodiments; or it may exist independently and not assembled into the system. The storage medium carries one or more computer programs that, when executed by a processor of a system, cause the system to implement the methods provided in the above embodiments.

[0152] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0153] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0154] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0155] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for detecting the operating condition of an automatic discharge valve in a drying kettle using multi-view image fusion, characterized in that, The method includes: Under high temperature conditions after the discharge valve of the drying kettle is opened and the discharge action is completed, multi-frame time-series images of the valve body surface at multiple angles are continuously acquired. Identify the rigid texture features of the valve body surface within a reference frame in the multi-frame temporal image; Using the rigid texture features of the valve body surface in the reference frame as a reference, calculate the geometric transformation parameters of the other time-series images relative to the reference frame; Spatial registration is performed on the multi-frame temporal images according to the geometric transformation parameters to generate a spatially aligned corrected image sequence; Temporal statistical filtering is performed on the grayscale value of the same pixel coordinate point in the time axis direction in the corrected image sequence to obtain a thermal real-world image; Acquire real-time temperature data at the outlet of the drying kettle and pre-stored cold-state reference diagram of the valve body; Based on the real-time temperature data and the preset thermal expansion coefficient of the valve body material, the cold state reference image of the valve body is scaled and deformed to generate a thermal expansion compensation image that matches the size of the hot state real scene image. The difference operation is performed between the thermal real-world image and the thermal expansion compensation image to obtain a grayscale difference image; Perform connected component analysis on the grayscale difference graph and count the connected areas of non-zero regions; If the connected area exceeds a preset threshold, it is determined that foreign matter remains.

2. The method for detecting the operating condition of an automatic discharge valve in a drying autoclave using multi-view image fusion as described in claim 1, characterized in that, The identification of the rigid texture features of the valve body surface within the reference frame in the multi-frame temporal images specifically includes: Calculate the grayscale gradient magnitude of each pixel in the reference frame in the horizontal and vertical directions; Regions with grayscale gradient magnitudes less than a preset gradient filtering threshold are marked as flat regions; regions with grayscale gradient magnitudes not less than the preset gradient filtering threshold are marked as high-frequency texture regions. Isolated noise regions with fewer than a preset noise area threshold connected pixels in the high-frequency texture region are removed to obtain the rigid texture features of the valve body surface.

3. The method for detecting the operating condition of an automatic discharge valve in a drying autoclave using multi-view image fusion as described in claim 1, characterized in that, The identification of the rigid texture features of the valve body surface within the reference frame in the multi-frame temporal images specifically includes: Several non-overlapping rectangular detection sub-windows are divided on the reference frame; Calculate the variance statistic of the pixel grayscale values ​​within each of the rectangular detection sub-windows; Rectangular detection sub-windows whose variance statistics fall within a preset texture saliency range are selected as candidate feature blocks; Calculate the spatial distance between each candidate feature block, and remove redundant candidate feature blocks whose spatial distance is less than a preset overlap suppression threshold to obtain alternative feature blocks; The candidate feature blocks distributed at the edge contour of the valve body are determined as rigid texture features on the valve body surface.

4. The method for detecting the operating condition of an automatic discharge valve in a drying autoclave using multi-view image fusion as described in claim 1, characterized in that, After determining the presence of foreign matter residue, the method further includes: Based on the grayscale difference map, connected regions with grayscale values ​​greater than a preset difference threshold are extracted as suspected foreign object regions; Using an edge detection operator, the first gradient direction angle of each pixel in the suspected foreign object region of the thermal real-world image and the second gradient direction angle of the corresponding pixel in the suspected foreign object region of the thermal expansion compensation image are calculated respectively. Calculate the cosine average of the angle differences between the first gradient direction angle and the second gradient direction angle of all pixels in the suspected foreign object region; The cosine average value is used as the texture structure similarity of the suspected foreign object region; If the texture structure similarity is greater than a preset consistency threshold, the suspected foreign object region is determined to be an artifact region caused by the nonlinear deformation of the valve body. Remove the artifact region from the foreign matter residue.

5. The method for detecting the operating condition of an automatic discharge valve in a drying autoclave using multi-view image fusion as described in claim 4, characterized in that, The calculation of the cosine average of the angle difference between the first gradient direction angle and the second gradient direction angle of all pixels in the suspected foreign object region specifically includes: Using the edge detection operator, the gradient magnitude of each pixel in the suspected foreign object region of the thermal real-world image is calculated; Calculate the cosine of the difference between the first gradient direction angle and the second gradient direction angle for each pixel; The weighted similarity sum is obtained by summing the product of the gradient magnitude and the cosine value of all pixels in the suspected foreign object region. Calculate the sum of the gradient magnitudes of all pixels within the suspected foreign object region to obtain the total weight; Calculate the ratio of the weighted similarity sum to the weight sum, and determine the ratio as the cosine average.

6. The method for detecting the operating condition of an automatic discharge valve in a drying autoclave using multi-view image fusion as described in claim 4, characterized in that, After removing the artifact region from the foreign matter residue, the method further includes: If the texture structure similarity is determined to be no greater than the preset consistency threshold, low-pass filtering is performed on the image data of the suspected foreign object region and the image data of the thermal expansion compensation map at the corresponding positions to obtain the first low-frequency feature map and the second low-frequency feature map respectively. Calculate the structural trend similarity between the first low-frequency feature map and the second low-frequency feature map; If the structural trend similarity is determined to be greater than the preset trend tolerance threshold, the suspected foreign object region is determined to be an artifact caused by thermal deformation of the valve body's three-dimensional structure, and the foreign object attribute of the suspected foreign object region is excluded.

7. The method for detecting the operating condition of an automatic discharge valve in a drying autoclave using multi-view image fusion as described in claim 6, characterized in that, The calculation of the structural trend similarity between the first low-frequency feature map and the second low-frequency feature map specifically includes: Obtain the gray values ​​of all pixels in the first low-frequency feature map to construct a first gray-level matrix, and obtain the gray values ​​of all pixels in the second low-frequency feature map to construct a second gray-level matrix; The correlation coefficient between the first gray matrix and the second gray matrix is ​​calculated using a normalized cross-correlation algorithm, and the correlation coefficient is determined as the structural trend similarity.

8. A multi-view image fusion-based automatic discharge valve condition detection system for a drying kettle, characterized in that, The system includes: One or more processors and a memory; the memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors calling the computer instructions to cause the system to perform a method for detecting the operating condition of an automatic discharge valve of a drying kettle using multi-view image fusion as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the processor, the processor performs a method for detecting the operating condition of an automatic discharge valve in a drying kettle using multi-view image fusion as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the processor, the processor executes a method for detecting the working condition of an automatic discharge valve in a drying kettle using multi-view image fusion as described in any one of claims 1-7.

Images