Image data management method and system for high-precision size visual inspection

Abstract

The invention discloses an image data management method and system for high-precision size visual inspection, and relates to the technical field of image data, the image data management method for high-precision size visual inspection comprises the following steps: S1, collecting a data set, and carrying out region segmentation processing; s2, correcting and enhancing the data set, and extracting contour feature parameters; s3, mapping association is carried out, and screening classification is carried out; s4, parameters are optimized and adjusted, and a standardized size detection feature set is generated; s5, constructing a size detection model to obtain a target size quantification detection result; and S6, performing accuracy verification, and establishing an image feature size parameter association database. According to the invention, through synchronous acquisition of a multi-view image data set and a physical size calibration data set, and in combination with calibration reference of a geometric quantity standard appliance, the problems of incomplete information and large physical size mapping deviation of a traditional single-view image are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data, in particular, is especially related to an image data management method and system for high-precision size visual detection. BACKGROUND

[0002] As a core means of realizing non-contact precision measurement in the field of industrial manufacturing, the core of high-precision size visual detection technology is to complete the quantification of the size parameters of precision parts through image data acquisition and processing, which has become a key link to guarantee product qualification rate and maintain production line stability. This technology is widely used in scenes with strict requirements on size precision, such as automobile manufacturing, semiconductor packaging, and high-end equipment.

[0003] Efficient image data management is a core prerequisite for the landing of high-precision size visual detection technology. Image data management not only solves the problems of traditional image data acquisition chaos and inefficient processing, helps enterprises reduce detection errors caused by data errors, reduces rework and scrap costs, but also promotes the upgrading of detection processes from manual intervention to full-automatic closed loop, replaces traditional manual inspection or contact detection, greatly improves production line detection efficiency, and provides a unified data standard for multi-line collaborative detection, reducing the inconsistency of cross-line detection results.

[0004] However, the existing image data management method and system for high-precision size visual detection rely on single-view image acquisition in the data acquisition link, ignoring the value of multi-view image in covering all-around size information of parts, which is prone to cause local size (such as inside recess of parts, irregular curved surface) missed detection due to visual angle blind area, and cannot match the detection needs of complex structure parts. At the same time, the region segmentation processing mostly uses fixed threshold algorithm, without considering the influence of light change and oil stain interference on image gray distribution in industrial scene, resulting in fuzzy boundary between target region and background region, and even target region disconnection, which lays hidden dangers for subsequent feature extraction. No effective solution has been proposed for the problems in the related art. SUMMARY

[0005] In view of the problems in the related art, the present application proposes an image data management method and system for high-precision size visual detection to overcome the above technical problems existing in the prior art.

[0006] In order to achieve the above purpose, the specific technical scheme adopted by the present application is as follows: According to one aspect of the present application, an image data management method for high-precision size visual detection is provided, comprising the following steps: S1, acquiring a multi-view image data set and a physical size calibration data set by a high-precision size detection device, and performing region segmentation processing on the multi-view image data set; As a preferred solution, the collecting a multi-view image dataset and a physical size calibration dataset by a high-precision size detection device and performing region segmentation processing on the multi-view image dataset comprises the following steps: S11, setting a high-precision industrial camera and a light source system to form a detection device, and adjusting a shooting angle to collect multi-view images of a target object to form a multi-view image dataset; S12, using an existing geometric standard instrument as a calibration reference, and synchronously collecting images and physical size parameters of a standard block under each view to form a physical size calibration dataset; S13, performing region segmentation on the multi-view image dataset by using an adaptive threshold segmentation algorithm to obtain a target region dataset and a background region dataset; As a preferred solution, the performing region segmentation on the multi-view image dataset by using an adaptive threshold segmentation algorithm to obtain a target region dataset and a background region dataset comprises the following steps: S131, removing image noise in the multi-view image dataset; S132, presetting an adaptive threshold calculation window set, and matching the adaptive threshold calculation window set with the denoised multi-view image dataset; S133, extracting gray scale distribution features and target size parameters in the multi-view image dataset based on the matched adaptive threshold calculation window; S134, performing binaryzation segmentation on the gray scale distribution features and the target size parameters to obtain the target region dataset and the background region dataset.

[0007] S14, extracting edge contours of the target region dataset and the background region dataset, and filling internal cavities of the regions.

[0008] As a preferred solution, the extracting edge contours of the target region dataset and the background region dataset, and filling internal cavities of the regions comprises the following steps: S141, presetting an edge detection algorithm library and a morphological processing parameter library, and setting an edge intensity threshold and an internal cavity area threshold of a region; S142, extracting gray scale gradient feature values of the target region dataset and the background region dataset, and selecting an edge detection algorithm in the edge detection algorithm library based on the gray scale gradient feature values; S143, applying the selected edge detection algorithm to the target region dataset and the background region dataset, and performing screening adjustment based on the edge intensity threshold to generate edge contours; S144, identifying internal cavities of the edge contours, and calling a closing operation parameter in the morphological processing parameter library and the internal cavity area threshold of the region to fill internal cavities of the edge contours.

[0009] S2, correcting and enhancing the multi-view image data set after the region segmentation processing, and extracting the contour feature parameters of the multi-view image data set after the correction and enhancement; S3, presetting a feature effectiveness evaluation index, mapping and correlating the contour feature parameters with the physical size calibration data, and screening and classifying the mapped and correlated contour feature parameters according to the feature effectiveness evaluation index; As a preferred solution, the preset feature effectiveness evaluation index, mapping and correlating the contour feature parameters with the physical size calibration data, and screening and classifying the mapped and correlated contour feature parameters according to the feature effectiveness evaluation index includes the following steps: S31, presetting a feature effectiveness evaluation index, and combining the contour feature parameters with the physical size calibration data to obtain a data contour sample set; S32, constructing a contour mapping correlation model by using a fuzzy logic reasoning algorithm, and iteratively optimizing the contour mapping correlation model through the data contour sample set; As a preferred solution, the contour mapping correlation model is constructed by using a fuzzy logic reasoning algorithm, and the contour mapping correlation model is iteratively optimized through the data contour sample set, including the following steps: S321, defining the input variables and output variables of the contour mapping correlation model, and constructing fuzzy sets and membership functions based on the input variables and output variables; S322, setting the contour mapping correlation model according to the fuzzy sets and membership functions; S323, presetting fuzzy rules, and fuzzy processing the data contour sample set based on the fuzzy rules to obtain a fuzzy contour sample set; S324, respectively inputting the fuzzy contour sample set and the data contour sample set into the contour mapping correlation model to calculate fuzzy contour parameters and data contour parameters; S325, presetting a model iteration threshold, calculating contour error parameters of the fuzzy contour parameters and the data contour parameters, and threshold matching the contour error parameters with the model iteration threshold; S326, iteratively optimizing the contour mapping correlation model based on the threshold matching result.

[0010] S33, inputting the contour feature parameters into the iteratively optimized contour mapping correlation model to generate a feature size mapping table, and quantitatively scoring each group of contour feature parameters in the feature size mapping table according to the feature effectiveness evaluation index; As a preferred solution, the contour feature parameters are input into the iteratively optimized contour mapping correlation model to generate a feature size mapping table, and each group of contour feature parameters in the feature size mapping table is quantitatively scored according to the feature effectiveness evaluation index, including the following steps: S331, input the profile feature parameters into the profile mapping correlation model after iteration optimization to calculate the physical size correlation value, and combine the profile feature parameter physical size correlation value to form a feature size mapping table; S332, set feature scoring standards and weight allocation standards for each indicator of the feature effectiveness evaluation indicators; S333, extract the correlation data of each group of profile feature parameters in the feature size mapping table, and calculate the quantitative scores of each evaluation indicator based on the feature scoring standards and weight allocation standards.

[0011] S34, preset scoring screening rules, screen the profile feature parameters after quantitative scoring based on the scoring screening rules, and classify them according to feature types.

[0012] S4, preset a multi-scale feature fusion strategy, and optimize and adjust the screened and classified profile feature parameters according to the multi-scale feature fusion strategy to generate a standardized size detection feature set; As a preferred solution, presetting a multi-scale feature fusion strategy and optimizing and adjusting the screened and classified profile feature parameters according to the multi-scale feature fusion strategy to generate a standardized size detection feature set includes the following steps: S41, preset a multi-scale feature fusion strategy, wherein the multi-scale feature fusion strategy includes a hierarchical fusion strategy library and an alignment method database, and divides a fusion scale level threshold and a feature space alignment threshold; S42, normalize the multi-scale profile feature parameters after screening and classification; S43, extract the fusion scale level value and the profile feature space alignment value of the normalized multi-scale profile feature parameters, and select the hierarchical fusion strategy and the space alignment method in the hierarchical fusion strategy library and the alignment method database based on the fusion scale level value and the profile feature space alignment value; S44, optimize and adjust the profile feature parameters based on the hierarchical fusion strategy and the space alignment method to generate a standardized size detection feature set.

[0013] S5, construct a size detection model based on a deep learning fusion algorithm, and input the standardized size detection feature set into the size detection model for iterative training to obtain a target size quantitative detection result; As a preferred solution, the size detection model is constructed based on a deep learning fusion algorithm, and the standardized size detection feature set is input into the size detection model for iterative training to obtain a target size quantitative detection result, which includes the following steps: S51, preset a deep learning fusion algorithm architecture, wherein the deep learning fusion algorithm architecture includes a network layer configuration library and a training parameter database, and constructs a size detection model according to the network layer configuration library and the training parameter database; S52, data enhancement processing is performed on the standardized size detection feature set, and the standardized size detection feature set is divided into a training set, a verification set and a test set; S53, the training set, the verification set and the test set are input into a size detection model for iterative training; S54, the standardized size detection feature set is input into the size detection model after iterative training, and a target size quantization detection result is obtained.

[0014] S6, accuracy verification is performed on the target size quantization detection result, and an image feature size parameter correlation database is established according to the target size quantization detection result after verification.

[0015] According to another aspect of the application, an image data management system for high-precision size visual detection is provided, which comprises a data acquisition segmentation module, an image correction extraction module, a feature evaluation screening module, a feature optimization generation module, a model construction training module and a result verification database construction module. The data acquisition segmentation module is used to acquire a multi-view image data set and a physical size calibration data set through a high-precision size detection device, and to perform regional segmentation processing on the multi-view image data set. The image correction extraction module is used to perform correction enhancement on the multi-view image data set after regional segmentation processing, and to extract contour feature parameters of the multi-view image data set after correction enhancement. The feature evaluation screening module is used to preset a feature effectiveness evaluation index, to map and correlate the contour feature parameters with the physical size calibration data, and to screen and classify the contour feature parameters after mapping and correlation according to the feature effectiveness evaluation index. The feature optimization generation module is used to preset a multi-scale feature fusion strategy, to optimize and adjust the contour feature parameters after screening and classification according to the multi-scale feature fusion strategy, and to generate a standardized size detection feature set. The model construction training module is used to construct a size detection model based on a deep learning fusion algorithm, to input the standardized size detection feature set into the size detection model for iterative training, and to obtain a target size quantization detection result. The result verification database construction module is used to perform accuracy verification on the target size quantization detection result, and to establish an image feature size parameter correlation database according to the target size quantization detection result after verification.

[0016] The application has the following advantages: 1、The present application solves the problems of incomplete traditional single-view image information and large physical size mapping deviation by synchronous acquisition of multi-view image data set and physical size calibration data set, combined with the calibration reference of geometric quantity standard instrument, while through adaptive threshold segmentation algorithm and edge contour optimization, image noise is removed, target and background area are accurately distinguished, the effectiveness of the original data and the accuracy of the region segmentation are improved, laying a high-quality data foundation for subsequent feature extraction, and introducing fuzzy logic reasoning algorithm to construct contour mapping correlation model, through data sample iterative optimization to realize dynamic mapping of contour feature parameters and physical size calibration data, avoiding the subjectivity of manually setting mapping rules in traditional feature correlation, and combining with the quantitative score and screening rule of feature effectiveness evaluation index, accurately screening high-value contour features, solving the problems of feature redundancy and insufficient effectiveness, and improving the feature quality of the input model.

[0017] 2、The present application solves the problems of strong heterogeneity of different scale features and low fusion adaptation by normalizing and optimizing the screened contour features through a multi-scale feature fusion strategy, while through dividing the fusion scale level threshold and the space alignment threshold, the standardization degree of the feature set is improved, the adaptation ability of the features to different detection scenes is enhanced, and a size detection model is constructed based on a deep learning fusion algorithm, and the standardized feature set is used for iterative training, solving the limitations of traditional detection models in capturing complex image features and low detection precision.

[0018] 3、The present application realizes the traceability and reuse of detection data by accuracy verification and establishment of image feature size parameter correlation database, solves the problems of scattered data management and artificial decision-making in traditional detection, and promotes the automation of high-precision size visual detection from data acquisition to result determination through intelligent data processing and model output in the whole process, improves the efficiency and intelligent level of industrial detection, and through hierarchical verification of training set, verification set and test set, improves the generalization ability of the model to different target objects and different shooting conditions, realizes high-precision quantitative output of target size. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 is a method flow chart of an image data management method for high-precision size visual detection according to an embodiment of the present application; Figure 2is a system block diagram of an image data management system for high-precision size visual inspection according to an embodiment of the present application.

[0021] In the drawings: 1, data acquisition segmentation module; 2, image correction extraction module; 3, feature evaluation and screening module; 4, feature optimization generation module; 5, model construction and training module; 6, result verification and database construction module. DETAILED DESCRIPTION

[0022] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present application, but are not intended to limit the scope of the present application.

[0023] Therefore, the detailed description of the embodiments of the present application provided in the drawings below is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0024] According to an embodiment of the present application, an image data management method and system for high-precision size visual inspection are provided.

[0025] The present application will be further described in conjunction with the drawings and specific embodiments. According to an embodiment of the present application, as shown in Figure 1 The image data management method for high-precision size visual inspection according to an embodiment of the present application includes the following steps: S1, collecting a multi-view image data set and a physical size calibration data set by a high-precision size detection device, and performing region segmentation processing on the multi-view image data set; In the present application, the collection of the multi-view image data set and the physical size calibration data set by the high-precision size detection device and the region segmentation processing on the multi-view image data set include the following steps: S11, setting a high-precision industrial camera and a light source system to form a detection device, and adjusting the shooting angle to collect multi-view images of the target object, forming a multi-view image data set; Specifically, a high-resolution (such as a million-pixel level) industrial camera with a low distortion lens is selected to match the target size detection accuracy requirement; the light source system selects a ring-shaped, strip-shaped or coaxial light source according to the target material (reflective / dull), adjusts the brightness and color temperature to ensure uniform illumination of the target area, avoids high light or shadow interference, and fixes the relative positions of the camera and the light source during device construction, uses a rigid bracket to ensure stability, and calibrates the camera internal and external parameters through a calibration plate to eliminate imaging distortion.

[0026] When adjusting the shooting angle to collect multi-view images, based on the geometric characteristics (such as plane, curved surface, hole distribution) of the target object, at least 3 or more different viewing angles are set through a multi-axis adjusting frame, covering the key detection surfaces such as front, side and oblique, and the adjacent viewing angles need to retain 15%-30% of the feature overlap area to ensure the integrity of subsequent splicing, and the camera parameters (focal length, exposure time) are unified during collection, and the images are collected in batches through a trigger synchronization mechanism, and after removing the blurred and overexposed images, the multi-view image dataset covering the full-dimensional characteristics of the target is formed by classifying and integrating according to the viewing angle.

[0027] S12, using existing geometric quantity standard instruments as calibration reference, and synchronously collecting images and physical size parameters of the standard gauge block under each viewing angle to form a physical size calibration dataset; Specifically, when using existing geometric quantity standard instruments as calibration reference, high-precision standard gauge blocks (such as first-class or second-class gauge blocks) and calibration rules certified by the legal metrology institution need to be selected, and the physical size parameters (length, width, diameter, etc.) thereof need to be clear and accompanied by uncertainty certificates to ensure the authority of the calibration reference.

[0028] Before collection, the standard gauge block is fixed on the same detection platform as the target object, and the placement posture (such as horizontal or vertical) is kept consistent with the subsequent target to be detected, so as to avoid introducing calibration errors due to position deviation. When synchronously collecting, clear images of the standard gauge block are shot under the set shooting viewing angles using the same industrial camera and light source parameters (focal length, exposure time, brightness, etc.) as those used for collecting the multi-view image dataset, so as to ensure the consistency of imaging conditions.

[0029] At the same time, the physical size parameters (accurate to microns) in the standard gauge block certificate are extracted, and the image number corresponding to each viewing angle and the parameter value are recorded to form the correlation data of viewing angle-image-physical size, and finally the physical size calibration dataset is constructed by classifying and integrating according to the viewing angle, removing the samples with blurred images or incorrect parameter matching, and providing an accurate reference for the mapping and correlation of the profile feature and the actual size.

[0030] S13, using an adaptive threshold segmentation algorithm to segment the multi-view image dataset to obtain a target region dataset and a background region dataset; In the embodiment of the application, the adaptive threshold segmentation algorithm is used to segment the multi-view image dataset to obtain a target region dataset and a background region dataset, which includes the following steps: S131, removing image noise in the multi-view image dataset; Specifically, to remove the image noise in the multi-view image dataset, first, the noise type (such as Gaussian noise, salt and pepper noise) and intensity are analyzed by image gray histogram, noise variance calculation and other methods, so as to select the appropriate method for subsequent processing. If it is salt and pepper noise, median filtering is preferred, and by setting a 3x3 or 5x5 filter kernel, the pixel gray value is replaced by the median value of the neighborhood pixels, and isolated noise points are efficiently removed. If it is Gaussian noise, Gaussian filtering is used, and the standard deviation of the filter kernel is adjusted according to the noise variance to smooth the noise while reducing the detail blur.

[0031] For detection scenes that require to retain edge features, bilateral filtering is selected, and through double weight calculation in spatial domain and gray domain, the target edge profile is protected while denoising. The filtering parameters (such as kernel size, standard deviation) of the multi-view images need to be unified during processing to avoid feature deviation of the same target in different views due to parameter differences. Peak signal-to-noise ratio and structural similarity index are used to verify the denoising effect, and images that are still blurred or lose key details after denoising are removed to ensure that the processed multi-view images have no obvious noise and retain high-precision texture and edge information required for size detection.

[0032] S132, a set of adaptive threshold calculation windows is preset, and the set of adaptive threshold calculation windows is matched with the denoised multi-view image dataset; Specifically, when the set of adaptive threshold calculation windows is preset, the core parameters of the window are first determined according to the resolution, target object size and gray distribution characteristics of the denoised multi-view image. Small windows such as 3x3 and 5x5 are selected for small components or edge details, and large windows such as 11x11 and 15x15 are selected for large size targets or uniform background. At the same time, different shaped windows such as rectangles and circles are used to adapt to irregular edges of the target (such as holes and corners). The step size is set to 1-2 pixels to balance the calculation efficiency and local gray capture accuracy. Finally, a window set containing 5-8 groups of differentiated parameters is formed, and the size, shape and step size information of each group of windows are labeled to facilitate subsequent matching and tracing.

[0033] The matching process needs to be executed in three steps: first, the multi-view images after denoising are subjected to gray scale normalization processing, the gray scale values of all images are uniformly mapped to the 0-255 interval, the gray scale deviation caused by the light intensity difference under different angles is eliminated, and a uniform reference is provided for matching; second, the images are processed one by one according to the angle, for a single image, each group of windows in the window set is called to slide on the image with a set step, the local gray scale mean and variance in each window are calculated in real time, and the local gray scale feature matrix of the corresponding window is generated; third, the optimal matching window is selected based on the local gray scale feature matrix: for the area with high proportion of target region in the image, small size window is preferred to match to retain details; for the area with high proportion of background, large size window is selected to improve the anti-interference ability, and the window matching results of the same target under different angles are compared to ensure the compatibility of the window parameters of adjacent angles (such as the size difference is not more than 2 levels), and to avoid the target region segmentation fracture caused by the difference in window selection. After the matching is completed, the optimal window parameters of each image are recorded to provide accurate local threshold calculation basis for subsequent binary segmentation.

[0034] S133, based on the matching, the adaptive threshold calculation window extracts the gray scale distribution features and target size parameters in the multi-view image data set; Specifically, for each denoised image, the matching adaptive threshold calculation window is used as a unit, the image is traversed along the set step, the local gray scale data in each window is collected in real time, the pixel gray scale mean (reflecting the brightness of the region), variance (reflecting the gray scale fluctuation amplitude) and gray scale histogram (presenting the gray scale distribution probability) in the window are calculated, and the window features of the target and background interface area are marked, and the coordinates of each window are recorded to establish the association of window-gray scale feature-coordinate.

[0035] Then, the local features are integrated to form the global gray scale distribution features, the gray scale mean distribution range, variance extreme value and histogram peak interval of all windows of a single image are counted, the gray scale feature difference between the target region (usually continuous gray scale interval) and the background region (gray scale distribution is relatively discrete) is distinguished, the common gray scale features of the same target under different angles (such as the gray scale mean fluctuation range of the target core region) are compared for the multi-view image set, the local feature difference caused by the angle deviation is eliminated, and a unified gray scale distribution feature template is formed.

[0036] The target size parameter is extracted, the target region in each view image is located based on the gray distribution feature template, the size (known pixel number) of the matching window is taken as a benchmark, the key size parameter of the target is calculated by combining the pixel-physical size conversion coefficient (such as 1 pixel corresponding to 0.01 mm) obtained by camera calibration, the pixel span of the target edge is determined through the window coordinates, and the physical length and width are converted, for circular, hole and other features, the contour is fitted through the peak area of the window gray histogram, and the diameter, radius and other parameters are calculated, and the deviation of the same size parameter under different views is checked (controlled within the detection accuracy threshold), the parameter consistency is ensured, and the structured data of view-gray distribution feature-target size parameter is formed.

[0037] In S134, the gray distribution feature and the target size parameter are binarized and segmented to obtain a target region data set and a background region data set.

[0038] Specifically, the initial binarization threshold is determined based on the gray distribution feature, the global gray mean, variance and window local gray histogram extracted in the early stage are combined, the dynamic threshold method is used to calculate the basic segmentation threshold, the image pixels are divided into two categories of foreground (suspected target) and background, for the area with uneven gray distribution (such as the transition area between the target edge and the background), the local gray feature of the matching window is further called to set differential sub-thresholds for different areas, so that the edge segmentation fracture caused by the global threshold is avoided, and the initial binarization image is obtained.

[0039] Secondly, the target size parameter is introduced to optimize the segmentation result, the target size parameter (such as the minimum / maximum physical area, length range) extracted in the early stage is used to establish a size screening rule, the pixel area (combined with the pixel-physical size conversion coefficient) of the connected region in the initial binarization image is calculated, the noise region (such as dust and light spot) with an area smaller than the minimum size of the target and the background region (such as the edge of the detection platform) with an area larger than the maximum size of the target are removed, and the matching degree of the shape feature (such as the aspect ratio) of the connected region and the target size parameter is checked, the interference region (such as irregular background shadow) with an inconsistent shape is excluded, and the suspected target region is preliminarily separated.

[0040] Then, the segmentation is improved through multi-view checking, morphological operations (first corrosion to remove small burrs, and then inflation to fill small holes in the region) are performed on the suspected target region, the region integrity after segmentation is repaired, the segmentation results of the same target under different views are compared, the size and contour consistency of the target region in each view is ensured (the deviation is controlled within the detection accuracy threshold), the single-view misjudgment is avoided, finally, the connected regions meeting the gray and size double standards are classified as the target region data set, and the remaining regions are classified as the background region data set according to the view classification and integration.

[0041] In S14, the edge contour of the target region data set and the background region data set is extracted, and the internal holes of the regions are filled.

[0042] In the embodiments of the present application, the edge profile of the target region data set and the background region data set is extracted, and the internal cavity of the region is filled, including the following steps: S141, presetting an edge detection algorithm library and a morphological processing parameter library, and setting an edge intensity threshold and an internal cavity area threshold of a region; Specifically, the edge detection algorithm library and the morphological processing parameter library need to be preset according to the scene demand of high-precision size detection, and the two types of thresholds are scientifically set according to the algorithm / parameter classification and scene adaptation logic construction.

[0043] When the edge detection algorithm library is preset, multiple mainstream algorithms need to be included and the applicable scene is specified: the core includes Canny algorithm (adapted to weak edge and noise interference scene, adjustable parameters include high and low threshold, Gaussian filter kernel size), Sobel algorithm (adapted to scene requiring to distinguish horizontal / vertical edge, adjustable parameters include convolution kernel size, direction weight), Laplacian algorithm (adapted to scene requiring to detect zero-crossing edge, adjustable parameters include filter kernel type, gray level change sensitivity); at the same time, the calling priority of each algorithm is labeled (such as preferentially using Canny to process complex contour target, and using Sobel to process regular rectangular / circular target), and an algorithm parameter fine-tuning interface is reserved for subsequent adaptation according to image features.

[0044] When the morphological processing parameter library is preset, the parameters need to be stored according to the operation type: for the erosion operation, 3x3, 5x5, etc. Different size rectangular / circular structural elements (small kernel adapts to remove small burrs, and large kernel adapts to eliminate wide interference) and 1-2 times of iteration number are stored; for the dilation operation, the structural elements matched with the erosion are stored (used to repair the edge shrinkage caused by erosion); for the closing operation (filling cavity), 5x5-11x11 structural elements and 2-3 times of iteration number are stored; all parameters need to be associated with target size parameters (such as the size of the structural element is not more than 1 / 5 of the minimum pixel size of the target), to avoid excessive processing damage to the target contour.

[0045] When setting the edge intensity threshold, first count the average gray scale gradient of the target edge in the multi-view sample image (e.g. 100-150), set the lower limit threshold of the edge intensity (e.g. 60) as 60% of the average, filter the noise pseudo-edge; set the upper limit threshold (e.g. 180) as 120% of the average, retain the strong edge; the middle interval is supplemented by edge continuity check (e.g. the number of continuous edge pixels is greater than or equal to 5). When setting the area internal cavity area threshold, based on the target minimum physical area, set the filling threshold as 1% of the target minimum area (e.g. 1 pixel) to ensure that small cavities are eliminated; at the same time, set the upper limit threshold (e.g. 5% of the target minimum area) to avoid filling the target's own hole (e.g. part assembly hole), and adjust the threshold through multi-view cavity consistency check (the same cavity appears in ≥3 views) to form a reusable threshold standard.

[0046] S142, extract the gray scale gradient feature value of the target region data set and the background region data set, and select an edge detection algorithm in the edge detection algorithm library based on the gray scale gradient feature value; Specifically, the target and background region data sets are preprocessed: the images of the two types of regions are uniformly converted to gray scale images and normalized by mean value to eliminate the gray scale deviation caused by different view light sources; then a 3x3 Sobel gradient operator is used to traverse the region pixels, the gradient components of each pixel in x (horizontal) and y (vertical) directions are calculated, the gray scale gradient amplitude matrix and direction matrix of the whole region are obtained, then the core gradient feature values are extracted: the gradient amplitude range of the target region (e.g. 50-200, reflecting the edge strength), the gradient direction distribution (e.g. concentrated in 0° / 90°, indicating that the edge is mostly horizontal / vertical) and the edge density (the proportion of pixels in the target region whose gradient amplitude exceeds the threshold, reflecting the edge richness), and the gradient features of the background region (usually low amplitude and discrete distribution) are compared, the unique features of the target edge are determined, and a region type-gradient feature value table is formed.

[0047] When selecting an edge detection algorithm based on gray gradient characteristic value, the algorithm characteristics and feature requirements need to be matched: if the overall gradient amplitude of the target area is low (there are many weak edges) and contains a small amount of noise, the Canny algorithm is preferentially called from the algorithm library (its Gaussian filter can reduce noise, and the double threshold can retain weak edges), and the high threshold of the algorithm is set to 80% of the upper limit of the target gradient amplitude, and the low threshold is set to 50% of the lower limit. If the target gradient direction is concentrated (such as the rectangular edge of a part), the Sobel algorithm is selected (which can calculate the horizontal / vertical gradient in a targeted manner and strengthen the directional edge), and the corresponding convolution kernel is matched. If the gradient of the target area changes sharply (such as sharp corners), and there is no obvious noise, the Laplacian algorithm is selected (which can quickly detect zero-crossing edges and does not need to be directionally filtered). After selection, the edge extraction effect of the single-view image needs to be tested to ensure edge continuity (the number of broken pixels is less than or equal to 3) and completeness (covering more than 95% of the target outline), and then it is extended to the entire data set to form an adaptive rule for gradient feature algorithms.

[0048] S143, apply the selected edge detection algorithm to the target area data set and the background area data set, and perform screening and adjustment based on the edge intensity threshold to generate an edge contour; Specifically, first, pre-processing before application is performed: the images of the target and background area data sets are uniformly adjusted to the same resolution, and the gray values are simultaneously normalized to the 0-255 interval, eliminating algorithm response deviations caused by multi-view light source differences or inconsistent resolutions; for the target area data set, a region mask is additionally generated (only the target pixels are retained, and the background pixels are shielded), to avoid background noise interference with the capture of the target edge by the algorithm, and to provide a pure data environment for algorithm application.

[0049] Then, the edge detection algorithm is applied: the Canny algorithm is selected, the high and low thresholds of the algorithm are set based on the gray gradient characteristic value (such as the gradient mean) of the target area (the high threshold is 1.2 times the gradient mean, and the low threshold is 0.5 times the gradient mean), and then Gaussian filter denoising, gradient amplitude and direction calculation, and non-maximum suppression (edge thinning) are performed on the target area image with the mask, and finally the double threshold edge connection is performed. If the Sobel algorithm is selected, the corresponding convolution kernel (x direction or y direction) is selected for the main direction of the target edge (such as horizontal / vertical), the target area image is convolved to calculate the gradient response, and the directional edge signal is strengthened. The background area data set is only processed by the basic algorithm, which is mainly used for comparison and exclusion of interference contours other than the target edge.

[0050] Subsequently, edge intensity threshold screening adjustment: call the preset edge intensity lower threshold (such as gradient amplitude ≥ 60), filter the pseudo edges (such as false response generated by background micro noise) below the threshold in the algorithm output; check the continuity of the screened edge pixels (continuous edge pixel number ≥ 5), and remove isolated edge points; if there is a broken edge area (such as a weak edge segment), compare the edge distribution of adjacent viewing angles, and complete the broken part by interpolation. Finally, perform post-processing: perform morphological thinning operation on the screened edge pixels to remove redundant edge pixels and generate clear contours of single-pixel width; check the matching degree of the contour and the target area (contour area deviation from target area area ≤ 3%), and ensure that there is no obvious deviation, and finally obtain complete and accurate target edge contour, and the background area retains its basic edge information for subsequent comparison and analysis.

[0051] S144, identify the internal cavity of the edge contour, and call the closing operation parameters in the morphological processing parameter library and the internal cavity area threshold to fill the internal cavity of the edge contour.

[0052] Specifically, when identifying the internal cavity, first perform connected region analysis on the generated single-pixel edge contour to mark all closed regions within the contour; through contour level judgment (outer contour as parent level, inner contour as child level), filter out the child level regions that are completely surrounded by the outer contour and not connected with the background area, and preliminarily determine them as internal cavities; then, combined with the preset internal cavity area threshold, calculate the pixel area of each suspected cavity (through pixel counting method), remove the regions with area exceeding the upper limit of the threshold (such as the functional holes of the target itself, which need to be associated with the target size parameter in advance), and retain the real cavities with area within the threshold range (such as 1-20 pixels) to form a cavity position-area table.

[0053] Then call the closing operation parameters in the morphological processing parameter library: match the corresponding parameters according to the area size of the cavities in the table, use the closing operation parameters of 3×3 circular structural elements and 1 iteration in the library for small cavities with area ≤ 10 pixels to avoid excessive expansion of the edge during filling; use the parameters of 5×5 circular structural elements and 2 iterations for larger cavities with area 10-20 pixels to ensure complete filling of the cavities; if there is an irregularly shaped cavity, preferentially select a rectangular or elliptical structural element similar to the cavity contour to reduce the deformation of the contour after filling.

[0054] Finally, the closing operation is performed to fill and check the effect: the matching closing operation parameters are applied to the edge contour image, the contour pixels are first expanded by the dilation operation, the holes and openings are closed, and then the original size of the contour is restored by the erosion operation to complete the hole filling; after filling, the area of the region inside the contour is recalculated to check whether the holes are completely eliminated, and the edge contour accuracy before and after filling is compared (such as the edge pixel coordinate deviation ≤1 pixel) to avoid distortion of the target size caused by filling; for multi-view images, the hole filling effect of the same target is synchronously checked to ensure the consistency of the filling of each view, and finally the target edge image with no internal hole and complete contour is obtained to provide an accurate contour basis for subsequent feature extraction.

[0055] S2, the multi-view image data set after region segmentation processing is corrected and enhanced, and the contour feature parameters of the multi-view image data set after correction and enhancement are extracted; Specifically, in the correction and enhancement stage, geometric correction is first performed: the intrinsic parameters (focal length, principal point coordinates) and extrinsic parameters (rotation matrix, translation vector) obtained by camera calibration are called to project each view image to a unified world coordinate system through perspective transformation, eliminate the scale deviation caused by lens distortion (radial / tangential distortion) and view angle tilt, and ensure that the pixel span of the same physical size under different views is consistent; for the corrected image, a light enhancement algorithm is used to optimize the image quality, CLAHE (Contrast Limited Adaptive Histogram Equalization) is applied to the area with uneven lighting to improve the local contrast while avoiding noise amplification; gamma correction (γ value 0.6-0.8) is performed on the low brightness area, and gray scale stretching is performed on the overexposed area, so that the gray difference between the target and the background is more significant, and an enhanced image set with aligned views and clear details is obtained.

[0056] Contour feature parameter extraction needs to be based on the enhanced image: first, the target closed contour is located through the edge detection result, and the chain code is used to record the continuous coordinates of the contour pixels; then, the basic geometric parameters are calculated, including the contour perimeter (pixel distance accumulation), the enclosed area (using the polygon area formula), the longest / shortest axis length (through the minimum circumscribed rectangle fitting), and the axis ratio; shape feature parameters such as the number of curvature extreme points (inflection points), the average curvature (reflecting the contour smoothness), and the convex hull defects (identifying the concave area) are extracted; the moment features (such as Hu moment) are calculated to describe the overall shape of the contour, and the parameter correspondence of the same contour under different views (such as the mapping value of the longest axis in the world coordinate system) is recorded. After extraction, the parameters are normalized (scaled according to the maximum size of the target), and abnormal values (such as parameters exceeding 3 times the standard deviation) are removed to form a contour feature parameter set containing geometric shape multi-view related information.

[0057] S3, a preset feature effectiveness evaluation index is used to map and associate the contour feature parameters with the physical size calibration data, and the mapped and associated contour feature parameters are classified and selected according to the feature effectiveness evaluation index; In the embodiments of the present application, the preset feature effectiveness evaluation index is mapped and associated with the physical size calibration data, and the profile feature parameters after mapping and association are screened and classified according to the feature effectiveness evaluation index, including the following steps: S31, a preset feature effectiveness evaluation index is mapped and associated with the physical size calibration data, and the profile feature parameters after mapping and association are screened and classified according to the feature effectiveness evaluation index, including the following steps: Specifically, when the evaluation index is preset, the core dimension and the calculation method are first determined: first, the mapping error rate is used to measure the association accuracy of the profile feature parameters and the physical size, the absolute deviation of the predicted physical size corresponding to a single set of profile feature parameters (such as profile perimeter, curvature extreme value) from the physical size calibration data is calculated, and the percentage of the absolute deviation in the calibration data is set to be less than or equal to 1% for excellent, and 1%-3% for good; second, the multi-angle consistency is used to evaluate the stability of the same feature under different angles, the coefficient of variation (standard deviation / mean) of the same feature parameter under multiple angles is calculated, and the coefficient of variation is less than or equal to 5% for qualified, so as to avoid feature failure caused by angle deviation; third, the size sensitivity is used to test the response ability of the feature to the change of the physical size, the change rate of the corresponding feature parameter is calculated by slightly adjusting the size of the standard block (such as ±0.01mm), and the change rate is greater than or equal to 80% for high sensitive feature. At the same time, the weights of the indexes (mapping error rate 40%, consistency 30%, sensitivity 30%) are allocated to form a complete evaluation system.

[0058] When the combined data generates the sample set, the corresponding relationship of angle-feature-calibration data is first established: according to the same shooting angle, the corrected and enhanced profile feature parameters (such as perimeter, number of inflection points, Hu moment) are matched with the physical size calibration data (such as actual length, diameter, angle) under the same angle one by one, so that a single sample contains 1 set of angle information + 5-8 profile feature parameters + corresponding physical size value; then the data is integrated through a structured table, and the parameter source (such as image number, detection area) and the calibration data accuracy level (such as first-class block ±0.001mm) are labeled; finally, data screening is performed to remove samples with mapping error rate >5% and multi-angle consistency not meeting the standard, and feature-calibration data pairs of different size specifications of standard blocks are supplemented to ensure that the sample covers the key size nodes in the detection range, and finally a data profile sample set with diversity and accuracy is formed, which provides high-quality training data for subsequent profile mapping and association model construction.

[0059] S32, a fuzzy logic reasoning algorithm is used to construct a profile mapping and association model, and the profile mapping and association model is iteratively optimized through the data profile sample set. In the embodiments of the present application, the fuzzy logic reasoning algorithm is used to construct a profile mapping and association model, and the profile mapping and association model is iteratively optimized through the data profile sample set, including the following steps: S321, define the input variables and output variables of the contour mapping correlation model, and construct fuzzy sets and membership functions based on the input variables and output variables; Specifically, when defining the input variables and output variables of the contour mapping correlation model, the core parameter category is determined: the input variables are the contour feature parameters extracted after correction enhancement, including geometric features (such as contour perimeter, enclosed area, longest axis length), shape features (such as the number of inflection points, average curvature, and the number of convex hull defects), and the value range of each parameter (such as perimeter 50-500 pixels, inflection points 2-15) needs to be determined based on the data contour sample set; the output variables are the corresponding physical size calibration data, such as actual length (0.5-50mm), diameter (0.3-30mm), and angle (30°-150°), and the value range needs to match the calibration range of the standard gauge block.

[0060] When constructing fuzzy sets and membership functions based on input and output variables, first divide the fuzzy subsets according to the variable distribution characteristics: the input variables such as contour perimeter are divided into three subsets of small, medium and large, and the output variables such as actual length are divided into three subsets of short, medium and long, the subset boundary needs to cover the full range of variables and retain 10%-20% overlap to ensure smooth transition. Then configure the membership function for each subset: for variables in the data distribution set (such as standard part perimeter), use a triangular function, with the vertex corresponding to the sample statistical mean (such as the medium perimeter vertex being the sample average perimeter), and the bottom edge covering the ±2 standard deviation range; for variables with discrete distribution (such as complex contour curvature), use a Gaussian function, with the mean set to the center value of the subset and the standard deviation adjusted according to the data dispersion degree (such as 10%-15% of the center value). Finally, the variable-fuzzy subset-membership function correlation table is formed.

[0061] S322, set the contour mapping correlation model according to the fuzzy sets and membership functions; Specifically, the input and output variables of the model are determined: the input is the contour feature parameter (such as perimeter, number of inflection points, curvature), and the output is the corresponding physical size (such as actual length, diameter), and the variables are divided into subsets based on fuzzy sets (such as input perimeter divided into small / medium / large, and output length divided into short / medium / long), ensuring that the subsets cover the full range of parameters.

[0062] Then define the membership function: for each variable subset, select a triangular or Gaussian membership function (such as a left-biased triangle for small perimeter and a symmetric triangle for medium perimeter), and adjust the vertex coordinates (such as the small perimeter vertex set to the sample minimum perimeter value) and width according to the parameter distribution in the data contour sample set, so that the function can accurately describe the degree to which the variable belongs to each subset, forming a variable-fuzzy subset-membership function correspondence table.

[0063] Constructing fuzzy reasoning mechanism: based on the association between variable subsets (such as if the perimeter fuzzy value is medium, the length fuzzy value is medium), preset fuzzy rule base; configure the reasoning algorithm to input the characteristic parameters after fuzzy by membership function, and substitute into the rule base for logical reasoning, output fuzzy results; At the same time, set the defuzzification method (such as the barycenter method), convert the fuzzy results into accurate physical size value, complete the model infrastructure setting, and ensure that the model can realize the fuzzy mapping of contour features to physical size.

[0064] S323, preset fuzzy rules, and fuzzy processing of data contour sample set based on fuzzy rules to obtain fuzzy contour sample set; Specifically, the preset fuzzy rules need to be based on the fuzzy set association law of input and output variables, combined with domain experience and sample data characteristics: taking the corresponding relationship between input variables (such as contour perimeter small / medium / large) and output variables (such as actual length short / medium / long) as the core, constructing if-then form rules, for example, if the contour perimeter fuzzy value is small, the actual length fuzzy value is short, and if the number of inflection points fuzzy value is medium, the angle fuzzy value is medium. The rules need to cover the main feature association scenarios (the number is usually 2-3 times the number of input subsets), and avoid conflicts (such as the same input condition corresponding to a unique output tendency), forming a structured fuzzy rule base.

[0065] When fuzzy processing of data contour sample set based on fuzzy rules, accurate parameters need to be converted sample by sample: for each group of input (such as perimeter 50 pixels) and output (such as length 0.5mm) in the sample, call the preset membership function to calculate the membership degree of each fuzzy subset (such as 50 pixels belonging to small perimeter with membership degree 0.9, and medium perimeter with membership degree 0.1), generate a three-dimensional fuzzy vector of parameter-fuzzy subset-membership degree; Then match the association between input fuzzy vector and output fuzzy vector according to the rule base, supplement the rule adaptation mark, integrate the fuzzy vector of all samples and the rule matching results, classify according to the angle and features, and form the fuzzy contour sample set containing fuzzy parameters and associated rule information.

[0066] S324, respectively input the fuzzy contour sample set and the data contour sample set into the contour mapping association model to calculate the fuzzy contour parameters and the data contour parameters; Specifically, when inputting the fuzzy contour sample set and the data contour sample set to calculate the corresponding parameters respectively, the fuzzy reasoning and defuzzification mechanism of the model need to be combined to process according to the differentiated process: When the input fuzzy profile sample set is input, the input fuzzy vector in each group of the sample (such as the membership degree distribution of the profile perimeter, the number of inflection points) is extracted and substituted into the profile mapping correlation model. The model calls the preset fuzzy rule library, matches the input fuzzy vector with the rule through the Mamdani inference method, calculates the fuzzy relation matrix, generates the fuzzy set of the output end (such as the short / medium / long membership degree distribution corresponding to the physical size), directly takes the fuzzy set as the fuzzy profile parameter, and temporarily stores the correlation data of the sample number-fuzzy set, thereby retaining the integrity of the fuzzy characteristics.

[0067] When the input data profile sample set is input, the accurate profile feature parameters (such as a perimeter of 50 pixels, 8 inflection points) in the sample are extracted first. The model converts the parameters into fuzzy vectors through a preset membership function (completes the fuzzification of the accurate values), and then repeats the above fuzzy inference process to obtain the output fuzzy set. Subsequently, the defuzzification module is started, the barycentric method is used to calculate the barycenters of the fuzzy set, the barycenters are converted into accurate physical size values (such as a length of 0.5 mm), and the accurate values are the data profile parameters. Finally, the two groups of parameters are stored according to the sample type, and a corresponding relationship table of the fuzzy profile parameters and the data profile parameters is formed, thereby laying a foundation for subsequent calculation of the profile error parameters.

[0068] S325, a preset model iteration threshold is set, and the profile error parameters of the fuzzy profile parameters and the data profile parameters are calculated, and the profile error parameters are matched with the model iteration threshold; Specifically, the preset model iteration threshold needs to be combined with the detection accuracy requirement and the data characteristics. The uncertainty (such as ±0.001 mm) of the physical size calibration data in the data profile sample set is taken as a reference, the maximum deviation allowed by the model inference is superimposed, the iteration threshold is set to 1.2-1.5 times (such as ±0.0015 mm) of the calibration accuracy, and an error rate threshold (such as ≤1.5%) is supplemented, thereby forming a double-dimensional threshold standard of absolute error and relative error, and ensuring that the threshold covers a reasonable error range and can constrain the model accuracy.

[0069] When the profile error parameters are calculated, the fuzzy profile parameters (fuzzy set) are defuzzified through the barycentric method to convert into accurate physical size values. Then, the absolute error (|fuzzy converted value-data parameter value|) and the relative error (absolute error / data parameter value x 100%) of each sample are calculated based on the data profile parameters (accurate physical size). The mean and variance of the errors of multiple samples are calculated, and a profile error parameter set containing the error of a single sample and the overall error distribution is formed.

[0070] When the threshold is matched, the absolute error and the relative error of each sample are compared with the preset iteration threshold respectively: if both errors are less than the threshold, it is determined that the matching is qualified, and the sample is used for model verification; if any error exceeds the threshold, it is marked as unqualified, and the sample characteristics (such as a specific viewing angle and a characteristic parameter) exceeding the error are recorded to provide error traceability basis for subsequent model iteration optimization, and the threshold matching process is completed.

[0071] S326, based on the threshold matching result, iteratively optimizing the contour mapping correlation model.

[0072] S33, inputting the contour characteristic parameters into the contour mapping correlation model after iteration optimization, generating a feature size mapping table, and quantitatively scoring each group of contour characteristic parameters in the feature size mapping table according to the feature validity evaluation index; In the embodiment of the present application, the inputting of the contour characteristic parameters into the contour mapping correlation model after iteration optimization, the generation of the feature size mapping table, and the quantitative scoring of each group of contour characteristic parameters in the feature size mapping table according to the feature validity evaluation index include the following steps: S331, inputting the contour characteristic parameters into the contour mapping correlation model after iteration optimization to calculate the physical size correlation value, and combining the contour characteristic parameter physical size correlation value to form a feature size mapping table; Specifically, inputting the contour characteristic parameters into the contour mapping correlation model after iteration optimization and forming the feature size mapping table need to follow the process of parameter preprocessing-model calculation-structured combination to ensure correlation accuracy and data specification. First, the contour characteristic parameters are processed: the corrected and enhanced parameters (such as contour perimeter, number of inflection points, and curvature extreme value) are extracted, normalized according to the parameter range (such as perimeter 0-1000 pixels, inflection points 2-20) during model training, and the dimensional difference is eliminated; at the same time, abnormal values (such as parameters exceeding 3 times the standard deviation) are removed to avoid interfering with model reasoning, and an adapted feature parameter set is obtained.

[0073] Then, the model calculates the physical size correlation value: after the iteration optimized model is substituted into the adapted parameters, the model first completes the parameter fuzzification, and then outputs the corresponding physical size correlation value (such as a perimeter of 500 pixels corresponding to a length of 5 mm, and 8 inflection points corresponding to a diameter of 3 mm) through accurate reasoning, and the correlation value accuracy needs to meet the preset detection threshold (such as absolute error ≤0.002 mm).

[0074] Finally, the feature size mapping table is formed: according to the structure of sample number-profile feature parameter group-physical size associated value-shooting angle, each group of parameters is integrated one by one corresponding to the associated value, the parameter group needs to completely contain geometric and shape features (such as circumference, number of convex hull defects), and the associated value is marked with units. At the same time, the data in the table is checked, some associated values are compared with the physical size calibration data, and items with error exceeding the limit are removed, and finally a structured and high-precision feature size mapping table is formed.

[0075] S332, setting feature score criteria and weight allocation criteria for each of the feature validity evaluation indicators; Specifically, the core evaluation indicators (mapping error rate, multi-angle consistency, and size sensitivity) are determined, and then the scoring criteria are developed one by one: for the mapping error rate, according to the error value, three grades are divided: ≤1% (feature and physical size association is accurate) is 100 points, 1%-3% (association interval) is 60 points, and >3% (association failure) is 0 points. At the same time, features with error exceeding 3% are directly removed; for multi-angle consistency, based on the coefficient of variation, ≤5% (feature is stable across view angles) is 100 points, 5%-10% (slight fluctuation) is 60 points, and >10% (fluctuation is severe) is 0 points; for size sensitivity, according to the parameter change rate, ≥80% (sensitive to size change) is 100 points, 60%-80% (general response) is 60 points, and <60% (response is sluggish) is 0 points.

[0076] The weight allocation needs to be combined with the influence of the indicators on the detection accuracy: the mapping error rate directly determines the size association accuracy, and the weight is set to 40% (the highest); the multi-angle consistency affects the reliability of the feature across view angles, and the weight is set to 30%; the size sensitivity guarantees the ability of the feature to capture subtle size changes, and the weight is set to 30%, and the sum of the three is 100%. Finally, the rationality is verified through expert review or sample test, if a certain type of feature (such as a curved part) requires higher sensitivity, the sensitivity weight is adjusted to 35%, and the consistency weight is correspondingly reduced to 25%, to ensure that the standard adapts to the actual detection scene.

[0077] S333, extracting the associated data of each group of profile feature parameters in the feature size mapping table, and calculating the quantitative scores of each evaluation indicator based on the feature score criteria and weight allocation criteria.

[0078] Specifically, each set of contour feature parameters (such as circumference, number of inflection points, curvature) is retrieved from the mapping table one by one, its corresponding physical size correlation value (such as actual length, diameter) is synchronously associated, and the physical size calibration data (for calculating error) of the source of the parameter set, the parameter set of the same feature under multiple perspectives (for calculating consistency), and the change value of the feature parameter after size fine-tuning (such as ±0.01 mm) (for calculating sensitivity) are matched to form a complete correlation data set of contour feature parameters-physical size correlation value-calibration data-multiple perspective parameters-fine-tuning change value, ensuring that each set of data can be traced back to a specific sample and perspective.

[0079] The calculation of the quantitative score needs to be divided into three steps: first, calculate the single indicator score according to the feature scoring standard, calculate the mapping error rate by / physical size correlation value-calibration data / calibration data x 100%, and get 100 / 60 / 0 points according to the standard; the consistency of multiple perspectives is calculated by the coefficient of variation (standard deviation / mean) of the multiple perspective parameters, and the corresponding score is obtained according to the standard; the size sensitivity is calculated by / feature parameter after fine-tuning-original parameter / original parameter x 100%, and the score is matched according to the standard. Second, combine the weight allocation standard (such as mapping error rate 40%, consistency 30%, sensitivity 30%) to multiply the single indicator score by the corresponding weight. Third, sum to get the comprehensive quantitative score of each parameter set (such as 100 x 40% + 80 x 30% + 90 x 30% = 91 points), and eliminate invalid feature parameter sets with a comprehensive score less than 60 points to ensure that the output score accurately reflects the effectiveness of the feature.

[0080] S34, a preset scoring screening rule is set, and the contour feature parameters after quantitative scoring are screened based on the scoring screening rule and classified according to the feature type.

[0081] Specifically, the preset scoring screening rule needs to be based on the core of guaranteeing the effectiveness and practicality of the feature, and a double screening standard is developed in combination with the quantitative score dimension: first, the comprehensive score threshold, which sets the comprehensive score ≥60 points as the basic pass line to ensure that the feature as a whole meets the detection requirements; second, the single indicator bottom line, which requires that the mapping error rate, the consistency of multiple perspectives, and the size sensitivity cannot be 0 points, and the score of the mapping error rate (core indicator) is ≥60 points to avoid features with key dimensions being invalid entering the subsequent process; at the same time, the multi-perspective stability clause is supplemented, and the score of the same feature under ≥80% perspective needs to be ≥60 points to exclude unstable features that are accidentally qualified under single perspective.

[0082] The screening is performed according to the process of first overall and then local: first, batch filter contour feature parameter sets with a comprehensive score <60 points to preliminarily eliminate invalid features; second, check the single indicators of the remaining qualified sets one by one to eliminate parameter sets with 0 points indicators or mapping error rate <60 points; third, verify the multi-perspective score distribution of the same feature, and delete features with a perspective qualified proportion <80%.

[0083] The classification stage is divided by feature type: geometric features such as contour perimeter, enclosed area, and longest axis length are classified into one category, and shape features such as the number of inflection points, average curvature, and convex hull defect number are classified into another category; the average comprehensive score, the highest score, and the proportion of qualified viewing angles of each feature are annotated, forming a classification table of feature type-parameter name-score information, which facilitates subsequent on-demand calling of high-effectiveness features.

[0084] S4, a multi-scale feature fusion strategy is preset, and the profile feature parameters after screening and classification are optimized and adjusted according to the multi-scale feature fusion strategy to generate a standardized size detection feature set; In the embodiments of the present application, the multi-scale feature fusion strategy is preset, and the profile feature parameters after screening and classification are optimized and adjusted according to the multi-scale feature fusion strategy to generate a standardized size detection feature set, which includes the following steps: S41, a multi-scale feature fusion strategy is preset, wherein the multi-scale feature fusion strategy includes a hierarchical fusion strategy library and an alignment method database, and a fusion scale level threshold and a feature space alignment threshold are divided; Specifically, two basic libraries are first constructed: one is a hierarchical fusion strategy library, which fuses features according to their scales. Low-scale features (such as edge burrs and small hole features) are fused at the pixel level (weighted average method, with weights assigned according to feature scores) to preserve details; medium-scale features (such as local corners and convex hull structures) are fused at the feature level (feature splicing method, with splicing dimensions including geometry and shape parameters) to enhance local correlation; high-scale features (such as overall contour and bounding box) are fused at the decision level (voting method, with weights assigned according to multi-view feature effectiveness scores) to ensure global consistency. The applicable feature types and computational complexity of each strategy are also labeled. The second is an alignment method database, which includes three types of methods: spatial alignment (perspective transformation, suitable for coordinate deviation after geometric correction), feature alignment (Hu moment registration, suitable for shape feature distribution differences), and grayscale alignment (histogram matching, suitable for grayscale deviation caused by uneven lighting). The applicable scenarios (such as perspective transformation for multi-view geometric feature alignment) and precision indicators (such as spatial alignment pixel deviation ≤2) of each method are also specified.

[0085] Then, key thresholds are divided: the fusion scale level threshold is divided according to feature pixel size, with a low-scale threshold <10 pixels (corresponding to detail features), a medium-scale threshold 10-50 pixels (corresponding to local features), and a high-scale threshold >50 pixels (corresponding to global features). Each threshold is associated with the corresponding method in the strategy library. The feature space alignment threshold has double standards: the spatial alignment pixel deviation threshold is ≤2 pixels (to ensure geometric position consistency), and the Hu moment distance threshold for feature alignment is ≤0.1 (to ensure similar shape distribution). If the threshold is exceeded, the optimization process in the alignment method library is triggered to ensure fusion accuracy.

[0086] S42, normalize the multi-scale contour feature parameters after screening and classification; S43, extract the fusion scale level value and contour feature space alignment value of the multi-scale contour feature parameters after normalization, and select the level fusion strategy and space alignment method in the level fusion strategy library and alignment method database based on the fusion scale level value and contour feature space alignment value; Specifically, when extracting the fusion scale level value, first calculate the actual pixel span of each contour feature parameter after normalization (such as the pixel length of local burr and the pixel surrounding range of overall contour), and then compare the span with the preset fusion scale level threshold (low <10 pixels, medium 10-50 pixels, high >50 pixels) to determine the level value of each type of feature, for example, pixel span 6 corresponds to low scale, span 35 corresponds to medium scale, forming an association table of features and level values.

[0087] When extracting the contour feature space alignment value, calculate from the two dimensions of geometry and shape: in the geometric dimension, compare the coordinate deviation of the same feature under multiple perspectives (based on the world coordinate system of camera calibration) to obtain the spatial position alignment value; in the shape dimension, calculate the Hu moment distance of the features under multiple perspectives (measure the shape distribution similarity) to obtain the shape alignment value, and the two together constitute the contour feature space alignment value (such as coordinate deviation 1.5 pixels, Hu moment distance 0.08).

[0088] When selecting the strategy and method, match the level value with the strategy and the alignment value with the method: if the level value is low scale, select pixel-level weighted average method from the level fusion strategy library; if the level value is medium scale, select feature-level splicing method; if the level value is high scale, select decision-level voting method. If the coordinate deviation in the spatial alignment value is ≤2 pixels, select perspective transformation from the alignment method database; if the Hu moment distance is ≤0.1, select Hu moment registration; if the value exceeds the threshold, call the optimization method in the library (such as iterative adjustment of perspective transformation parameters) to ensure the fusion and alignment accuracy.

[0089] S44, based on the level fusion strategy and the space alignment method, optimize and adjust the contour feature parameters to generate a standardized size detection feature set.

[0090] Specifically, first perform spatial alignment optimization: based on the selected spatial alignment method (such as perspective transformation or Hu moment registration), correct the coordinates and shapes of the multi-perspective contour feature parameters, map the coordinates of geometric features under different perspectives to a unified world coordinate system through perspective transformation to make the spatial position deviation ≤2 pixels; for shape features, adjust the contour distribution using Hu moment registration to ensure that the shape alignment value (Hu moment distance) ≤0.1, and eliminate feature misplacement caused by perspective deviation.

[0091] Then integrate the features according to the hierarchical fusion strategy: low-scale features (such as edge burrs) are fused at the pixel level, and the weight is allocated according to the feature score (the weight of high-score features accounts for more than 60%), and the detail information is retained; medium-scale features (such as local corners) are spliced at the feature level, and the geometric parameters (such as corner angles) and shape parameters (such as curvature) are combined according to the dimension to form a structured feature set; high-scale features (such as overall contour) are fused by decision-level voting, and the most consistent global parameters (such as average perimeter, longest axis length) are selected according to the physical size correlation value of multi-view features.

[0092] Finally, standardize the adjustment: eliminate redundant features after fusion (such as parameters with a repetition rate of more than 80%), unify the dimension (such as physical size unit mm) and format (such as feature name-value-error range) of the feature, and check the correlation error (≤0.002mm) between the feature and the physical size calibration data to ensure accuracy. Finally, a standardized size detection feature set containing geometric, shape features and multi-view fusion information is generated, providing a unified data benchmark for subsequent size detection.

[0093] S5, constructing a size detection model based on a deep learning fusion algorithm, and inputting the standardized size detection feature set into the size detection model for iterative training to obtain a target size quantitative detection result; In the embodiment of the application, the size detection model is constructed based on the deep learning fusion algorithm, and the standardized size detection feature set is input into the size detection model for iterative training to obtain a target size quantitative detection result, which includes the following steps: S51, presetting a deep learning fusion algorithm architecture, wherein the deep learning fusion algorithm architecture includes a network layer configuration library and a training parameter database, and constructing a size detection model according to the network layer configuration library and the training parameter database; Specifically, when constructing the network layer configuration library, the feature processing logic is designed hierarchically: the feature extraction layer includes a multi-scale convolution module (such as 3x3 / 5x5 convolution kernel, suitable for different size features), and a residual connection (to alleviate gradient disappearance); the fusion layer includes an attention mechanism layer (such as channel attention, to highlight high-score feature weight), and a cross-scale splicing layer (to splice low / medium / high-scale feature maps); the output layer uses a fully connected layer + regression head (to output physical size parameters such as length and diameter), and labels the parameter range of each layer (such as convolution kernel number 32-256, according to feature dimension adjustment).

[0094] The training parameter database needs to cover key training variables: the optimizer is Adam (initial learning rate 0.001, stepwise decay), the batch size is 8-32 (adjust according to the sample size), the iteration number is 50-200 (according to the loss convergence), the loss function adopts the combination of MSE (physical size error) + Dice loss (contour matching degree), and the regularization parameters (such as L2 weight decay coefficient 0.0001) are stored.

[0095] When constructing the size detection model, select the adaptive module from the network layer configuration library: based on the dimension of the standardized size detection feature set (such as feature map size 256x256), configure 3-5 layers of convolution to extract features, access the attention fusion layer to integrate multi-scale information, and finally output the size prediction value through the output layer; call the training parameter database to initialize the model, jointly train with the fuzzy contour sample set and the data contour sample set, monitor the loss change in real time (target loss <0.001), and iterate the network weight and training parameter to form a stable size detection model.

[0096] S52, data enhancement processing is performed on the standardized size detection feature set, and the feature set is divided into a training set, a validation set and a test set; Specifically, during data enhancement processing, preferentially select methods that do not destroy the size accuracy: first, geometric transformation, perform small range translation, slight rotation (0°-10°, maintain shape integrity), and equal proportion scaling (0.9-1.1 times, maintain size proportion relationship) on the contour image in the feature set to ensure that the physical size label does not change after enhancement; second, gray scale enhancement, random gray scale jitter and slight Gaussian noise (standard deviation ≤0.05) are adopted to simulate the light fluctuation and sensor noise in actual detection, while avoiding hiding the edge features, and the original feature set needs to be synchronized during the enhancement process to prevent data distortion caused by excessive enhancement.

[0097] The data set division needs to follow the principles of random sampling and consistent distribution: according to the conventional ratio of 7:2:1, the enhanced feature set (including the original set) is randomly divided into a training set (70%, used for model parameter learning), a validation set (20%, used for hyperparameter tuning and model selection), and a test set (10%, used for final performance evaluation). The data order needs to be shuffled before division to ensure that the feature distribution (such as size range, feature type proportion) of the three types of data sets is consistent with the original set. If the sample size is small (<1000), 5-fold cross-validation is used instead of fixed division to improve data utilization. Finally, independent and representative three types of data sets are formed to support model training and evaluation.

[0098] S53, input the training set, the validation set and the test set into the size detection model for iterative training; S54, input the standardized size detection feature set into the size detection model after iterative training to obtain the target size quantitative detection result.

[0099] Specifically, first, input adaptation processing is performed: the format of the standardized feature set (such as feature map dimension, parameter arrangement order) is checked to be consistent with the input requirements during model training, if the feature set is image type data, the resolution needs to be adjusted to the model specified size (such as 256x256); if it is parameter type data (such as perimeter, curvature value), the normalization range (such as 0-1) needs to be confirmed to be the same as the pre-processing logic during training, to avoid inference errors caused by format deviation, and to filter out abnormal samples in the feature set (such as missing features, data beyond the range), to reduce the interference of invalid input on the results.

[0100] Then, model inference is performed: the optimal model weight saved after iteration training (usually the weight when the validation set loss is the lowest) is loaded, and the adapted feature set is batched, and the model automatically completes secondary feature extraction (through convolution layer), multi-scale fusion (through attention layer) and size regression calculation (through output layer), directly outputs the preliminary target size quantization prediction value (such as length 5.02mm, diameter 3.15mm), and synchronously outputs the prediction confidence (reflecting the reliability of the result).

[0101] Finally, the detection results are optimized: the accuracy of the preliminary prediction value is calibrated, combined with the error law of the physical size calibration data (such as the model predicting small size too small, corrected by a fixed coefficient of 1.02), the abnormal results with confidence <0.8 or prediction value beyond the reasonable range are removed, the dimension is unified (such as all sizes in mm), and the error range of each result is labeled (such as 5.02±0.003mm), and finally the accurate and traceable target size quantization detection results are obtained.

[0102] S6, the accuracy of the target size quantization detection result is verified, and an image feature size parameter correlation database is established according to the verified target size quantization detection result.

[0103] Specifically, first, select verification samples covering the full detection range (such as detection results of small, medium and large size standard blocks), calculate the absolute error (|detection value-calibration value|) and relative error (absolute error / calibration value x 100%) of each sample; then calculate the overall verification index, including the average absolute error, the root mean square error, if the average absolute error is ≤0.002mm, the RMSE is ≤0.003mm, and the single sample relative error is ≤1.5%, the result is accurate; for the out-of-tolerance results, trace back the model inference process (such as input adaptation, parameter calibration), eliminate the system error and re-verify.

[0104] When the image feature size parameter correlation database is established, a structured data framework needs to be constructed: the core fields include image feature information (normalized contour parameters, multi-scale fusion features, spatial alignment values), target size data (quantitative detection values, error ranges, confidence levels, verification results), physical calibration data (standard block true values, calibration accuracy levels), and metadata (detection angle, time, equipment number); first, abnormal data (such as unqualified detection results) is cleaned, then integrated according to the correlation of image features-size data-calibration data, stored using a relational database, ensuring that fields can be cross-table queried, and finally forming a correlation database supporting feature backtracking and size tracing, providing data support for subsequent model optimization and detection scene expansion.

[0105] According to another aspect of the present application, as shown in Figure 2 An image data management system for high-precision size visual detection is provided, which comprises a data acquisition segmentation module 1, an image correction extraction module 2, a feature evaluation screening module 3, a feature optimization generation module 4, a model construction training module 5, and a result verification database construction module 6. The data acquisition segmentation module 1 is used to acquire multi-angle image data sets and physical size calibration data sets through a high-precision size detection device, and to perform regional segmentation processing on the multi-angle image data sets. The image correction extraction module 2 is used to correct and enhance the multi-angle image data sets after regional segmentation processing, and to extract contour feature parameters of the multi-angle image data sets after correction and enhancement. The feature evaluation screening module 3 is used to preset a feature effectiveness evaluation index, to map and correlate the contour feature parameters with the physical size calibration data, and to screen and classify the contour feature parameters after mapping and correlation according to the feature effectiveness evaluation index. The feature optimization generation module 4 is used to preset a multi-scale feature fusion strategy, to optimize and adjust the contour feature parameters after screening and classification according to the multi-scale feature fusion strategy, and to generate a standardized size detection feature set. The model construction training module 5 is used to construct a size detection model based on a deep learning fusion algorithm, to input the standardized size detection feature set into the size detection model for iterative training, and to obtain a target size quantitative detection result. The result verification database construction module 6 is used to verify the accuracy of the target size quantitative detection result, and to establish an image feature size parameter correlation database according to the target size quantitative detection result after verification.

[0106] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An image data management method for high-precision dimensional visual inspection, characterized by, The method comprises the following steps: S1, collecting multi-view image data set and physical size calibration data set by high-precision size detection equipment, and performing region segmentation processing on the multi-view image data set; S2, correcting and enhancing the multi-view image data set after region segmentation processing, and extracting the contour feature parameters of the multi-view image data set after correction and enhancement; S3, presetting a feature validity evaluation index, mapping and associating the contour feature parameters with the physical size calibration data, and screening and classifying the contour feature parameters after mapping and association according to the feature validity evaluation index; S4, presetting a multi-scale feature fusion strategy, and optimizing and adjusting the contour feature parameters after screening and classification according to the multi-scale feature fusion strategy to generate a standardized size detection feature set; S5, constructing a size detection model based on a deep learning fusion algorithm, inputting the standardized size detection feature set into the size detection model for iterative training, and obtaining a target size quantitative detection result; S6, verifying the accuracy of the target size quantitative detection result, and establishing an image feature size parameter association database according to the verified target size quantitative detection result.

2. The image data management method for high-precision dimensional visual inspection according to claim 1, characterized in that, The method comprises the following steps: S11, setting a high-precision industrial camera and a light source system to form a detection equipment, adjusting the shooting angle to collect multi-view images of the target object, and forming a multi-view image data set; S12, using existing geometric standard instruments as calibration reference objects, synchronously collecting images and physical size parameters of the standard blocks under each view angle, and forming a physical size calibration data set; S13, using an adaptive threshold segmentation algorithm to perform region segmentation on the multi-view image data set to obtain a target region data set and a background region data set; S14, extracting the edge contour of the target region data set and the background region data set, and filling the internal cavities.

3. The image data management method for high-precision dimensional visual inspection according to claim 2, characterized in that, The method comprises the following steps: S31, presetting a feature validity evaluation index, combining the contour feature parameters with the physical size calibration data to obtain a data contour sample set; S32, constructing a contour mapping association model using a fuzzy logic reasoning algorithm, and iteratively optimizing the contour mapping association model through the data contour sample set; S33, inputting the contour feature parameters into the iteratively optimized contour mapping association model to generate a feature size mapping table, and quantitatively scoring each group of contour feature parameters in the feature size mapping table according to the feature validity evaluation index; S34, presetting a scoring screening rule, screening the contour feature parameters after quantitative scoring based on the scoring screening rule, and classifying them according to feature types.

4. The image data management method for high-precision dimensional visual inspection according to claim 1, characterized in that, The method comprises the following steps: S41, a multi-scale feature fusion strategy is preset, wherein the multi-scale feature fusion strategy includes a hierarchical fusion strategy library and an alignment method database, and a fusion scale level threshold and a feature space alignment threshold are divided; S42, the multi-scale contour feature parameters after screening and classification are normalized; S43, the fusion scale level value and the contour feature space alignment value of the multi-scale contour feature parameters after normalization are extracted, and a hierarchical fusion strategy and a space alignment method in the hierarchical fusion strategy library and the alignment method database are selected based on the fusion scale level value and the contour feature space alignment value; S44, the contour feature parameters are optimized and adjusted based on the hierarchical fusion strategy and the space alignment method to generate a standardized size detection feature set.

5. The image data management method for high-precision dimensional visual inspection according to claim 1, characterized in that, The depth learning fusion algorithm architecture is preset, wherein the depth learning fusion algorithm architecture includes a network layer configuration library and a training parameter database, and the size detection model is constructed according to the network layer configuration library and the training parameter database; S52, the standardized size detection feature set is subjected to data enhancement processing, and is divided into a training set, a verification set and a test set; S53, the training set, the verification set and the test set are input into the size detection model for iterative training; S54, the standardized size detection feature set is input into the size detection model after iterative training to obtain a target size quantitative detection result. The adaptive threshold segmentation algorithm is used to segment the multi-view image data set to obtain a target region data set and a background region data set, including the following steps:

6. The image data management method for high-precision dimensional visual inspection according to claim 2, characterized in that, S131, image noise in the multi-view image data set is removed; S132, an adaptive threshold calculation window set is preset, and the adaptive threshold calculation window set is matched with the denoised multi-view image data set; S133, based on the matched adaptive threshold calculation window, the gray distribution feature and the target size parameter in the multi-view image data set are extracted; S134, the gray distribution feature and the target size parameter are binarized and segmented to obtain a target region data set and a background region data set. The edge contour of the target region data set and the background region data set is extracted, and the internal cavity of the region is filled, including the following steps:

7. The image data management method for high-precision dimensional visual inspection according to claim 6, characterized in that, S141, an edge detection algorithm library and a morphological processing parameter library are preset, and an edge intensity threshold and a region internal cavity area threshold are set; S142, the gray gradient feature value of the target region data set and the background region data set is extracted, and an edge detection algorithm in the edge detection algorithm library is selected based on the gray gradient feature value; S143, the selected edge detection algorithm is applied to the target region data set and the background region data set, and an edge contour is generated by screening and adjusting based on the edge intensity threshold; S144, the internal cavity of the edge contour is identified, and a closing operation parameter in the morphological processing parameter library and the region internal cavity area threshold are called to fill the internal cavity of the edge contour. ​ 8. The image data management method for high-precision dimensional visual inspection according to claim 3, characterized in that, The method comprises the following steps of: S321, defining input variables and output variables of the contour mapping correlation model, and constructing fuzzy sets and membership functions based on the input variables and the output variables; S322, setting the contour mapping correlation model according to the fuzzy sets and the membership functions; S323, presetting fuzzy rules, and performing fuzzy processing on the data contour sample set based on the fuzzy rules to obtain a fuzzy contour sample set; S324, inputting the fuzzy contour sample set and the data contour sample set into the contour mapping correlation model to calculate fuzzy contour parameters and data contour parameters; S325, presetting a model iteration threshold, calculating contour error parameters of the fuzzy contour parameters and the data contour parameters, and performing threshold matching on the contour error parameters and the model iteration threshold; S326, iteratively optimizing the contour mapping correlation model based on the threshold matching result.

9. The image data management method for high-precision dimensional visual inspection according to claim 8, wherein, The method comprises the following steps of: S331, inputting the contour feature parameters into the iteratively optimized contour mapping correlation model to calculate physical size correlation values, and combining the contour feature parameter physical size correlation values to form a feature size mapping table; S332, setting feature scoring standards and weight allocation standards for each index of the feature effectiveness evaluation index; S333, extracting correlation data of each group of contour feature parameters in the feature size mapping table, and calculating quantitative scores of each evaluation index based on the feature scoring standards and the weight allocation standards.

10. An image data management system for high precision dimensional visual inspection for implementing the image data management method for high precision dimensional visual inspection according to any one of claims 1 to 9, characterized in that, The system comprises a data acquisition and segmentation module, an image correction and extraction module, a feature evaluation and screening module, a feature optimization and generation module, a model construction and training module, and a result verification and database construction module. The data acquisition and segmentation module is configured to acquire multi-view image data sets and physical size calibration data sets by using a high-precision size detection device, and perform region segmentation processing on the multi-view image data sets. The image correction and extraction module is configured to correct and enhance the multi-view image data sets after region segmentation processing, and extract contour feature parameters of the multi-view image data sets after correction and enhancement. The feature evaluation and screening module is configured to preset feature effectiveness evaluation indexes, map and correlate the contour feature parameters and the physical size calibration data, and screen and classify the mapped and correlated contour feature parameters according to the feature effectiveness evaluation indexes. The feature optimization and generation module is configured to preset a multi-scale feature fusion strategy, and optimize and adjust the screened and classified contour feature parameters according to the multi-scale feature fusion strategy to generate a standardized size detection feature set. The model construction and training module is configured to construct a size detection model based on a deep learning fusion algorithm, input the standardized size detection feature set into the size detection model for iterative training, and obtain a target size quantitative detection result. The result verification and database building module is used for verifying the accuracy of the target size quantification detection result, and building an image feature size parameter correlation database according to the verified target size quantification detection result.

Images