Mechanical casting detection method and system based on image data analysis and processing

Abstract

The invention relates to the technical field of image recognition, in particular to a mechanical casting detection method and system based on image data analysis processing, and the method comprises the following steps: obtaining a casting image gray value, classifying according to a gray scale standard, generating a multi-layer image structure and a layering graph, setting a grid, counting pixels to form a density sequence, extracting a trend region, and building an index table; and extracting multi-angle image boundary point matching intersection points to form a coordinate matrix, screening a stable intersection point set, searching an original image comparison trend coordinate to extract an overlapping region, and generating a defect annotation image result. According to the method, an image is subjected to gray scale division to form a multi-layer structure, texture differential expression is enhanced, a pattern layer grid statistics density sequence is used for extracting a trend region, abnormal distribution positioning is achieved, boundary mutation points in a multi-view image are uniformly mapped, boundary recognition consistency and robustness are improved, and a stable boundary and the trend region are compared and marked; defect expression intuition and credibility are enhanced, and detection precision and stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a mechanical casting detection method and system based on image data analysis and processing. Background Art

[0002] The field of image recognition technology encompasses a technical system that utilizes computer vision, image processing, and artificial intelligence methods to automatically identify and classify information in images. The core content of this technical field includes key technical links such as target detection, image segmentation, feature extraction, image classification, and image semantic understanding. Image recognition technology typically relies on numerical processing of low-level features such as grayscale, texture, and shape in image data, and combines classification and discrimination models to identify and analyze targets in the image. This technology is widely used in multiple industries, including medical image analysis, security monitoring, industrial quality inspection, autonomous driving, and remote sensing image interpretation. It possesses highly automated and large-scale processing capabilities, enabling a complete technical process from image acquisition to information recognition.

[0003] Among them, the mechanical casting inspection method based on image data analysis and processing refers to using an image acquisition device to obtain the surface image of the mechanical casting, and by identifying and processing the edge features, geometric shape and surface texture of the target area in the image, it is possible to determine whether the mechanical casting has defects. The subject of this patent mainly focuses on technical matters such as data acquisition, image preprocessing and feature extraction in the detection of surface defects of castings. Through image analysis methods such as image enhancement and denoising, edge contour extraction, and binarization processing, effective structural information of the target area is obtained. Then, combined with the recognition method based on feature vector matching, the defect type is classified and identified, thereby completing the detection of casting defects.

[0004] Grayscale, texture, and edge recognition based on low-level image features is highly dependent on a single image, making it difficult to avoid bias caused by factors such as surface lighting, shooting angle, and localized stains. Without the support of a multi-layered image representation structure, grayscale information appears uniformly distributed within the image, resulting in the appearance of features similar to those of normal areas in some edge transition regions, leading to increased false positives. Due to the lack of statistical analysis of regional density trends, existing technologies are unable to effectively capture small but continuously changing texture anomalies, easily missing early defect signals. Furthermore, methods that extract edges from a single perspective and then directly classify them struggle to accurately assess boundary stability, making boundary recognition susceptible to local noise and increasing the risk of false and missed detections. The lack of multi-image analysis methods based on spatial coordinate cross-verification and stable pixel selection results in unstable recognition results across multiple angles, making them difficult to adapt to image variations in production scenarios. For example, in industrial inspection, the edge appearance of the same casting varies significantly when viewed from different angles. Conventional methods are prone to boundary jitter and misaligned recognition, compromising the accuracy and completeness of defect annotation. Existing technologies fail to establish a coordinate mapping and verification mechanism between images, which limits the generalization ability of recognition effects in multi-angle and multi-scale scenarios. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a mechanical casting detection method and system based on image data analysis and processing.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a mechanical casting detection method based on image data analysis and processing, comprising the following steps:

[0007] S1: Acquire image data of the casting surface area, extract pixel grayscale values, and classify pixels according to grayscale according to the preset grayscale interval standard. The division results are organized into a multi-layer image structure in layer order to generate a grayscale layered atlas dataset;

[0008] S2: calling the grayscale layered atlas dataset, setting equal-area grid division in the layer, counting the number of pixels in the grid in each layer of the image to form a multi-layer density sequence, extracting continuous trend areas based on the change relationship in the density sequence, and generating a trend direction area index table;

[0009] S3: extracting images corresponding to the indicated area under multiple shooting angles according to the trend area index table, extracting pixels with sudden gradient changes at the boundary of the area in each image to form a boundary point sequence, matching the boundary points under multiple viewing angles according to position correspondence, and generating a boundary coincidence intersection coordinate matrix;

[0010] S4: Call the boundary coincidence intersection coordinate matrix, perform statistics on the spatial coordinate changes of the intersection under multiple images, perform stability judgment on the horizontal and vertical offsets, filter the intersection set with a change amplitude lower than the set standard, and generate a stable boundary pixel point set.

[0011] As a further solution of the present invention, the grayscale layered atlas data set includes a multi-layer image structure, grayscale interval standards, and pixel grayscale values; the trend direction area index table includes a multi-layer density sequence at the grid position, changes in the number of pixels in the area, and continuous trend areas; the boundary coincidence intersection coordinate matrix includes a boundary point sequence, matching results under multiple perspectives, and intersection mapping in a unified coordinate system; the stable boundary pixel point set includes a stable intersection set, an image area with coincident coordinates, and an intersection with a change amplitude lower than a set standard.

[0012] As a further solution of the present invention, the specific steps of S1 are:

[0013] S101: Acquire image data of the casting surface area, call the pixel array of the image grayscale channel, read the grayscale value corresponding to each pixel, establish a mapping relationship between the pixel position and the grayscale value, and generate the image pixel grayscale matrix value;

[0014] S102: Based on the grayscale matrix values ​​of the image pixels, according to a preset grayscale interval standard, determining the grayscale attribution of the pixels, classifying them into corresponding grayscale labels, sorting the pixel positions corresponding to the grayscales, and obtaining grayscale label division coordinate groups;

[0015] S103: Divide the coordinate groups according to the grayscale labels, mark the grayscale values ​​of corresponding positions in the image in sequence, organize the layer structure according to the grayscale label sequence, integrate the image data to form a multi-layer image structure, and obtain a grayscale layered atlas dataset.

[0016] As a further solution of the present invention, the specific steps of S2 are:

[0017] S201: calling the grayscale layered atlas dataset, dividing each layer of the image into equal-area grids, identifying the spatial distribution positions of pixels within the grids, establishing a correspondence between the grid positions and pixel distributions in the layer, and generating grid pixel distribution values;

[0018] S202: Based on the grid pixel distribution values, construct a sequence of the number of pixels at the same grid position in the layer, determine the continuous direction characteristics of the value change in the sequence, extract the layer segment with a single direction change, and generate a density sequence trend segment value;

[0019] S203: According to the density sequence trend segment value, the change direction identifiers of adjacent grid positions are compared, the region combinations with the same continuous direction are classified, the trend extension range in the layer is delineated, and a trend direction region index table is generated.

[0020] As a further solution of the present invention, the specific calculation formula for the continuous direction feature of the value change in the judgment sequence is:

[0021]

[0022] Calculate the directional consistency eigenvalue and generate the density sequence trend segment value;

[0023] Among them, H represents the eigenvalue of directional consistency, x f represents the number of pixels in the grid position f, s represents the total length of the sequence, |x f+1 -x f | represents the absolute difference in the number of adjacent grid pixels, (x f+1 -x f ) represents the direction of change in the number of adjacent grid pixels, represents the sum of squares of the sequence, and ∈ represents the minimum value to prevent zero constant.

[0024] As a further solution of the present invention, the specific steps of S3 are:

[0025] S301: Obtain a multi-angle image corresponding to a number in the trend area index table, locate the corresponding area in the image, extract pixel grayscale change characteristics within the area, mark boundary pixels according to grayscale value mutations, sequentially construct a boundary point set, and generate a region boundary point sequence value;

[0026] S302: Calling the boundary points under the angle in the region boundary point sequence value, extracting the coordinate corresponding points in the same region according to the position relationship in the image, constructing a matching pair set between view angles, and generating coordinate pair values ​​of the boundary corresponding points between view angles;

[0027] S303: Convert the coordinate pairs of the corresponding points on the boundaries between the perspectives and the image shooting parameters into the same coordinate system, calculate the intersection positions under multiple perspectives, and integrate to generate a boundary coincidence intersection coordinate matrix.

[0028] As a further solution of the present invention, the specific calculation formula for constructing the matching pair set between perspectives is:

[0029]

[0030] Calculate the perspective matching index and generate the coordinate pairs of corresponding points on the boundary between perspectives;

[0031] Among them, M ik Represents the matching weight value between the i-th perspective and the k-th perspective, Δx ij Represents the absolute difference between the jth boundary point at the i-th perspective and the reference x-coordinate, Δy ijRepresents the absolute difference between the jth boundary point of the i-th perspective and the reference y coordinate, θ jk Represents the projection inclination angle radian value of the j-th boundary point at the k-th viewing angle, Δφ ik Represents the change in the optical axis angle between the i-th viewing angle and the k-th viewing angle, Δψ jk Represents the curvature change of the j-th boundary point at the k-th perspective, Represents the three-dimensional unit direction vector of the i-th perspective, R represents the preset field of view overlap radius threshold, and n represents the number of valid boundary points in the current area.

[0032] As a further solution of the present invention, the specific steps of S4 are:

[0033] S401: Calling the pixel positions of the intersection points in the image sequence in the boundary coincidence intersection coordinate matrix, extracting the coordinate information in the image order, establishing a sequence set of image positions corresponding to the intersection points, integrating the image position information of the intersection points, and generating an intersection pixel displacement dataset;

[0034] S402: extracting the horizontal and vertical displacement amplitudes of the intersection points in the image sequence based on the coordinate changes in the intersection pixel displacement dataset, setting an upper limit for the displacement as a judgment criterion, screening a set of intersection points that meet the conditions, and generating a coordinate stability screening matrix;

[0035] S403: Calling the intersection point set in the coordinate stability screening matrix, integrating the horizontal and vertical position information under the image, generating an average position point set of the intersection points, constructing the boundary distribution according to the original index, and generating a stable boundary pixel point set.

[0036] As a further embodiment of the present invention, the method further comprises:

[0037] S5: Based on the stable boundary pixel point set, the corresponding position is searched in the original image, and the position coordinates are compared with the position coordinates in the trend direction area index table, and the image area where the coordinates of the two coincide with each other is extracted, and the regional image is annotated to generate a defect annotation visualization image result;

[0038] The defect annotation visualization image result includes the detection image after annotation processing, the original image fusion output result, and the visualization defect information;

[0039] The specific steps of S5 are:

[0040] S501: Based on the stable boundary pixel point set, position coordinate information in the original image is extracted, and the coordinates are compared point by point with the coordinates in the trend region index table, image positions with consistent coordinate values ​​are screened, and a coordinate coincidence region index value set is generated;

[0041] S502: Calling the coordinate coincident region index value set, extracting the image content of the corresponding region from the original image, screening the positions that meet the boundary clarity condition based on the edge feature distribution of the image channel, and generating a pixel position set of the labelable region;

[0042] S503: calling the pixel position set of the annotable area, completing pixel value replacement processing at the corresponding position in the original image, adding a unified boundary mark and performing image fusion to generate a defect annotation visualization image result.

[0043] The mechanical casting inspection system based on image data analysis and processing includes:

[0044] The image layering module acquires image data of the surface area of ​​mechanical castings, collects the grayscale values ​​of pixels in the image, makes attribution judgments based on preset grayscale interval standards, divides pixels into corresponding layers, organizes the layers according to the grayscale level order, and establishes a grayscale layered atlas dataset;

[0045] The density extraction module calls the grayscale layered atlas dataset, sets a grid of equal proportions within the image area, counts the number of pixels contained in each grid in the layer, arranges the pixel distribution sequence of the grid under the difference layer, extracts the trend area index based on the continuous change of the pixel distribution, and generates a trend direction area index table;

[0046] The trend construction module calls the trend direction area index table, extracts the image content of the corresponding area under multiple shooting angles, obtains the grayscale gradient mutation points within the boundary range, forms a boundary point sequence, matches the boundary points based on the corresponding relationship between the image perspective and position, maps the intersection position in a unified coordinate system, and establishes a boundary coincidence intersection coordinate matrix;

[0047] The boundary mapping module calls the coordinate matrix of the boundary coincidence intersection points, counts the spatial position changes of the intersection points under the perspective of the difference image, determines whether the horizontal and vertical offset amplitudes are lower than the stability benchmark, selects the intersection point set with a smaller position change range, and generates a stable boundary pixel point set;

[0048] The defect annotation module calls the stable boundary pixel point set, searches for the corresponding coordinates in the original image area, compares them with the positions in the trend area index table, extracts the coordinate overlapping area for image annotation processing, outputs the fused image, and generates a defect annotation visualization image result.

[0049] Compared with the prior art, the advantages and positive effects of the present invention are:

[0050] In the present invention, a multi-layer structure is formed by dividing the image by grayscale attribution to enhance the hierarchical expression of texture differences, and the trend direction area is extracted through the layer grid statistical density sequence to achieve accurate positioning of abnormal distribution. The boundary gradient mutation points in the multi-view image are uniformly mapped to generate an intersection coordinate matrix, and stable intersections are screened according to spatial offset to improve the consistency and robustness of boundary recognition. Finally, the stable boundary and trend area positions are compared and marked to improve the intuitiveness and credibility of the defect visualization expression. This process runs through grayscale stratification, trend extraction, spatial matching and precise labeling to enhance detection accuracy, stability and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the steps of the present invention;

[0052] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0054] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0055] See also Figure 1 The mechanical casting detection method based on image data analysis and processing includes the following steps:

[0056] S1: Acquire image data of the casting surface area, extract pixel grayscale values, and classify pixels according to grayscale according to the preset grayscale interval standard. The divided results are organized into a multi-layer image structure in layer order to generate a grayscale layered atlas dataset;

[0057] S2: Call the grayscale layered atlas dataset, set the equal-area grid division in the layer, count the number of pixels in the grid in each layer of the image, form a multi-layer density sequence at the grid position, extract the continuous trend area based on the change relationship in the density sequence, and generate the trend direction area index table;

[0058] S3: Call the trend area index table, extract the corresponding images of the indicated area under multiple shooting angles, extract the boundary gradient mutation pixels in the area from each image to form a boundary point sequence, match the boundary points under multiple perspectives based on the position correspondence, complete the intersection mapping in the unified coordinate system, and generate the boundary coincidence intersection coordinate matrix;

[0059] S4: Call the boundary coincidence intersection coordinate matrix, perform statistics on the spatial coordinate changes of the intersections under multiple images, perform stability judgment on the horizontal and vertical offsets, filter the intersection set with a change amplitude lower than the set standard, and generate a stable boundary pixel point set;

[0060] S5: Call the stable boundary pixel point set, find the corresponding position in the original image, compare it with the position coordinates in the trend area index table, extract the image area where the two coordinates coincide, perform annotation processing on the area image, and fuse it with the original image to output the detection image, generating a defect annotation visualization image result.

[0061] The grayscale layered atlas dataset includes multi-layer image structure, grayscale interval standards, and pixel grayscale values. The trend area index table includes multi-layer density sequences at grid positions, changes in the number of pixels in the area, and continuous trend areas. The boundary coincidence intersection coordinate matrix includes boundary point sequences, matching results under multiple perspectives, and intersection mapping in a unified coordinate system. The stable boundary pixel point set includes stable intersection sets, image areas with coincident coordinates, and intersections with a change amplitude lower than the set standard. The defect annotation visualization image results include the detection image after annotation processing, the original image fusion output results, and visualized defect information.

[0062] See also Figure 1 , the specific steps of S1 are:

[0063] S101: Acquire image data of the casting surface area, call the pixel array of the image grayscale channel, read the grayscale value corresponding to each pixel, establish a mapping relationship between the pixel position and the grayscale value, and generate the image pixel grayscale matrix value;

[0064] When acquiring image data of the casting surface area, an industrial camera with grayscale acquisition function is first used to capture the data. Commonly used devices are CCD type, with the camera resolution set to 1280×1024. The image is stored in grayscale 8-bit format, and the pixel value range is 0 to 255. After image acquisition is completed, the image is imported into the image processing system and converted into a grayscale pixel matrix, where each pixel corresponds to a unique coordinate position and grayscale value. The reading process uses a row-by-row scanning method to record the row and column coordinates of each pixel point and the corresponding grayscale value. For example, the grayscale of the pixel point at the coordinate position of 120 rows and 240 columns is 187, and the grayscale value of the position at the coordinate position of 500 rows and 300 columns is 72. The system extracts all pixel grayscale values ​​in sequence and combines them into a complete two-dimensional grayscale matrix. The matrix size is consistent with the image resolution, forming a pixel grayscale mapping structure for the casting surface area. This matrix structure provides the basis for subsequent grayscale interval determination and layer decomposition.

[0065] S102: Based on the grayscale matrix values ​​of the image pixels and according to a preset grayscale interval standard, the grayscale attribution of the pixels is determined, and the grayscale labels are classified. The pixel positions corresponding to the grayscales are sorted, and the grayscale label division coordinate groups are obtained.

[0066] Grayscale interval division criteria are set based on the grayscale matrix structure. For example, five grayscale level labels can be used, covering the full grayscale range from 0 to 255. The specific intervals are 0 to 50 for the first level, 51 to 100 for the second level, 101 to 150 for the third level, 151 to 200 for the fourth level, and 201 to 255 for the fifth level. All positions in the pixel matrix are traversed, each grayscale value is obtained, and its corresponding grayscale affiliation is determined. For example, a pixel with a grayscale value of 187 is classified as level 4, and a pixel with a grayscale value of 45 is classified as level 1. The system groups all pixel coordinates under the same grayscale label into a coordinate group. For example, all pixels with a grayscale value between 101 and 150 are classified as level 3. The corresponding positions form a coordinate set for subsequent layer generation. The entire grayscale determination process compares grayscale value intervals sequentially based on classification conditions, achieving pixel classification and mapping operations, ultimately resulting in a coordinate group data structure under the five grayscale labels.

[0067] S103: Divide the coordinates into groups according to the grayscale labels, mark the grayscale values ​​of the corresponding positions in the image in sequence, organize the layer structure according to the grayscale label sequence, integrate the image data to form a multi-layer image structure, and obtain a grayscale layered atlas dataset;

[0068] Based on the assigned grayscale labels and their corresponding coordinate sets, multiple layers are constructed, each representing a grayscale level. In the initial step, an empty layer is created that matches the original image size. Each layer is populated with the grayscale values ​​from the original image only at pixel locations within the layer's grayscale range, while all other locations remain at a zero background. For example, in the third-level layer, only pixels with grayscale values ​​between 101 and 150 have grayscale values; all others are zero. In this way, five layers are constructed, each representing one of the five grayscale intervals. Once the layers are constructed, they are organized sequentially according to grayscale level and merged into a hierarchical image dataset. This structure is used for subsequent image analysis tasks such as regional analysis, distribution recognition, or anomaly identification. Each layer can be loaded and reviewed layer by layer through an image viewer. The layer dataset can be exported as a multi-page image format for further processing and archiving.

[0069] See also Figure 1 , the specific steps of S2 are:

[0070] S201: Calling the grayscale layered atlas dataset, dividing each layer of the image into equal-area grids, identifying the spatial distribution positions of pixels within the grids, establishing a correspondence between the grid positions and pixel distribution in the layer, and generating grid pixel distribution values;

[0071] First, each layer of the image needs to be converted into a two-dimensional grayscale matrix representation, where each matrix element represents the grayscale value of a pixel. The layer is divided into multiple equal-area grids. For example, an image of size 1024×1024 can be divided into a 32×32 grid structure, with each grid containing 1024 pixels. The spatial distribution of the pixels in each grid is then identified. This can be done by recording the relative position of the pixels to the grid center and clustering them based on their grayscale values. For example, they can be divided into three categories to form a preliminary distribution feature model. The number of pixels in each category and the degree of concentration of their spatial distribution are then counted to measure the pixel distribution value of the grid. The pixel distribution value can be obtained by using methods such as average grayscale, grayscale standard deviation, or information entropy. For example, if the grayscale of pixels in a grid is concentrated between 120 and 140, the average grayscale value is 130 and the standard deviation is less than 10, indicating that the grayscale distribution in this area is relatively concentrated. Based on this value, a mapping table is established for all grids to achieve the correspondence between grid positions in the layer and pixel spatial distribution. This process is suitable for scenarios such as remote sensing images and geological survey maps.

[0072] S202: Based on the grid pixel distribution values, construct a sequence of the number of pixels at the same grid position in the layer, determine the continuous direction characteristics of the value change in the sequence, extract the layer segments with a single direction change, and generate a density sequence trend segment value;

[0073] The specific calculation formula for determining the continuous direction characteristics of the value change in the sequence is:

[0074]

[0075] Calculate the directional consistency eigenvalue and generate the density sequence trend segment value;

[0076] Among them, H represents the eigenvalue of directional consistency, x f represents the number of pixels in the grid position f, s represents the total length of the sequence, |x f+1 -x f | represents the absolute difference in the number of adjacent grid pixels, (x f+1 -x f ) represents the direction of change in the number of adjacent grid pixels, represents the sum of squares of the sequence, ∈ represents the minimum value to prevent zero constant;

[0077] Data collection and parameter setting:

[0078] Monitor the grid area of ​​a certain image layer and collect the pixel number sequence values ​​at grid positions f = 1 to f = 5:

[0079] x f =[12,15,18,14,20];

[0080] The total sequence length s=5 is directly obtained through grid division.

[0081] Adjacent difference calculation:

[0082] The absolute difference in the number of adjacent grid pixels:

[0083] |x f+1 -x f |=[3,3,4,6];

[0084] The direction of change of the number of adjacent grid pixels (sign):

[0085] (x f+1 -x f )=[3,3,-4,6];

[0086] Sequential sum of squares calculation:

[0087]

[0088] Zero prevention constant setting:

[0089] ∈=0.001 is set according to the IEEE floating-point arithmetic standard to avoid the denominator being zero.

[0090] Molecular computing:

[0091]

[0092] Denominator calculation:

[0093]

[0094] Final result:

[0095]

[0096] Result analysis:

[0097] The directional consistency eigenvalue H≈1.058 is greater than zero, indicating that the trend of pixel number change is dominated by positive growth. When generating density sequence trend segment values, it is necessary to prioritize the extraction of continuously increasing areas.

[0098] S203: Based on the density sequence trend segment values, the change direction identifiers of adjacent grid positions are compared, and the region combinations with the same continuous direction are classified to define the trend extension range in the layer, and generate a trend direction region index table;

[0099] First, the trend direction of each grid needs to be quantified. The angular difference between the distribution values ​​of adjacent positions can be used to calculate the direction identity to determine whether the trend directions of two adjacent grids are consistent. A threshold for determining direction consistency is set, for example, a direction change angle of less than 15 degrees is considered consistent. Adjacent grids with consistent directions are then combined to form continuous regions. These regions are classified into the same trend group. Subsequently, they are numbered and the position index of each region is recorded, including the coordinates of the upper left and lower right corners, which are used to mark the extension range of the region in the layer. Finally, a trend direction region index table is generated. The index table records basic information such as the number, range boundary, average direction, and area of ​​each trend region. For example, a region with trend number T01 extends from coordinates (0,0) to (128,256), with a direction of 10 degrees north-east, and covers 128 grids. This information can be used in layer analysis to identify areas with continuous structural extension, such as identifying rock formation trend sections with consistent trend in rock formation images.

[0100] See also Figure 1 , the specific steps of S3 are:

[0101] S301: Obtain a multi-angle image corresponding to the number in the trend area index table, locate the corresponding area in the image, extract the grayscale change characteristics of the pixels in the area, mark the boundary pixels according to the grayscale value mutation, construct a boundary point set in sequence, and generate a regional boundary point sequence value;

[0102] To obtain the image corresponding to a number, the first step is to extract the number from the region index table and parse the multi-view image information associated with that number. For example, the number Z023 corresponds to three images, named A_1, A_2, and A_3, with the corresponding viewpoints recorded using camera identifiers. After loading the image data, the target region within the image is located based on the pixel range defined in the index table. For example, if the rectangular boundary points are 124,210 in the upper left corner and 242,305 in the lower right corner, this range can be cropped to generate the region image. After grayscale conversion, all pixels within the region are traversed, and the grayscale variation between adjacent pixels is calculated. A preset threshold is used to determine whether a pixel constitutes an edge point. Pixels with a grayscale gradient exceeding 20 are considered candidate boundary points. The edge tracking algorithm is then used to sequentially extract the coordinates of boundary pixels from left to right. These boundary points are organized sequentially to form a boundary point set. For example, 85 boundary pixels are extracted from the A_1 image for region Z023. Their two-dimensional coordinates are recorded and arranged sequentially to form a complete boundary point sequence.

[0103] S302: Calling the boundary points under the angle in the region boundary point sequence value, extracting the coordinate corresponding points in the same region according to the position relationship in the image, constructing a matching pair set between viewpoints, and generating the coordinate pair values ​​of the boundary corresponding points between viewpoints;

[0104] The specific calculation formula for constructing a set of matching pairs between perspectives is:

[0105]

[0106] Calculate the perspective matching index and generate the coordinate pairs of corresponding points on the boundary between perspectives;

[0107] Among them, M ik Represents the matching weight value between the i-th perspective and the k-th perspective, Δx ij Represents the absolute difference between the jth boundary point at the i-th perspective and the reference x-coordinate, Δy ij Represents the absolute difference between the jth boundary point of the i-th perspective and the reference y coordinate, θ jk Represents the projection inclination angle radian value of the j-th boundary point at the k-th viewing angle, Δφ ik Represents the change in the optical axis angle between the i-th viewing angle and the k-th viewing angle, Δψ jk Represents the curvature change of the j-th boundary point at the k-th perspective, Represents the three-dimensional unit direction vector of the i-th perspective, R represents the preset field of view overlap radius threshold, and n represents the number of valid boundary points in the current area;

[0108] Δx ijObtained by measuring the image coordinate system, the x coordinate of the jth boundary point at the i-th viewpoint is subtracted from the x coordinate of the reference coordinate system. The reference coordinate system is determined by the coordinates of the center point of the calibration plate (512,384), and the measured Δx 12 =3.2 pixels, Δx 23 =5.7 pixels.

[0109] Δy ij Obtained in the same way, the measured Δy 12 =2.1 pixels, Δy 23 =4.3 pixels.

[0110] θ jk Calculated by the multi-view geometric projection model, the θ of the second boundary point at the third perspective is 23 =0.785 radians (45 degrees), with an error range of ±0.05 radians.

[0111] Δφ ik Obtained through binocular vision system calibration, Δφ of the first and third viewing angles 13 =0.349 radians (20 degrees), which meets the lens field angle constraint.

[0112] Δψ jk Through point cloud curvature analysis, the Δψ of the second boundary point at the third perspective is calculated. 23 =0.062 radians. The curvature is calculated using the mean angle method of the normal vectors of 5 adjacent point clouds.

[0113] 3D unit direction vector v i The first angle of view v1 = (0.707, 0.707, 0) and the third angle of view v3 = (0.866, 0.5, 0) are calculated by the inertial measurement unit and normalized.

[0114] The R value is preset to 120 mm based on the device parameters, which corresponds to 1.2 times the diagonal length of the CMOS sensor.

[0115] n=5 is determined by the number of valid boundary points output by the image segmentation algorithm.

[0116] Example calculation:

[0117] When j = 2, the numerator is calculated as:

[0118] (|3.2×sin0.785|+|2.1×cos0.785|)=(3.2×0.707+2.1×0.707)=3.763;

[0119] The denominator is calculated as:

[0120]

[0121] The result of the first fraction is:

[0122] 3.763 / 0.501≈7.511;

[0123] Vector cross product:

[0124]

[0125] The result of the latter fraction is:

[0126] 0.707 / 120≈0.00589;

[0127] The single loop items are:

[0128] 7.511×0.00589≈0.0443;

[0129] After traversing n=5 boundary points and accumulating, the measured values ​​of j=1 to j=5 are 0.0382, 0.0443, 0.0517, 0.0401, and 0.0489, respectively. The total value M_13=0.0382+0.0443+0.0517+0.0401+0.0489=0.223;

[0130] The result shows that the matching weight value of the first and third perspectives is 0.223. When Mik exceeds the preset threshold of 0.5, a valid matching pair is generated. The current result needs to continue to optimize the matching pair set.

[0131] S303: Converting the coordinate pairs of the corresponding points on the boundaries between the perspectives and the image capture parameters into a unified coordinate system, calculating the intersection positions under multiple perspectives, and integrating them to generate a boundary coincidence intersection coordinate matrix;

[0132] Based on the image coordinate values ​​and shooting parameter information of the matching point pairs, a coordinate system conversion operation is required. First, the camera intrinsic parameters such as focal length, principal point position, and pixel density are read for each image. At the same time, the spatial position and orientation information of the camera are obtained to calculate their relative relationship. The two-dimensional points in the image are back-projected into three-dimensional rays. The intersection point is then calculated with the back-projected rays of the matching points from another perspective to form the three-dimensional coordinate values. The intersection position is determined by the closest distance principle. If the two rays are not coplanar, the midpoint of the closest point between them is used as the estimated position. This intersection calculation is performed for each set of matching points. For example, after calculating the 65 boundary point pairs in area Z023, 65 three-dimensional spatial coordinate points are obtained. Each coordinate point is described by three sets of real values, forming a matrix set used to describe the spatial boundary structure.

[0133] See also Figure 1 , the specific steps of S4 are:

[0134] S401: Calling the pixel positions of the intersection points in the image sequence in the boundary coincidence intersection coordinate matrix, extracting the coordinate information in the image order, establishing a sequence set of intersection point corresponding image positions, integrating the image position information of the intersection points, and generating an intersection point pixel displacement dataset;

[0135] First, the image sequence and the intersection coordinate matrix must be read. The coordinate matrix records the pixel positions of each intersection in different image frames. For example, the first intersection has two-dimensional coordinates 100, 200 to 110, 210 in frames 1 through 10, respectively. By sequentially parsing the image frames, the position of each intersection in the image is extracted frame by frame and recorded as a pair of horizontal and vertical coordinates. These coordinates are then organized sequentially by frame number to form a coordinate sequence for each intersection. For example, if the image is 1920 × 1080 pixels, each frame may contain multiple intersections, each corresponding to a two-dimensional coordinate in each frame. All intersections must be traversed and data extracted in frame order, and their corresponding image position data is constructed into a set of coordinate sequences. The coordinate information of all intersections in the image sequence is then aggregated into a dataset with a sequence dimension. The pixel trajectory information of all intersections in the image frame sequence is recorded, and each intersection is assigned an image sequence identification number, forming the basic data set for image displacement analysis.

[0136] S402: Extracting the horizontal and vertical displacement amplitudes of the intersection points in the image sequence based on the coordinate changes in the intersection pixel displacement dataset, setting an upper limit for the displacement as a judgment criterion, screening a set of intersection points that meet the conditions, and generating a coordinate stability screening matrix;

[0137] Based on the pixel data from the image frame sequence obtained in the previous step, the magnitude of the change in the horizontal and vertical coordinates of each intersection over the number of frames is calculated. The pixel change magnitude between consecutive frames can be obtained by inter-frame coordinate difference. For each intersection, the horizontal displacement magnitude between adjacent frames is first calculated. For example, the horizontal coordinates of the first and second frames are subtracted, and then the interpolation calculation is performed frame by frame. The same is true for the vertical coordinates. The absolute value of all inter-frame differences is then taken to obtain the average displacement magnitude of each intersection in the image sequence. Filtering is performed based on a set pixel displacement judgment standard. For example, if the average horizontal and vertical displacements do not exceed 3 pixels, the intersection is considered stable in the image sequence. For example, if the horizontal coordinate of an intersection changes from 100 to 99 across five frames, the horizontal difference sequence is 1, 1, 2, and 1. The average of these values ​​is 1.25 pixels, which is below the upper limit of 3 pixels and meets the stability requirement. By determining whether the displacement of each intersection exceeds the offset limit, a Boolean screening result is generated. Each intersection is recorded to determine whether it belongs to the stable set, and a stability judgment matrix is ​​generated to prepare input for subsequent boundary reconstruction.

[0138] S403: Calling the intersection point set in the coordinate stability screening matrix, integrating the horizontal and vertical position information under the image, generating an average position point set of the intersection points, constructing the boundary distribution according to the original index, and generating a stable boundary pixel point set;

[0139] Extract the numbers of all intersections marked as stable from the stability judgment matrix, and then calculate the mean of all coordinate values ​​of these intersections in the image sequence to obtain the average of their horizontal and vertical coordinates. The average value is calculated by dividing the sum by the number of frames. For example, if the horizontal coordinates of an intersection in 10 frames are 100 to 109, the average horizontal coordinate is 104.5. The vertical coordinates are processed similarly. After sorting, the average position point set of stable intersections is obtained. Next, it is necessary to reorder these stable intersections in space according to the numbering order in the original data and establish the connection relationship of the boundary contour. For example, the stable points numbered 2, 4, and 6 are connected in the original order to form a boundary segment. Finally, the average position of all stable points and the reconstructed boundary relationship are combined to generate a set of boundary pixel points consisting of stable intersections for further image processing or boundary reconstruction.

[0140] See also Figure 1 , the specific steps of S5 are:

[0141] S501: Based on the stable boundary pixel point set, position coordinate information in the original image is extracted, and the coordinates are compared point by point with the coordinates in the trend area index table, and image positions with consistent coordinate values ​​are selected to generate a coordinate coincidence area index value set;

[0142] First, edge points are extracted from the original image. Conventional edge detection methods, such as gradient difference or local grayscale change, are used to determine edge attributes. Points with similar locations and consistent gradient changes in multiple consecutive images are classified into a stable boundary set. Then, based on this set, the horizontal and vertical coordinates of each pixel are extracted and compared point by point with the coordinate values ​​of each region in the trend region index table. The trend region index table divides the entire image into several logical regions, each containing a fixed number of pixel coordinates. For example, a 512×512 image can be divided into 40 equal regions, each containing approximately 200 pixels. The matching process achieves fast comparison by establishing an index structure. For example, spatial partitioning is used to accelerate the identification of whether a boundary point belongs to a certain region. When the coordinates of a stable pixel point are detected to be exactly the same as those in a region, the index number of that region is recorded in the set. After comparing all stable boundary points, a set of region index numbers is obtained, pointing to all image regions with overlapping boundary coordinates.

[0143] S502: Calling the coordinate coincident region index value set, extracting the image content of the corresponding region from the original image, screening the positions that meet the boundary clarity conditions based on the edge feature distribution of the image channel, and generating a pixel position set of the labeled region;

[0144] According to the set of region index numbers, the image content of all corresponding regions is extracted from the original image. These regions can be subdivided into image blocks. Then, the pixel content in the image blocks is decomposed into channels, such as splitting into three basic channels of red, green and blue. The edge change characteristics are analyzed on each channel. Specifically, the edge clarity can be judged by observing the continuity and mutation degree of the grayscale change within the channel. The average amplitude of the change value is counted and compared with the set reference value. For example, the benchmark is set to 30. When the average amplitude of the change of any channel in an image block exceeds this value, it is determined to be a valid area with a clear boundary. All pixel positions of such image blocks are recorded and added to the set of annotable areas. Assume that among the 8 image blocks extracted from the previous step, 5 meet the conditions. Each block is 64×64 pixels, and a total of approximately 20,480 pixels are included in the annotation preparation queue. These pixel position sets constitute the basic index for subsequent image modification and fusion.

[0145] S503: Calling the pixel position set of the annotable area, completing pixel value replacement processing at the corresponding position in the original image, adding a unified boundary mark and performing image fusion to generate a defect annotation visualization image result;

[0146] After obtaining the pixel coordinate set of the marked area, the corresponding pixel position in the original image is located one by one, and a color replacement operation is performed at that position. The original pixel value is replaced by a unified boundary color, such as red is used for identification. Each target pixel is directly replaced with the set color value, and the original image is retained for subsequent image fusion processing. The fusion process performs weighted superposition of the two image contents according to the set ratio. For example, the weight is set to 70% to retain the original image content, and the remaining 30% is superimposed with the identification image content. The weight distribution of different areas can be automatically adjusted according to the boundary stability. Areas with clearer boundaries correspond to a higher proportion of identification image superposition intensity to enhance the visual performance. Finally, a fused image is generated, which contains the content of all areas where pixel replacement is completed and superimposed with identification, forming an intuitive and visible presentation of the defect location.

[0147] See also Figure 2 , a mechanical casting inspection system based on image data analysis and processing, including:

[0148] The image layering module acquires image data of the surface area of ​​mechanical castings, collects the grayscale values ​​of pixels in the image, makes attribution judgments based on preset grayscale interval standards, divides pixels into corresponding layers, organizes the layers according to the grayscale level order, and establishes a grayscale layered atlas dataset;

[0149] The density extraction module calls the grayscale layered atlas dataset, sets a grid of equal proportions within the image area, counts the number of pixels contained in each grid in the layer, organizes the pixel distribution sequence of the grid under the difference layer, extracts the trend area index based on the continuous change of the pixel distribution, and generates a trend direction area index table;

[0150] The trend construction module calls the trend direction area index table, extracts the image content of the corresponding area under multiple shooting angles, obtains the grayscale gradient mutation points within the boundary range, forms a boundary point sequence, matches the boundary points based on the correspondence between image perspective and position, maps the intersection positions in a unified coordinate system, and establishes a boundary coincidence intersection coordinate matrix;

[0151] The boundary mapping module calls the coordinate matrix of the boundary coincidence intersection points, counts the spatial position changes of the intersection points under the perspective of the difference image, determines whether the horizontal and vertical offset amplitudes are lower than the stability benchmark, and selects the intersection set with the smallest position change range to generate a stable boundary pixel point set;

[0152] The defect annotation module calls the stable boundary pixel point set, searches for the corresponding coordinates in the original image area, compares them with the positions in the trend area index table, extracts the coordinate overlapping area for image annotation processing, outputs the fused image, and generates a defect annotation visualization image result.

[0153] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A mechanical casting detection method based on image data analysis and processing, characterized in that: The following steps are involved: S1: Acquire image data of the casting surface area, extract pixel grayscale values, and classify pixels according to grayscale according to the preset grayscale interval standard. The division results are organized into a multi-layer image structure in layer order to generate a grayscale layered atlas dataset; S2: calling the grayscale layered atlas dataset, setting equal-area grid division in the layer, counting the number of pixels in the grid in each layer of the image to form a multi-layer density sequence, extracting continuous trend areas based on the change relationship in the density sequence, and generating a trend direction area index table; S3: extracting images corresponding to the indicated area under multiple shooting angles according to the trend area index table, extracting pixels with sudden gradient changes at the boundary of the area in each image to form a boundary point sequence, matching the boundary points under multiple viewing angles according to position correspondence, and generating a boundary coincidence intersection coordinate matrix; S4: Call the boundary coincidence intersection coordinate matrix, perform statistics on the spatial coordinate changes of the intersection under multiple images, perform stability judgment on the horizontal and vertical offsets, filter the intersection set with a change amplitude lower than the set standard, and generate a stable boundary pixel point set.

2. The mechanical casting detection method based on image data analysis and processing according to claim 1 is characterized in that: The grayscale layered atlas dataset includes a multi-layer image structure, grayscale interval standards, and pixel grayscale values. The trend area index table includes a multi-layer density sequence at the grid position, changes in the number of pixels in the area, and continuous trend areas. The boundary coincidence intersection coordinate matrix includes a boundary point sequence, matching results under multiple perspectives, and intersection mapping in a unified coordinate system. The stable boundary pixel point set includes a stable intersection set, an image area with coincident coordinates, and an intersection with a change amplitude lower than a set standard.

3. The mechanical casting detection method based on image data analysis and processing according to claim 1 is characterized in that: The specific steps of S1 are: S101: Acquire image data of the casting surface area, call the pixel array of the image grayscale channel, read the grayscale value corresponding to each pixel, establish a mapping relationship between the pixel position and the grayscale value, and generate the image pixel grayscale matrix value; S102: Based on the grayscale matrix values ​​of the image pixels, according to a preset grayscale interval standard, determining the grayscale attribution of the pixels, classifying them into corresponding grayscale labels, sorting the pixel positions corresponding to the grayscales, and obtaining grayscale label division coordinate groups; S103: Divide the coordinate groups according to the grayscale labels, mark the grayscale values ​​of corresponding positions in the image in sequence, organize the layer structure according to the grayscale label sequence, integrate the image data to form a multi-layer image structure, and obtain a grayscale layered atlas dataset.

4. The mechanical casting detection method based on image data analysis and processing according to claim 3 is characterized in that: The specific steps of S2 are: S201: calling the grayscale layered atlas dataset, dividing each layer of the image into equal-area grids, identifying the spatial distribution positions of pixels within the grids, establishing a correspondence between the grid positions and pixel distributions in the layer, and generating grid pixel distribution values; S202: Based on the grid pixel distribution values, construct a sequence of the number of pixels at the same grid position in the layer, determine the continuous direction characteristics of the value change in the sequence, extract the layer segment with a single direction change, and generate a density sequence trend segment value; S203: According to the density sequence trend segment value, the change direction identifiers of adjacent grid positions are compared, the region combinations with the same continuous direction are classified, the trend extension range in the layer is delineated, and a trend direction region index table is generated.

5. The mechanical casting detection method based on image data analysis and processing according to claim 4 is characterized in that: The specific calculation formula for the continuous direction feature of the value change in the judgment sequence is: Calculate the directional consistency eigenvalue and generate the density sequence trend segment value; Among them, H represents the eigenvalue of directional consistency, x f represents the number of pixels in the grid position f, s represents the total length of the sequence, |x f+1 -x f | represents the absolute difference in the number of adjacent grid pixels, (x f+1 -x f ) represents the direction of change in the number of adjacent grid pixels, represents the sum of squares of the sequence, and ∈ represents the minimum value to prevent zero constant.

6. The mechanical casting detection method based on image data analysis and processing according to claim 4 is characterized in that: The specific steps of S3 are: S301: Obtain a multi-angle image corresponding to a number in the trend area index table, locate the corresponding area in the image, extract pixel grayscale change characteristics within the area, mark boundary pixels according to grayscale value mutations, sequentially construct a boundary point set, and generate a region boundary point sequence value; S302: Calling the boundary points under the angle in the region boundary point sequence value, extracting the coordinate corresponding points in the same region according to the position relationship in the image, constructing a matching pair set between view angles, and generating coordinate pair values ​​of the boundary corresponding points between view angles; S303: Convert the coordinate pairs of the corresponding points on the boundaries between the perspectives and the image shooting parameters into the same coordinate system, calculate the intersection positions under multiple perspectives, and integrate to generate a boundary coincidence intersection coordinate matrix.

7. The mechanical casting detection method based on image data analysis and processing according to claim 6 is characterized in that: The specific calculation formula for constructing the matching pair set between perspectives is: Calculate the perspective matching index and generate the coordinate pairs of corresponding points on the boundary between perspectives; Among them, M ik Represents the matching weight value between the i-th perspective and the k-th perspective, Δx ij Represents the absolute difference between the jth boundary point of the i-th perspective and the reference x-coordinate, Δy ij Represents the absolute difference between the jth boundary point of the i-th perspective and the reference y coordinate, θ jk Represents the projection inclination angle radian value of the j-th boundary point at the k-th viewing angle, Δφ ik Represents the change in the optical axis angle between the i-th viewing angle and the k-th viewing angle, Δψ jk Represents the curvature change of the j-th boundary point at the k-th perspective, Represents the three-dimensional unit direction vector of the i-th perspective, R represents the preset field of view overlap radius threshold, and n represents the number of valid boundary points in the current area.

8. The mechanical casting detection method based on image data analysis and processing according to claim 6, characterized in that: The specific steps of S4 are: S401: Calling the pixel positions of the intersection points in the image sequence in the boundary coincidence intersection coordinate matrix, extracting the coordinate information in the image order, establishing a sequence set of image positions corresponding to the intersection points, integrating the image position information of the intersection points, and generating an intersection pixel displacement dataset; S402: extracting the horizontal and vertical displacement amplitudes of the intersection points in the image sequence based on the coordinate changes in the intersection pixel displacement dataset, setting an upper limit for the displacement as a judgment criterion, screening a set of intersection points that meet the conditions, and generating a coordinate stability screening matrix; S403: Calling the intersection point set in the coordinate stability screening matrix, integrating the horizontal and vertical position information under the image, generating an average position point set of the intersection points, constructing the boundary distribution according to the original index, and generating a stable boundary pixel point set.

9. The mechanical casting detection method based on image data analysis and processing according to claim 1, characterized in that: The method further comprises: S5: Based on the stable boundary pixel point set, the corresponding position is searched in the original image, and the position coordinates are compared with the position coordinates in the trend direction area index table, and the image area where the coordinates of the two coincide with each other is extracted, and the regional image is annotated to generate a defect annotation visualization image result; The defect annotation visualization image result includes the detection image after annotation processing, the original image fusion output result, and the visualization defect information; The specific steps of S5 are: S501: Based on the stable boundary pixel point set, position coordinate information in the original image is extracted, and the coordinates are compared point by point with the coordinates in the trend region index table, image positions with consistent coordinate values ​​are screened, and a coordinate coincidence region index value set is generated; S502: Calling the coordinate coincident region index value set, extracting the image content of the corresponding region from the original image, screening the positions that meet the boundary clarity condition based on the edge feature distribution of the image channel, and generating a pixel position set of the labelable region; S503: calling the pixel position set of the annotable area, completing pixel value replacement processing at the corresponding position in the original image, adding a unified boundary mark and performing image fusion to generate a defect annotation visualization image result.

10. A mechanical casting detection system based on image data analysis and processing, characterized in that: The mechanical casting detection method based on image data analysis and processing according to any one of claims 1 to 9, wherein the system comprises: The image layering module acquires image data of the surface area of ​​mechanical castings, collects the grayscale values ​​of pixels in the image, makes attribution judgments based on preset grayscale interval standards, divides pixels into corresponding layers, organizes the layers according to the grayscale level order, and establishes a grayscale layered atlas dataset; The density extraction module calls the grayscale layered atlas dataset, sets a grid of equal proportions within the image area, counts the number of pixels contained in each grid in the layer, arranges the pixel distribution sequence of the grid under the difference layer, extracts the trend area index based on the continuous change of the pixel distribution, and generates a trend direction area index table; The trend construction module calls the trend direction area index table, extracts the image content of the corresponding area under multiple shooting angles, obtains the grayscale gradient mutation points within the boundary range, forms a boundary point sequence, matches the boundary points based on the corresponding relationship between the image perspective and position, maps the intersection position in a unified coordinate system, and establishes a boundary coincidence intersection coordinate matrix; The boundary mapping module calls the coordinate matrix of the boundary coincidence intersection points, counts the spatial position changes of the intersection points under the perspective of the difference image, determines whether the horizontal and vertical offset amplitudes are lower than the stability benchmark, selects the intersection point set with a smaller position change range, and generates a stable boundary pixel point set; The defect annotation module calls the stable boundary pixel point set, searches for the corresponding coordinates in the original image area, compares them with the positions in the trend area index table, extracts the coordinate overlapping area for image annotation processing, outputs the fused image, and generates a defect annotation visualization image result.

Images