PD chip detection system based on shape hierarchy

Abstract

The application discloses a PD chip detection system based on shape hierarchy, and relates to the field of chip appearance defect detection. The system comprises: an image acquisition module: acquiring a PD chip image to obtain an original image; a die positioning module: based on the original image, using a template matching algorithm to obtain a die segmentation image; a feature extraction module: using an edge detection algorithm to extract geometric features to obtain a straight line feature set and a circular feature set; a feature matching module: matching according to a standard area template to generate boundary information set of the current die; a region structure confidence calculation module: based on the boundary information set, calculating a structure confidence score; a defect detection module: based on the boundary information set, the structure confidence score and the die segmentation image, performing defect detection to generate a defect judgment result. Through the fusion of shape hierarchical analysis and intelligent parameter evaluation, the application creatively realizes stable, accurate and adaptive automatic detection of appearance defects of the PD chip.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of chip appearance defect detection, in particular to a PD chip detection system based on shape hierarchy. BACKGROUND

[0002] With the large number of optical modules applied in data centers, the quality stability of PD chips is an important basis for stable work of optical modules, and high requirements are put forward for quality stability. Chip appearance defects are the key factors affecting its reliability.

[0003] At present, the PD chip appearance detection mainly faces the following challenges: the traditional manual visual inspection is low in efficiency and poor in consistency; and the current mainstream automatic optical detection algorithm has obvious limitations. Although the algorithm based on traditional image processing is effective for simple defects, it is difficult to cope with complex defects and lacks generalization ability; although the algorithm based on deep learning is widely applicable, it seriously depends on a large amount of data, and cannot adapt to the inherent batch-to-batch geometric size variation and color fluctuation in the production process of PD chips, resulting in poor detection stability and high maintenance cost.

[0004] In addition, PD chip detection also faces two technical difficulties: one is that the slight deformation caused by process fluctuation is difficult to distinguish from real defects; the other is that the detection accuracy of defects such as cracks is extremely high. Therefore, there is an urgent need in the art for a solution that can adapt to process fluctuations, does not depend on a large amount of data training, and can realize high-precision detection. SUMMARY

[0005] Based on the above-mentioned shortcomings of the prior art, the purpose of the present application is to provide a PD chip detection system based on shape hierarchy to solve the above technical problems.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a PD chip detection system based on shape hierarchy, comprising: an image acquisition module, a die positioning module, a feature extraction module, a feature matching module, a region structure confidence calculation module and a defect detection module;

[0007] The image acquisition module: uses an imaging system to acquire PD chip images to obtain original images;

[0008] The die positioning module: obtains a die segmentation image based on the original image in combination with a template matching algorithm;

[0009] The feature extraction module: extracts geometric features using an edge detection algorithm based on the die segmentation image, and performs geometric feature analysis to obtain a straight line feature set and a circular feature set;

[0010] The feature matching module: matches the straight line feature set and the circular feature set in combination with a pre-set standard region template to generate a boundary information set of the current die;

[0011] Region structure confidence calculation module: based on the boundary information set, calculate the structure confidence score of each to-be-detected region in combination with the standard region template;

[0012] Defect detection module: based on the boundary information set, the structure confidence score and the grain segmentation image, perform defect detection on each to-be-detected region to generate a defect judgment result.

[0013] The application further provides that the image acquisition module comprises:

[0014] The imaging system comprises a high-resolution camera, a high-magnification metallographic lens and a high-illumination point light source.

[0015] The imaging system is used to acquire the image of the PD chip within a preset exposure time to obtain an original image.

[0016] The application further provides that the grain positioning module comprises:

[0017] The size of the Gaussian kernel and the down-sampling ratio are determined according to the magnification of the imaging lens.

[0018] Based on the original image, the Gaussian blur processing and the down-sampling processing are performed in combination with the Gaussian kernel and the down-sampling ratio to obtain a down-sampled image.

[0019] The normalized cross-correlation algorithm or the linear template matching algorithm based on gradient features is used on the down-sampled image to obtain the coarse position and angle of the grain.

[0020] Based on the coarse position and angle of the grain, the corresponding image region is cut from the original image to obtain a grain segmentation image.

[0021] The application further provides that the feature extraction module comprises:

[0022] Based on the grain segmentation image, more than two preset fixed threshold combinations are used to perform Canny edge detection on the same grain segmentation image to obtain multiple edge images.

[0023] For each edge image, the probabilistic Hough transform algorithm is applied to detect straight line segments in the image, the support point set corresponding to each straight line segment is recorded, and the union set operation is performed on the straight line support point sets of all edge images to construct a straight line feature set.

[0024] Meanwhile, for each edge image, the Hough transform circle detection algorithm is applied to detect circles in the image, the support point set corresponding to each circle is recorded, and the union set operation is performed on the circular support point sets of all edge images to construct a circular feature set.

[0025] The application further provides that the feature matching module comprises:

[0026] A standard region template is preset, wherein the standard region template comprises: die outline information, each target line segment constituting the outline, ideal geometric positions of each target line segment, and connection order between line segments;

[0027] For each target line segment, a search range is constructed with the ideal geometric position of the target line segment as the center and a preset physical tolerance as the radius.

[0028] The application is further configured to perform line segment matching based on the set of straight line features and the set of circular features within the search range, search for the best matching object, and perform region reconstruction according to the matching result:

[0029] If no matching line segment is found within the search range, the region to which the current target line segment belongs is marked as feature missing;

[0030] If all line segments are successfully matched within the search range, the successfully matched line segments are geometrically extended to intersect with adjacent matching line segments, and the coordinates of the intersection points are calculated;

[0031] According to the predefined line segment connection order in the standard region template, the intersection point coordinates are sequentially connected to form a boundary polygon of the region to be detected;

[0032] All successfully reconstructed boundary polygons of the region to be detected are integrated to construct a boundary information set.

[0033] The application is further configured that the region structure confidence calculation module comprises:

[0034] Based on the comparison between the boundary information set and the standard region template, a basic difference measure of each region in geometric features is calculated, wherein the basic difference measure comprises: position deviation of each matching line segment endpoint and corresponding endpoint of the standard template, angle or curvature deviation of each matching line segment and corresponding line segment of the standard template, and continuity deviation of the connection point between adjacent matching line segments;

[0035] Based on the predefined topological relationship between each region, a region relationship graph model is constructed;

[0036] Based on the basic difference measure, an initial confidence is calculated for each region;

[0037] Based on the region relationship graph model and the initial confidence, a relationship consistency propagation algorithm is used to make the confidence of each region affected by the confidence of its adjacent regions and iteratively updated;

[0038] The final value of each region after propagation is stabilized is set as the structure confidence score of the corresponding region.

[0039] The application is further configured to have the defect detection module include a shape and size detection unit, an edge defect detection unit, a multi-scale texture stability calculation unit, a crack detection unit, a color anomaly detection unit, and a result generation unit.

[0040] The shape and size detection unit includes, based on the boundary information set, using a structure confidence score to dynamically adjust a size threshold for region shape integrity judgment and dynamic size detection, and generating a shape and size discrimination result.

[0041] The edge defect detection unit includes, based on the boundary information set, using a pointPolygonTest algorithm to measure the maximum convex height and the recess depth, and judging according to a preset concave-convex value to generate an edge concave-convex discrimination result.

[0042] The application is further configured to have the multi-scale texture stability calculation unit include:

[0043] Based on the to-be-detected region image defined by the boundary information set in the grain segmentation image, a Gaussian pyramid is constructed.

[0044] Based on the Gaussian pyramid, a Gaussian difference texture energy map under different scales is calculated.

[0045] By analyzing the correlation between the texture energy maps of each scale, a multi-scale texture stability index is generated.

[0046] The crack detection unit includes:

[0047] The multi-scale texture stability index is compared with a preset sensitivity threshold.

[0048] If the multi-scale texture stability index is lower than the sensitivity threshold, the sensitivity level of crack detection is started or improved.

[0049] Based on the started or improved sensitivity level, the to-be-detected region image is subjected to difference processing of median filtering and Gaussian filtering to obtain enhanced crack features.

[0050] Based on the enhanced crack features, a preset analysis parameter corresponding to the current sensitivity level is used to perform connected region analysis to extract linear defects and generate a crack detection result.

[0051] The application is further configured to have the color anomaly detection unit include:

[0052] Based on the to-be-detected region image, a pyrMeanShiftFiltering algorithm is used to segment the to-be-detected region image to obtain a color segmentation image.

[0053] Based on the color segmentation image, the floodFill algorithm is called to perform region connection, and the main color region and the abnormal color region are obtained, and the color anomaly detection result is generated by comparing the standard region template.

[0054] The result generation unit comprises:

[0055] The shape and size discrimination result, the edge concave-convex discrimination result, the crack detection result, the color anomaly detection result and the structure confidence score are integrated to generate a final defect judgment result.

[0056] The present application provides a PD chip detection system based on shape hierarchy, which comprises an image acquisition module, a die positioning module, a feature extraction module, a feature matching module, a region structure confidence calculation module and a defect detection module.

[0057] The detection mechanism based on shape hierarchy and region structure confidence effectively overcomes the process fluctuation interference: through the hierarchical analysis of "geometric feature matching-region reconstruction-confidence evaluation", the prior shape knowledge is integrated into the detection process, and the region structure confidence is innovatively used to dynamically adjust the detection sensitivity, thereby intelligently distinguishing between process deformation and real defects, ensuring high detection rate while significantly reducing false positives.

[0058] Using the multi-scale texture stability index and its trigger detection mechanism, the precise capture of hidden defects is realized: for weak and complex texture defects, the multi-scale texture stability index is used as the decision basis, so that the system can automatically trigger or enhance the crack and color anomaly detection of specific regions according to the multi-scale texture stability index, forming an intelligent closed loop of "evaluation-decision-action", and greatly improving the detection accuracy and efficiency of such defects.

[0059] The autonomous system combined with model and data driving realizes high stability detection without relying on massive data: the deep learning path relying on big data training is abandoned, and a transparent and interpretable detection system is established by constructing an analysis and decision framework with regional structure confidence and multi-scale texture stability index as the core. The dependence on labeled data is fundamentally reduced, solving the common problem of "small sample, high precision and high stability" in industrial scenes, and ensuring the long-term reliability and low maintenance cost of the system.

[0060] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the following specific embodiments of the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. 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. In the drawings:

[0062] Figure 1 The structure diagram of the PD chip detection system based on shape hierarchy shown in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0063] The embodiments of the present application will be described below with reference to the drawings and preferred embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the specification. The present application can also be implemented or applied by different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, and are not intended to limit the protection scope of the present application.

[0064] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only the components related to the present application are shown in the diagrams, not the number, shape and size of the components when actually implemented. The type, number and proportion of each component in actual implementation can be arbitrarily changed, and the layout type of the components can also be more complex.

[0065] In the following description, numerous specific details are discussed in order to provide a thorough explanation of embodiments of the present application. However, it will be apparent to one of ordinary skill in the art that embodiments of the present application can be practiced without these specific details. In other instances, well-known structures and devices are not described in detail in order to avoid obscuring embodiments of the present application.

[0066] Embodiments:

[0067] The PD chip detection system based on shape hierarchy, as shown in Figure 1 includes:

[0068] An image acquisition module: using an imaging system to acquire PD chip images to obtain original images;

[0069] A die positioning module: based on the original images, combining a template matching algorithm to obtain a die segmentation image;

[0070] A feature extraction module: based on the die segmentation image, using an edge detection algorithm to extract geometric features, performing geometric feature analysis to obtain a straight line feature set and a circular feature set;

[0071] A feature matching module: according to the straight line feature set and the circular feature set, combining a pre-set standard region template to perform matching to generate boundary information set of the current die;

[0072] A region structure confidence calculation module: based on the boundary information set, combining the standard region template to calculate structure confidence scores of each region to be detected;

[0073] A defect detection module: based on the boundary information set, the structure confidence scores and the die segmentation image, performing defect detection on each region to be detected to generate defect judgment results.

[0074] The present application is further provided that the image acquisition module includes:

[0075] The imaging system includes: a high-resolution camera, a high-magnification metal lens and a high-illumination point light source to form an imaging system;

[0076] The imaging system acquires images of the PD chip within a preset exposure time to obtain the raw image. Specifically, the image acquisition module consists of a high-resolution camera, a high-magnification metallurgical lens, and a high-illuminance point light source. Depending on the inspection requirements, the camera target size is 1.2 inches or larger, the frame rate is 70 frames per second or higher, the lens can use a 10x, 20x, or 50x high-resolution metallurgical microscope, and the light source is a white fiber optic point light source with an illuminance exceeding 10,000 lux. This ensures that the exposure time is less than 1ms to prevent image blurring caused by machine vibration. For example, in implementation, the camera and lens are first connected via a C-interface, and the fiber optic light guide is connected to the microscope's vertical illumination path. The initial exposure time is set to 800 microseconds using the camera software development kit. Then, the PD chip wafer is placed on the stage for inspection. In the software preview, the light source is adjusted to maximum brightness, and the exposure time is gradually shortened to 500 microseconds until the image is free of motion blur. Finally, the camera acquires the image within the optimized exposure time, which is then converted by the sensor and digitally processed by the image processor to output a raw image without motion blur to the subsequent processing module.

[0077] The present invention is further configured such that the grain positioning module includes:

[0078] The Gaussian kernel size and downsampling ratio are determined based on the magnification of the imaging lens;

[0079] Based on the original image, Gaussian blurring and downsampling are performed by combining the Gaussian kernel and the downsampling ratio to obtain a downsampled image.

[0080] On the downsampled image, a normalized cross-correlation algorithm or a gradient feature-based linear template matching algorithm is used to obtain the approximate position and angle of the grains;

[0081] Based on the approximate position and angle of the grains, corresponding image regions are extracted from the original image to obtain a grain segmentation image. Specifically, the core purpose of the grain localization module is to solve the matching error problem caused by changes in grain morphology due to process fluctuations under high-magnification imaging. Matching accuracy is ensured by first downsampling the image to eliminate shape errors, and then performing precise localization. In practice, the preprocessing parameters are first determined based on the magnification of the current imaging lens. When using a 10x lens, the default Gaussian kernel size is 3, and the default downsampling ratio is 0.5; for a 20x lens, the kernel size is 7, and the downsampling ratio is 0.25; for a 50x lens, the kernel size is 9, and the downsampling ratio is 0.1. These default values ​​are set because higher magnification lenses amplify image deformation caused by process fluctuations, requiring stronger blurring and downsampling to eliminate subtle deformation noise while preserving the main contour features of the grains. Processing flow: First, the original image is Gaussian blurred using the GaussianBlur function from the OpenCV library. Convolution is performed using a kernel size corresponding to the current lens magnification, and the standard deviation is automatically calculated based on the kernel size. Then, bilinear interpolation downsampling is performed on the blurred image using the previously determined downsampling ratio, resulting in a downsampled image with reduced resolution but more stable features. A normalized cross-correlation algorithm is used for template matching on the downsampled image. The optimal matching position is found by calculating the normalized cross-correlation coefficient between the template image and the search region, exhibiting good stability to illumination changes. The confidence threshold for successful matching is set to 0.75 by default. This default threshold was determined through multiple sample tests to balance the risks of false and missed matches and can be modified appropriately based on actual deployment. As an alternative, the Linemod 2D matching algorithm based on gradient features can also be used. The Linemod 2D matching algorithm has better robustness to complex backgrounds and partial occlusion. After successful matching, the approximate position coordinates and rotation angle of the grain are obtained. Finally, based on these parameters, the original high-resolution image is located in reverse through coordinate transformation and affine transformation. The precisely aligned grain segmentation image is obtained through image cropping, providing high-quality input data for the subsequent feature extraction module.

[0082] The present invention is further configured such that the feature extraction module includes:

[0083] Based on the grain segmentation image, two or more sets of preset fixed threshold combinations are used to perform Canny edge detection on the same grain segmentation image to obtain multiple sets of edge images.

[0084] For each edge image, the probabilistic Hough transform algorithm is applied to detect line segments in the image, and the support point set corresponding to each line segment is recorded. The union operation is performed on the support point sets of all edge images to construct a line feature set.

[0085] Simultaneously, for each edge image, the Hough transform circle detection algorithm is applied to detect circles in the image, and the support point set corresponding to each circle is recorded. A union operation is then performed on the support point sets of all edge images to construct a circular feature set. Specifically, the feature extraction module addresses the problem of stably extracting the geometric features of the PD chip under different lighting and contrast conditions. By employing a multi-threshold edge detection and multi-path feature fusion strategy, it ensures that straight line and circular features can be extracted completely and accurately under various imaging environments. In specific implementation, three fixed Canny edge detection threshold combinations are used. The default values ​​are three sets of high and low thresholds (15, 30), (20, 40), and (30, 60) to process the images, resulting in three sets of images. The high and low thresholds are determined through testing with a large number of qualified samples under different lighting conditions. The low threshold from 15 to 30 covers weak edge information, while the high threshold from 30 to 60 ensures the integrity of strong edges. The three combinations together cover various imaging situations from weak to strong contrast, avoiding the insufficient adaptability of a single threshold under complex imaging conditions. Processing flow: First, Canny edge detection is performed three times in parallel on the same segmented image of the same chip. Each time, a set of preset thresholds is used to generate three sets of edge images containing different levels of detail. For each edge image, the probabilistic Hough transform algorithm is applied to detect line segments. The probabilistic Hough transform algorithm detects lines in the image through an accumulator. The default settings are a minimum line length of 30 pixels and a maximum line segment gap of 10 pixels. The minimum vote threshold is set to 30 by default, which is used to balance detection sensitivity and false positive rate. During algorithm execution, the support point set corresponding to each line segment is recorded, i.e., the original edge pixel coordinates constituting the line segment. Simultaneously, the Hough transform circle detection algorithm is applied to each edge image to detect circular features. The algorithm defaults to a minimum circle radius of 5 pixels, a maximum circle radius of 100 pixels, and a gradient threshold of 50. These default parameters are set based on the actual size range of the PD chip to effectively filter noise interference. During the detection process, the support point set corresponding to each circle is recorded synchronously. Finally, the straight line and circular features are fused separately: all straight line segments and their support point sets detected in the three edge images are combined, and duplicate line segments are removed to construct a unified straight line feature set; similarly, all circular features detected in the three images are combined to construct a unified circular feature set. The multi-path fusion mechanism used here ensures that even if some features are missed under certain lighting conditions, they can be compensated for under other threshold conditions, greatly improving the robustness and completeness of feature extraction.

[0086] The present invention is further configured such that the feature matching module includes:

[0087] A standard region template is preset, which includes: grain contour information, each target line segment constituting the contour, the ideal geometric position of each target line segment, and the connection order between the line segments;

[0088] For each target line segment, a search range is constructed with the ideal geometric position of the target line segment as the center and a preset physical tolerance as the radius. Specifically, the creation of the preset standard region template is based on the geometric specifications of a qualified PD chip. The standard region template contains four levels of definition information: 1. Diode contour information defines the overall outer shape boundary and internal functional area boundary of the chip, clearly distinguishing between the outer and inner contours; 2. Each target line segment constituting the contour is used to decompose the complete contour into a series of interconnected basic geometric elements, including straight line segments and arc segments; 3. The ideal geometric position of each target line segment is used to accurately mark the coordinate parameters of each line segment in the standard coordinate system in micrometers. Straight line segments are defined by the coordinates of the start and end points, and arc segments are defined by the coordinates of the center, radius, and start and end angles; 4. The connection order between line segments strictly defines the ordered connection relationship of all geometric elements to ensure that the complete region contour can be correctly reconstructed. In the specific implementation process: For each target line segment, a circular search area within a preset physical tolerance range is constructed based on the center point marked by its ideal geometric position. The default physical tolerance is 5 micrometers. This default value is based on statistical analysis of fluctuations in PD chip manufacturing processes, covering the maximum possible positional deviation under normal production conditions, while effectively avoiding the risk of mismatches due to an excessively large search range. In practical applications, this tolerance value can be adjusted appropriately according to the process capabilities of different products, typically within a range of 3 to 8 micrometers. This combination of precise geometric templates and tolerance search establishes a stable and reliable comparison benchmark for subsequent feature matching, ensuring both matching accuracy and providing necessary process adaptability.

[0089] The present invention is further configured to perform line segment matching based on the straight line feature set and the circular feature set within the search range, search for the best matching object, and reconstruct the region based on the matching result:

[0090] If no matching line segment is found within the search range, the region to which the current target line segment belongs is marked as having missing features;

[0091] If all line segments are successfully matched within the search range, the successfully matched line segments are geometrically extended to intersect with adjacent matched line segments, and the coordinates of the intersection point are calculated.

[0092] Based on the predefined line segment connection order in the standard area template, the coordinates of each intersection point are connected sequentially to form the boundary polygon of the area to be detected.

[0093] The system integrates the boundary polygons of all successfully reconstructed regions to be detected, constructing a boundary information set. Specifically, based on the constructed search range, each geometric element in the straight line feature set and the circular feature set is matched and evaluated against the current target line segment. The criterion for determining the best match comprehensively considers distance deviation and angle deviation, with a default weight of 0.6 for distance deviation and 0.4 for angle deviation. This weight allocation is based on the principle that positional accuracy is more important than directional accuracy in practical applications, and has been verified through matching experiments. In specific implementation, it can be fine-tuned according to the actual deployment scenario. When no matching line segment is found within the search range, the system marks the complete region to which the current target line segment belongs as having a missing feature state. This determination immediately triggers the interruption of the region reconstruction process and passes this state to the subsequent defect detection module for specialized processing. For all successfully matched line segments, the system performs precise geometric reconstruction. First, each successfully matched line segment is geometrically extended along its direction until it intersects with an adjacent matching line segment. The precise intersection coordinates are calculated by solving the equation of the line intersection point. Subsequently, strictly following the predefined line segment connection order in the standard region template, all calculated intersection coordinates are sequentially connected to form a closed boundary polygon of the region to be detected. During the region integrity verification phase, a default minimum threshold of 70% is set for successful region reconstruction. This means that when the number of successfully matched line segments in a region reaches 70% of the total defined in the template, the region is considered reconstructed successfully. This 70% threshold is determined through statistical analysis to tolerate a certain degree of feature loss while ensuring reconstruction reliability. Finally, the system integrates all successfully reconstructed boundary polygons of the regions to be detected, constructing a complete boundary information set to provide accurate geometric foundation data for subsequent detection modules.

[0094] The present invention is further configured such that the region structure confidence calculation module includes:

[0095] Based on the comparison between the boundary information set and the standard region template, the basic difference measure of each region in terms of geometric features is calculated. The basic difference measure includes: the positional deviation of each matching line segment endpoint from the corresponding endpoint of the standard template, the angle or curvature deviation of each matching line segment from the corresponding line segment of the standard template, and the continuity deviation of the connection points between adjacent matching line segments.

[0096] A regional relationship graph model is constructed based on predefined topological relationships between regions.

[0097] Based on the basic difference measure, an initial confidence level is calculated for each region;

[0098] Based on the regional relationship graph model and the initial confidence level, the relationship consistency propagation algorithm is used to make the confidence level of each region affected by the confidence level of its neighboring regions and to update iteratively.

[0099] After propagation stabilization, the final value of each region is set as the structural confidence score for that region. Specifically, based on the comparison between the boundary information set and the standard region template, three basic difference metrics are calculated: 1. The positional deviation between the endpoints of each matched line segment and the corresponding endpoints of the standard template is calculated using Euclidean distance to reflect the accuracy of the geometric position; 2. The angular or curvature deviation between each matched line segment and the corresponding line segment of the standard template is calculated, with the angular difference for straight segments and the relative error of the radius of curvature for arc segments, to evaluate shape fidelity; 3. The continuity deviation of the connection points between adjacent matched line segments is evaluated by calculating the change in the tangent direction at the connection point to assess contour smoothness. Based on the predefined topological relationships between regions, a region relationship graph model is constructed. The region relationship graph model defines each region to be detected as a graph node, and the adjacency relationship between regions as an edge. The default value of the connection weight is 1.0, and the connection weight for regions with inclusion relationships is increased to 1.5. This parameter setting is based on the consideration that inclusion relationships have stronger structural constraints, and the specific value can also be adaptively changed according to actual deployment needs. When calculating the initial confidence score based on the basic difference metric, the default weight for positional deviation is 0.5, the weight for angle or curvature deviation is 0.3, and the weight for continuity deviation is 0.2. This weighting scheme is determined through regression analysis of multiple sets of samples, based on the principle that positional accuracy has the greatest impact on structural stability. The initial confidence score is calculated using the weighted harmonic mean method, which is more sensitive to extreme values ​​and can better highlight severely mismatched areas. When iteratively updating using the relational consistency propagation algorithm, the default value for the damping factor is set to 0.2. This default value is determined through convergence testing and is used to balance convergence speed and stability. During the actual algorithm execution, the confidence score of each region is adjusted by weighting the confidence scores of its neighboring regions. The influence weight of the confidence scores of neighboring regions is dynamically calculated based on the inter-regional connection strength. The iteration termination condition is set to the change in confidence score of all regions being less than 0.001 in three consecutive iterations. This threshold is used to achieve a balance between computational efficiency and accuracy requirements. Finally, the final value of each region after propagation stabilization is output as the structural confidence score of the corresponding region. The structural confidence score is standardized to a range of 0 to 1. The higher the score, the higher the structural reliability of the region, which is used to provide an important quality assessment basis for the subsequent defect detection module.

[0100] The present invention is further configured such that the defect detection module includes: a shape and size detection unit, an edge defect detection unit, a multi-scale texture stability calculation unit, a crack detection unit, a color anomaly detection unit, and a result generation unit;

[0101] The shape and size detection unit includes: based on the boundary information set, dynamically adjusting the size threshold using the structural confidence score to judge the integrity of the region shape and perform dynamic size detection, generating shape and size discrimination results;

[0102] The edge defect detection unit includes: based on the boundary information set, using the pointPolygonTest algorithm, measuring the maximum protrusion height and indentation depth, judging according to preset protrusion and indentation values, and generating edge protrusion and indentation discrimination results. Specifically, the core function of the shape and size detection unit is to evaluate the integrity and dimensional compliance of the region contour, and to achieve adaptive adjustment of the detection standard through structural confidence score. During implementation, the region shape integrity is first judged by verifying the closure and continuity of the boundary polygons of each region in the boundary information set, and detecting whether there are missing or broken line segments. When any non-closed contour or missing key line segments is found, the region is immediately judged as a shape defect. The size detection stage adopts a dynamic threshold adjustment mechanism. The basic size threshold is set according to the product design specifications, with default values ​​for length and width tolerances of ±2 micrometers. The default value of the dynamic adjustment coefficient is 0.5, which is determined through testing multiple sets of process samples to effectively balance detection sensitivity and false alarm rate. In actual use, the threshold can also be calculated using the formula "basic threshold × (1 + adjustment coefficient × (1 - structural confidence score))". When the structural confidence score is low, the tolerance is automatically relaxed to avoid misjudging process fluctuations as defects; when the confidence score is high, a strict standard is adopted to ensure accurate detection of precise structures. The edge defect detection unit quantifies and evaluates the smoothness of the edge contour using geometric calculation methods. During implementation, calculations are performed based on the ideal boundary polygon in the boundary information set, combined with the actual edge point set extracted from the original image. The `pointPolygonTest` function from the OpenCV library calculates the shortest signed distance from each actual edge point to the corresponding ideal boundary polygon. In the distance calculation, positive values ​​represent convexities, and negative values ​​represent concavities. The system calculates the maximum positive value among all edge points as the maximum convex height and the minimum negative value as the maximum concavity depth. The preset convexity / concavity thresholds are determined by the product smoothness specifications. The default value for the convexity / concavity threshold is 1.2 micrometers, and the default value for the concavity / concavity threshold is 1.5 micrometers. The default values ​​for the preset convexity / concavity thresholds are based on chip packaging process requirements; the slightly larger concavity tolerance is due to the relatively small impact of concavities on performance in the process characteristics. Finally, the edge convexity / concavity discrimination result is generated based on the comparison between the measurement results and the preset thresholds. If either the height of a protrusion or the depth of a depression exceeds the corresponding threshold, the edge is deemed defective; otherwise, it is deemed acceptable. All judgment results, along with specific measurement values, are output to provide a quantitative basis for subsequent comprehensive analysis.

[0103] The present invention is further configured such that the multi-scale texture stability calculation unit includes:

[0104] A Gaussian pyramid is constructed based on the region to be detected defined by the boundary information set in the grain segmentation image;

[0105] Based on the Gaussian pyramid, calculate the energy map of Gaussian difference texture at different scales;

[0106] By analyzing the correlation between texture energy maps at different scales, a multi-scale texture stability index is generated.

[0107] The crack detection unit includes:

[0108] The multi-scale texture stability index is compared with a preset sensitivity threshold.

[0109] If the multi-scale texture stability index is lower than the sensitivity threshold, activate or increase the sensitivity level of crack detection.

[0110] Based on the sensitivity level of activation or enhancement, median filtering and Gaussian filtering are performed on the image of the region to be detected to obtain enhanced crack features.

[0111] Based on enhanced crack features, a connected component analysis is performed using preset analysis parameters corresponding to the current sensitivity level to extract linear defects and generate crack detection results. Specifically, the multi-scale texture stability calculation unit quantifies surface uniformity through multi-scale texture analysis, providing a decision basis for defect detection. In practice, a Gaussian pyramid with three scales is first constructed based on the image of the region to be detected defined by the boundary information set in the grain segmentation image. The scale parameters default to the original image scale, a 2x downsampling scale, and a 4x downsampling scale. The default setting of the scale parameters is based on the requirement of covering the main frequency components in texture analysis theory, which is used to effectively capture texture features from fine to macroscopic. Based on the Gaussian pyramid, Gaussian difference texture energy maps at different scales are calculated. Gaussian difference processing uses two sets of standard deviation parameters, 1.5 and 3.0, which are determined through frequency response analysis to enhance medium-frequency texture information. Texture energy calculation uses the local standard deviation method, with the window size set to 7×7 pixels by default, to achieve a balance between texture detail and statistical stability. A multi-scale texture stability index is generated by analyzing the correlation between texture energy maps at various scales. The Pearson correlation coefficient is used for calculation, and the average of the pairwise correlation coefficients between the three scales is taken as the final index. The default correlation coefficient threshold is set to 0.85, determined based on the statistical distribution of qualified samples, to effectively distinguish between normal and abnormal textures. The crack detection unit employs an intelligent detection mechanism based on texture stability assessment to achieve adaptive allocation of detection resources. In practice, the multi-scale texture stability index is first compared with a preset sensitivity threshold, which is set to 0.7. This threshold is determined through ROC curve analysis to control false alarms while ensuring detection rate. When the multi-scale texture stability index is lower than the sensitivity threshold, the system automatically activates or increases the sensitivity level of crack detection. Sensitivity levels are divided into two levels: normal mode and enhanced mode. Normal mode corresponds to an index higher than the threshold, while enhanced mode corresponds to an index lower than the threshold. In enhanced mode, the detection depth is increased by approximately 40% by default. The default increase is determined through defect detection experiments, and the specific value can be adaptively adjusted according to the actual deployment scenario. Based on the sensitivity level at startup or enhancement, median filtering and Gaussian filtering are applied to the image of the region to be detected. The median filter kernel size is set to 15 by default to effectively suppress speckle noise; the Gaussian filter kernel size is set to 5 by default to smooth the background and preserve crack features. The difference processing is achieved by subtracting the original image from the filtered image, which can highlight linear defect features. Based on the enhanced crack features, connected component analysis is performed using analysis parameters corresponding to the current sensitivity level. In normal mode, the minimum crack length threshold is set to 30 pixels and the aspect ratio threshold is set to 5; in enhanced mode, the corresponding thresholds are adjusted to 20 pixels and 4 to improve sensitivity to fine cracks.Finally, crack detection results are generated through linear defect feature extraction, providing a professional basis for subsequent comprehensive analysis.

[0112] The present invention is further configured such that the color anomaly detection unit includes:

[0113] Based on the image of the region to be detected, the pyrMeanShiftFiltering algorithm is used to segment the image of the region to be detected, and a color segmentation image is obtained.

[0114] Based on the color segmentation image, the floodFill algorithm is called to connect regions to obtain the main color region and abnormal color region. The color anomaly detection result is generated by comparing with the standard region template.

[0115] The result generation unit includes:

[0116] The final defect assessment result is generated by combining the results of shape and size discrimination, edge concavity and convexity discrimination, crack detection, color anomaly detection, and structural confidence score. Specifically, the color anomaly detection unit uses color clustering and region connectivity techniques to detect surface color anomalies. In practice, based on the image of the area to be detected, the pyrMeanShiftFiltering algorithm is used for color segmentation. The default value for the spatial window radius is set to 7, and the default value for the color window radius is set to 11. This parameter combination is determined through color separation effect testing and is used to effectively smooth uniform areas while maintaining the clarity of color boundaries. Based on the color segmented image, the floodFill algorithm is called for region connectivity. The default value for the color tolerance threshold is set to 15. The default value for the color tolerance threshold is set based on the human eye's perception limit of color differences and is used to effectively distinguish between true color anomalies and minor color differences caused by illumination. After the connectivity operation, the main color area and the abnormal color area are obtained. The main color area is determined based on the principle of the highest area proportion. In the color anomaly assessment stage, the color features of the abnormal area are compared with the standard color range in the standard area template. The default color difference threshold is set to ΔE10. This default value is determined based on the CIELAB color space standard and reflects color differences perceptible to the human eye. Simultaneously, the abnormal area threshold is set to 3% of the total area. This threshold is determined through statistical analysis and effectively filters out insignificant micro-color spots. The result generation unit uses a weighted decision-making mechanism to synthesize the various detection results and generate the final judgment. In practice, each detection result is first assigned a weight coefficient: shape and size discrimination results have a default weight of 0.3, crack detection results have a weight of 0.25, color anomaly detection results have a weight of 0.25, and edge unevenness discrimination results have a weight of 0.2. This weight allocation is based on the assessment of the impact of each defect type on product reliability. Among these, shape and size defects have the most significant impact and are therefore assigned the highest weight. During the comprehensive judgment process, the structural confidence score is used as a correction factor in the calculation. When the structural confidence score is below 0.6, the system automatically increases the stringency of each defect judgment, raising the judgment threshold by approximately 20%. This correction mechanism is based on the consideration of potential structural risks in low-confidence areas. The final defect judgment threshold is set to 0.85 by default. This default value is determined through precision-recall curve analysis to ensure a high detection rate while keeping the false positive rate within an acceptable range. All judgment results are processed according to the number of each crystal grain, generating the test results and original images, along with detailed data for each test item, to form a complete quality inspection report. This report is then written to a MySQL database to provide data support for process improvement.

[0117] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A PD chip detection system based on shape hierarchy, characterized in that, include: Image acquisition module: Acquires images from the PD chip using an imaging system to obtain raw images; Grain localization module: Based on the original image, it obtains a grain segmentation image by combining a template matching algorithm; Feature extraction module: Based on the grain segmentation image, geometric features are extracted using an edge detection algorithm, and geometric feature analysis is performed to obtain a straight line feature set and a circular feature set. The feature extraction module includes: Based on the grain segmentation image, Canny edge detection is performed on the same grain segmentation image using two or more sets of preset fixed threshold combinations to obtain multiple sets of edge images. For each edge image, the probabilistic Hough transform algorithm is applied to detect straight line segments in the image, and the support point set corresponding to each straight line segment is recorded. The union operation is performed on the straight line support point sets of all edge images to construct a straight line feature set. At the same time, for each edge image, the Hough transform circle detection algorithm is applied to detect circles in the image, and the support point set corresponding to each circle is recorded. The union operation is performed on the circle support point sets of all edge images to construct a circle feature set. Feature matching module: Based on the straight line feature set and the circular feature set, and combined with a preset standard region template, a matching is performed to generate the boundary information set of the current grain. The feature matching module includes: A standard region template is preset, which includes: grain contour information, each target line segment constituting the contour, the ideal geometric position of each target line segment and the connection order between the line segments. The grain contour information defines the overall external boundary and internal functional region boundary of the chip. For each target line segment, a search range is constructed with the ideal geometric position of the target line segment as the center and a preset physical tolerance as the radius. Within the search range, line segment matching is performed based on the straight line feature set and the circular feature set to search for the best matching object. Region reconstruction is then performed based on the matching results. If no matching line segment is found within the search range, the region to which the current target line segment belongs is marked as having missing features; If all line segments are successfully matched within the search range, the successfully matched line segments are geometrically extended to intersect with adjacent matched line segments, and the coordinates of the intersection point are calculated. Based on the predefined line segment connection order in the standard area template, the coordinates of each intersection point are connected sequentially to form the boundary polygon of the area to be detected. Integrate the boundary polygons of all successfully reconstructed regions to be detected to construct a boundary information set; Regional structure confidence calculation module: Based on the boundary information set and combined with the standard regional template, the module calculates the structure confidence score of each region to be detected. The regional structure confidence calculation module includes: Based on the comparison between the boundary information set and the standard region template, the basic difference measure of each region in terms of geometric features is calculated. The basic difference measure includes: the positional deviation of each matching line segment endpoint from the corresponding endpoint of the standard template, the angle or curvature deviation of each matching line segment from the corresponding line segment of the standard template, and the continuity deviation of the connection points between adjacent matching line segments. A regional relationship graph model is constructed based on predefined topological relationships between regions. Based on the basic difference measure, an initial confidence level is calculated for each region; Based on the regional relationship graph model and the initial confidence level, the relationship consistency propagation algorithm is used to make the confidence level of each region affected by the confidence level of its neighboring regions and to update iteratively. Set the final value of each region after propagation stabilization as the structural confidence score of the corresponding region; Defect detection module: Based on boundary information set, structural confidence score and grain segmentation image, it performs defect detection on each region to be detected and generates defect judgment results.

2. The PD chip detection system based on shape hierarchy structure according to claim 1, characterized in that, The image acquisition module includes: The imaging system consists of a high-resolution camera, a high-magnification metallurgical lens, and a high-illuminance point light source. The imaging system acquires images of the PD chip within a preset exposure time to obtain the original image.

3. The PD chip detection system based on shape hierarchy structure according to claim 1, characterized in that, The grain positioning module includes: The Gaussian kernel size and downsampling ratio are determined based on the magnification of the imaging lens; Based on the original image, Gaussian blurring and downsampling are performed by combining the Gaussian kernel and the downsampling ratio to obtain a downsampled image. On the downsampled image, a normalized cross-correlation algorithm or a gradient feature-based linear template matching algorithm is used to obtain the approximate position and angle of the grains; Based on the approximate position and angle of the grains, the corresponding image region is extracted from the original image to obtain the grain segmentation image.

4. The PD chip detection system based on shape hierarchy structure according to claim 1, characterized in that, The defect detection module includes: a shape and size detection unit, an edge defect detection unit, a multi-scale texture stability calculation unit, a crack detection unit, a color anomaly detection unit, and a result generation unit; The shape and size detection unit includes: based on the boundary information set, dynamically adjusting the size threshold using the structural confidence score to judge the integrity of the region shape and perform dynamic size detection, generating shape and size discrimination results; The edge defect detection unit includes: based on the boundary information set, using the pointPolygonTest algorithm, measuring the maximum protrusion height and depression depth, judging according to the preset protrusion and depression values, and generating edge protrusion and depression discrimination results.

5. The PD chip detection system based on shape hierarchy structure according to claim 4, characterized in that, The multi-scale texture stability calculation unit includes: A Gaussian pyramid is constructed based on the region to be detected defined by the boundary information set in the grain segmentation image; Based on the Gaussian pyramid, calculate the energy map of Gaussian difference texture at different scales; By analyzing the correlation between texture energy maps at different scales, a multi-scale texture stability index is generated. The crack detection unit includes: The multi-scale texture stability index is compared with a preset sensitivity threshold. If the multi-scale texture stability index is lower than the sensitivity threshold, activate or increase the sensitivity level of crack detection. Based on the sensitivity level of activation or enhancement, median filtering and Gaussian filtering are performed on the image of the region to be detected to obtain enhanced crack features. Based on enhanced crack features, using preset analysis parameters corresponding to the current sensitivity level, connected region analysis is performed to extract linear defects and generate crack detection results.

6. The PD chip detection system based on shape hierarchy structure according to claim 5, characterized in that, The color anomaly detection unit includes: Based on the image of the region to be detected, the pyrMeanShiftFiltering algorithm is used to segment the image of the region to be detected, and a color segmentation image is obtained. Based on the color segmentation image, the floodFill algorithm is called to connect regions to obtain the main color region and abnormal color region. The color anomaly detection result is generated by comparing with the standard region template. The result generation unit includes: The final defect determination result is generated by combining the results of shape and size discrimination, edge concavity and convexity discrimination, crack detection, color anomaly detection, and structural confidence score.

Images