Semiconductor wire bonding high-precision 3D detection method based on monocular visible light camera

Abstract

The application relates to the technical field of semiconductor package detection, in particular to a high-precision 3D detection method for semiconductor lead bonding based on a monocular visible light camera, which comprises the following steps: obtaining an image sequence based on an ultralow depth-of-field optical system and Z-axis continuous scanning, determining a lead surface best focus point layer by layer through image alignment and analysis by means of focusing, reconstructing a complete three-dimensional model by combining lead trajectory tracking and cubic spline interpolation, and comparing the three-dimensional model with CAD data to realize high-precision detection of defects such as height, radian, coplanarity and fracture. Through the technical scheme, micron-level three-dimensional topography reconstruction can be realized under the premise of using only one visible light camera, and the method has the advantages of simple structure, strong anti-interference capability, no detection blind area and suitability for online detection of vehicle-mounted power devices.

Description

Technical Field

[0001] This invention relates to the field of semiconductor packaging inspection technology, and more specifically to a high-precision 3D inspection method for semiconductor wire bonding based on a monocular visible light camera. Background Technology

[0002] As semiconductor packaging technology rapidly advances towards higher density and performance, wire bonding, as a critical process in chip packaging, directly impacts the electrical performance and long-term reliability of devices. Especially in high-reliability applications such as automotive power semiconductors, industry standards impose mandatory requirements on the appearance quality inspection throughout the packaging process. Traditional quality control methods relying on pre-production equipment debugging and limited manual sampling are no longer sufficient to meet the demands for mass production consistency and defect traceability. Currently, automated inspection systems based on machine vision are gradually replacing human visual inspection, becoming the core means of ensuring wire bonding quality. Precise quantification of parameters such as the three-dimensional morphology, height, curvature, and spatial position of the wires has become an unavoidable technical focus.

[0003] Among these technologies, the detection techniques for two-dimensional features such as the position, shape, and area of ​​wire bonding points and the horizontal integrity of devices are relatively mature and can be efficiently identified using a monocular visible light camera. However, for key 3D parameters such as the vertical height of the wires, their three-dimensional spatial orientation, the position, area, and shape of the bonding points, and the coplanarity of the IC, traditional monocular imaging cannot effectively measure them due to a lack of depth information. Therefore, the industry commonly uses multi-view vision systems, which reconstruct the 3D model of the wires through multi-view synchronous imaging combined with triangulation principles. While this method possesses micron-level accuracy and online detection capabilities, its system architecture is complex, requiring high-precision calibration of multiple camera poses. Furthermore, it is susceptible to interference from factors such as strong reflections from metal surfaces, mutual occlusion between dense wires, or wire orientation parallel to the viewing axis, leading to feature matching failures or reconstruction distortions. This makes it difficult to stably apply to high-requirement scenarios such as automotive-grade power devices.

[0004] In existing technologies, the high cost, high maintenance difficulty, and blind spots under certain geometric conditions of multi-view solutions severely restrict their widespread application in large-scale packaging and testing production lines. Meanwhile, although monocular systems have a simple structure, their limited depth of field and lack of effective depth resolution mechanisms have prevented them from overcoming the bottleneck of 3D measurement. Therefore, there is an urgent need for a monocular 3D inspection method that balances high precision, high stability, and low cost, capable of accurately reconstructing the complete three-dimensional morphology of leads and bonding points without relying on multi-view synchronous imaging, to meet the increasingly stringent end-to-end quality control requirements of fields such as automotive semiconductors. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera. By combining a single visible light camera with an ultra-narrow depth-of-field objective lens and a Z-axis precision scanning mechanism, high-precision reconstruction of the three-dimensional morphology of semiconductor wires is achieved, breaking through the technical bottleneck that traditional monocular vision systems cannot acquire depth information.

[0006] To achieve the above objectives, the present invention provides a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, comprising: Step 1: Image acquisition. The lead frame to be tested is precisely positioned directly below the microscope lens using the X-axis and Y-axis motion platform. The Z-axis drive device is controlled to drive the camera assembly to continuously autofocus and capture images between the preset start height ZStart and end height ZEnd in step intervals of 5 to 10 micrometers, acquiring an image sequence covering the entire lead height range. Each image is accompanied by its corresponding Z-axis coordinate information and stored in the data buffer. Step 2: Image alignment processing. Extract the start frame and end frame of the image sequence respectively, obtain the clearest image of the corresponding ROI region, and perform image alignment processing using the clearest image, start frame and end frame; use the initial point of the image as the base point to compensate for the position of subsequent non-overlapping images. Step 3: Focusing analysis. For each target lead within the field of view, within its preset XY coordinate range, the brightness distribution and edge sharpness of the corresponding area in the image sequence are analyzed layer by layer along the Z-axis. The local bright area with the strongest surface reflection and the sharpest contour edge in the current layer is identified as the best focal point at that location. Its X, Y and corresponding Z-axis coordinates are recorded to form the initial three-dimensional point set. Step 4: Lead trajectory tracking. Starting from the height group where the initial bonding point BondStart is located, based on the identified focal positions, search for subsequent focal points on the same lead path in the adjacent N frames above and below. Set the search neighborhood to within ±20 micrometers of the current point's XY coordinates. Determine whether there are candidate points that meet the following criteria: high brightness intensity threshold greater than 180 gray units, edge gradient modulus greater than 15 gray levels / pixel, and shape conforming to cylindrical projection characteristics. If found, update the current position and continue tracking until the height group where the termination bonding point BondEnd is located is reached and its geometry is successfully identified. Step 5: 3D model reconstruction. Summarize the XYZ coordinate data of all tracked lead wire focal points, use cubic spline interpolation algorithm to smoothly fit the point series, generate a continuous lead wire 3D spatial direction curve, and reconstruct the bonding point morphology based on the maximum bright area of ​​each bonding point and the 2D projection contour to construct a complete 3D geometric model containing all leads and bonding points. Step 6: Defect detection and analysis. The reconstructed 3D model is compared with the pre-stored CAD design data item by item. The actual lead height, radius of curvature, tilt angle, coplanarity deviation, bonding point size, and whether there is breakage or offset are calculated. When any parameter exceeds the set tolerance range, it is marked as a defect and an inspection report is generated.

[0007] Preferably, the objective lens used in step 1 is an ultra-narrow depth-of-field objective lens with a depth of field of less than 10 micrometers. It is used with a 10x magnification for gold wire detection or a 5x magnification for aluminum wire detection to adapt to different wire diameters and field of view requirements. The diameter of the gold wire ranges from 10 micrometers to 25 micrometers, and the wire height ranges from 500 micrometers to 1500 micrometers. The diameter of the aluminum wire is greater than 50 micrometers, and the wire height ranges from 500 micrometers to 2000 micrometers. The light source system consists of a ring of white or blue LEDs fixed around the lens. The illumination angle is adjustable to ensure that the metal lead is significantly bright in the focal plane while the background remains dark. At the same time, a coaxial point light source is configured to enhance the imaging contrast of the bonding point area.

[0008] Preferably, in step 1, the reference plane setting process involves visually confirming the optimal focusing state of the LeadFrame surface. The Z-axis coordinate at this point is defined as BaseHeight, which serves as the reference origin for all subsequent height measurements. The height group GroupHeight is pre-divided according to the different device surface heights marked in the chip design drawings. The starting shooting position ZStart is set to the lowest height group minus α micrometers, and the ending shooting position ZEnd is set to the highest lead point plus α micrometers, where α is 30 micrometers. The Z-axis movement step size Step is set to 10 micrometers, and approximately 300 images are collected within a total height range of 2 millimeters.

[0009] Preferably, in step 2, the feature template search method performs template matching based on horizontal straight line alignment; when there are complex scenes, morphological feature matching is performed; a feature correspondence between the template image and the real-time image is established, and a global transformation matrix is ​​calculated to eliminate image offset caused by mechanical vibration or positioning error, ensuring that the image sequence has sub-pixel-level spatial consistency in subsequent analysis.

[0010] Preferably, the focus determination criteria in step 3 include maximizing local contrast, detecting the peak response of the Laplacian operator, and evaluating the Tenengrad gradient function. A corresponding sliding window is set according to the lead wire direction for sequential tracking processing. When a sphere exists, the center of the sphere is designated as the initial point for sequential tracking processing. Multiple analysis units are divided into multiple analysis units according to a preset grid within each field of view. The optimal focusing layer of the lead wire segment in each region is tracked, and the optimal Z value and its confidence score at the center point of each unit are output.

[0011] Preferably, in step 4, the lead tracking starting point selection strategy prioritizes searching from the bond point with the lower Z coordinate. The system initiates initial recognition below the BondStart height group. The number of image frames for initial recognition is set according to the 3D shape of the lead, and the search range is limited to a rectangular area of ​​±50 micrometers in its preset XY coordinates. Once recognition is successful, the iterative tracking process is started. Each iteration reduces the search range to ±20 micrometers. If the next focal point that meets the conditions is not found for 10 consecutive frames, it is determined to be a broken or severely bent failure.

[0012] Preferably, in step 5, the three-dimensional curve fitting adopts the weighted least squares method, which assigns higher weight coefficients to high-confidence focal points, and controls the fitting residual within ±2 micrometers. At the same time, a two-dimensional panoramic image is synthesized based on the local image blocks around each focal point in the image sequence, which unfolds along the lead line to assist in manual review and defect classification.

[0013] Preferably, in step 6, the defect judgment threshold is dynamically configured according to the product specifications. The allowable deviation of the lead height is ±15 micrometers, the radius of curvature is not less than 80 micrometers, the coplanarity error does not exceed the process threshold, and the bonding point area change rate is greater than ±25% as abnormal. All test results are summarized to generate a unit-level comprehensive test report, which can be exported to a standardized data format for MES system access. Among them, the process threshold is the chip feature threshold, which is 20-50 micrometers.

[0014] Preferably, it also includes a multi-module parallel detection architecture, which divides the entire LeadFrame into N multiplied by M independent detection units. Each unit is further subdivided into several field of view areas. Multiple identical detection modules synchronously execute the above process to achieve batch automated 3D detection of the entire lead frame. The cycle time of a single complete detection is controlled within 90 seconds, which meets the requirement of 10-15 seconds for the cycle time of online high-speed production lines.

[0015] Preferably, the method is applicable to automotive power semiconductor packaging production lines, capable of stably detecting fine-pitch lead layouts with line spacing less than 50 micrometers, and has good anti-interference capabilities for cross-line routing, dense arrangement and mirror-reflective gold lines. The axial height resolution reaches the 5-micrometer level, the spatial repeatability positioning accuracy is better than ±0.8 micrometers, the whole system does not require multi-camera synchronous calibration, and the maintenance cost is significantly lower than that of traditional multi-view vision solutions.

[0016] Compared with the closest existing technology, the present invention has the following advantages: By combining a single visible light camera with an ultra-narrow depth-of-field objective lens and a Z-axis precision scanning mechanism, high-precision reconstruction of the 3D morphology of semiconductor leads is achieved. This breakthrough overcomes the technical bottleneck of traditional monocular vision systems' inability to acquire depth information and avoids reconstruction distortion caused by problems such as pose calibration errors between cameras and feature matching failures due to strong reflections from metal surfaces in multi-view vision systems. Employing the focusing method principle, it utilizes the physical property of enhanced reflection within the focal plane of the metal lead. Through layer-by-layer image analysis, the optimal focus position of each point is determined, thereby accurately reconstructing its XYZ 3D coordinates. This method is particularly suitable for micron-level surface morphology detection of highly reflective materials such as gold and aluminum wires. The hardware structure is simple, requiring only one optical imaging system and a three-axis motion platform, significantly reducing equipment costs and maintenance difficulty. It is also unaffected by external vibrations and relative displacements, exhibiting excellent detection stability. By spatially aligning and tracking the image sequence, continuous modeling of the entire lead path is achieved, enabling comprehensive evaluation of various defect types such as lead height, curvature, coplanarity, breakage, and offset. Detection is blind-spot-free and adaptable to complex wiring scenarios. By combining multi-module parallel operation mode, the speed disadvantage of layer-by-layer scanning is effectively compensated for, meeting the online application requirements of large-scale packaging and testing production lines. The overall solution ensures micron-level detection accuracy while also featuring high stability, low cost, and strong adaptability, providing reliable technical assurance for the packaging quality of automotive-grade semiconductor devices. Attached Figure Description

[0017] Figure 1 This is a flowchart of a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera provided by the present invention; Figure 2 This is a schematic diagram of a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera provided by the present invention. Figure 3 This is a schematic diagram of lead changes in a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera provided by the present invention. Figure 4 This is a schematic diagram of wire bonding synthesis for a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera provided by the present invention.

[0018] Figure label: 1. High-resolution industrial camera; 2. 1X dedicated microscope barrel; 3. White point light source; 4. 10X ultra-low depth-of-field objective lens; 5. Blue / white ring light source. Detailed Implementation

[0019] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 This invention provides a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, such as... Figure 1 As shown, it includes: a three-axis motion platform, a monocular visible light imaging module, an image processing unit, and a defect analysis module. Among them, as... Figure 2 As shown, the system includes a high-resolution industrial camera 1, a 1X dedicated microscope barrel 2, a white point light source 3, a 10X ultra-low depth-of-field objective lens 4, and a blue / white ring light source 5. The three-axis motion platform includes an X-axis motion mechanism, a Y-axis motion mechanism, and a Z-axis drive device. The monocular visible light imaging module consists of an ultra-narrow depth-of-field objective lens, a high-resolution CMOS industrial camera, and an adjustable ring LED light source. The image processing unit performs image acquisition, alignment, focus analysis, trajectory tracking, and 3D reconstruction. The defect analysis module compares the image with the CAD model and outputs a test report. The core principle of this invention lies in utilizing the physical property of enhanced reflection within the focal plane of metal leads. By continuously scanning along the Z-axis to acquire image sequences and analyzing the optimal focus state of each pixel region layer by layer, the corresponding Z-axis coordinates are determined, ultimately achieving 3D topographic reconstruction.

[0022] In this embodiment, the high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera includes an adjustable ring LED light source. Specifically, the method employs a 10x (10×) objective lens for gold wire observation to match the fine structural features of the gold wire; a 5x (5×) objective lens for aluminum wire observation to cover the distribution range of the aluminum wire with a larger imaging field of view; a ring light source is configured as the core illumination component, which is coaxially fixed with the camera lens. The central aperture of the light source is precisely aligned with the lens optical path to ensure that light is uniformly projected onto the surface of the sample along the lens axis. The emitting surface of the ring light source maintains a preset distance from the objective lens end face, and the light illuminates the metal wire at a ring-shaped divergence angle, avoiding shadow occlusion caused by unidirectional illumination. Based on a linkage movement mechanism, the focal plane of the objective lens is always within the optimal illumination range of the ring light source: when the focal plane is positioned on the plane where the metal wire is located, the light forms a specular reflection on the wire surface, and the reflected light returns to the camera's photosensitive element along the optical path; light from non-focal plane areas cannot effectively enter the lens due to diffuse reflection or deviation in reflection angle. Ultimately, this maximizes the intensity of light reflection at the focal plane, resulting in a bright and clear outline and details of the metal lead in the image, while the background outside the focal plane is in a low-brightness state, significantly improving the imaging contrast and recognizability of the metal lead.

[0023] In the aforementioned high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, step 1 involves precisely positioning the wire frame under test directly below the microscope lens using X-axis and Y-axis motion platforms. The Z-axis drive device is then controlled to move the camera assembly in 5-10 micrometer step intervals, continuously autofocusing and capturing images between a preset starting height ZStart and an ending height ZEnd. This acquires an image sequence covering the entire wire height range, with each image accompanied by its corresponding Z-axis coordinate information, stored in a data buffer. Specifically, the objective lens used in step 1 is an ultra-narrow depth-of-field objective lens with a depth of field less than 10 micrometers. It is used with a 10x magnification for gold wire detection or a 5x magnification for aluminum wire detection to accommodate different wire diameters and field-of-view requirements. The gold wire diameter range is 10-25 micrometers, and the wire height range is 500-1500 micrometers. The aluminum wire diameter is greater than 50 micrometers, and the wire height range is 500-2000 micrometers. The light source system consists of a ring of white or blue LEDs fixed around the lens, with an adjustable illumination angle. This ensures that the metal leads are significantly bright within the focal plane while the background remains dark. A coaxial point light source is also used to enhance the imaging contrast of the bonding point area. The reference plane setup process involves visually confirming the optimal focus state of the LeadFrame surface. The Z-axis coordinate at this point is defined as BaseHeight and serves as the reference origin for all subsequent height measurements. The height group (GroupHeight) is pre-divided according to the different device surface heights indicated in the chip design drawings. The starting shooting position (ZStart) is set to the lowest height group minus α micrometers, and the ending shooting position (ZEnd) is set to the highest lead point plus α micrometers, where α is 30 micrometers. The Z-axis movement step is set to 10 micrometers, and approximately 300 images are acquired within a total height range of 2 millimeters. The Z-axis drive device uses a closed-loop controlled piezoelectric ceramic driver or a high-precision servo motor, with a repeatability better than ±0.8 micrometers, ensuring sub-micrometer accuracy for the height information corresponding to each frame. During image acquisition, the CMOS camera's exposure time, gain parameters, and light source brightness are synchronized to maintain the stability of the image's grayscale distribution and avoid focus judgment errors caused by lighting fluctuations. All images are stored in a high-speed solid-state drive data buffer, numbered sequentially along the Z-axis. Each frame is appended with metadata fields recording the Z-coordinate, exposure parameters, light source configuration, and timestamp.

[0024] In the aforementioned high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, step 2 involves extracting alignment features from each frame of the image sequence. Horizontal and vertical edge search or feature template matching methods are used to identify the geometric boundaries or specific texture patterns of ICs, diodes, or other rectangular components on the LeadFrame. Pixel-level comparison is used to calculate the positional deviation of each frame relative to the standard template image. An affine transformation algorithm is then used to perform spatial coordinate correction on all images, ensuring that the image sequence maintains a consistent spatial alignment in the XY plane. Specifically, in step 2, the feature template search method uses SIFT or ORB feature point matching algorithms to extract key corner points and edge structures in the image, establishing a feature correspondence between the template image and the real-time image. A global transformation matrix is ​​calculated to eliminate image offsets caused by mechanical vibration or positioning errors, ensuring sub-pixel-level spatial consistency of the image sequence during subsequent analysis. Image alignment processing first selects an intermediate frame from the image sequence as a reference frame, which typically corresponds to the optimal focus state on the LeadFrame substrate surface. The system identifies at least four stable and evenly distributed alignment feature points in the reference frame. These feature points originate from rectangular pads, chip edges, or dedicated alignment marks on the LeadFrame. For each frame of the image to be corrected, the system performs the same feature extraction operation within its corresponding region and eliminates mismatched points using the RANSAC algorithm, ultimately solving for the optimal affine transformation matrix. This transformation matrix contains translation, rotation, and scaling components, used to map the current frame to the reference frame coordinate system. The corrected image is resampled to a format perfectly aligned with the reference frame, with pixel offset errors controlled within 0.3 pixels. This process effectively eliminates the slight XY drift that may be introduced during Z-axis movement, providing spatially consistent basic data for subsequent focus analysis.

[0025] In the aforementioned high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, step 3 involves analyzing the brightness distribution and edge sharpness of the corresponding region in the image sequence layer by layer along the Z-axis within a preset range of the target wire's XY coordinates in each field of view. The method identifies the local bright area with the strongest surface reflection and sharpest contour edge in the current layer as the optimal focal point at that location, recording its X, Y, and corresponding Z-axis coordinates to form an initial three-dimensional point set. Specifically, the focusing criteria in step 3 include maximizing local contrast, detecting the peak response of the Laplacian operator, and evaluating the Tenengrad gradient function. The sliding window size is set to 50 x 50 pixels. Multiple analysis units are divided within each field of view according to a preset grid. The optimal focal layer for the wire segment in each region is tracked, and the optimal Z-value and its confidence score at the center point of each unit are output. The focusing analysis process first constructs a two-dimensional grid within a preset wire path region, with a grid spacing of 20 micrometers, covering all possible wire segments. For each grid node, the system traverses the entire Z-axis image sequence within a 50x50 pixel window around it, calculating the focus evaluation function value for each layer. The focus evaluation function comprehensively employs three metrics: local contrast, defined as the difference between the maximum and minimum grayscale values ​​within the window; the Laplacian operator response, the mean of the absolute values ​​after taking the second derivative of the image; and the Tenengrad gradient function, the sum of the squared gradients of the Sobel operator in the X and Y directions. These three metrics are normalized and then weighted, with weighting coefficients of 0.4, 0.3, and 0.3, respectively. The system records the maximum value of the focus evaluation function within the window and its corresponding Z-layer index, using this Z-value as a preliminary height estimate for the grid node. Simultaneously, the system calculates the sharpness of this peak, i.e., the sum of the differences between the maximum value and the evaluation function values ​​of its two adjacent layers, as a confidence score. Nodes with a confidence score below the threshold of 0.6 are marked as invalid and do not participate in subsequent processing. Ultimately, the X and Y coordinates of all valid grid nodes and their corresponding optimal Z values ​​together constitute the initial three-dimensional point set, with a point density of approximately 2,500 points per square millimeter.

[0026] In the above-mentioned high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, step 4 starts from the height group where the starting bonding point BondStart is located. Based on the identified focal position, it searches for subsequent focal points on the same wire path in adjacent upper and lower N frames of images. The search neighborhood is set to within ±20 micrometers of the current point's XY coordinates. It is determined whether there are candidate points that meet the following conditions: a high brightness intensity threshold greater than 180 gray units, an edge gradient modulus greater than 15 gray levels / pixel, and a shape that conforms to the cylindrical projection characteristics. If found, the current position is updated and tracking continues until the height group where the ending bonding point BondEnd is located is reached and its geometric shape is successfully identified. Specifically, in step 4, the lead tracing starting point selection strategy prioritizes searching from bond points with lower Z coordinates. The system initiates initial recognition within five consecutive frames below the BondStart height group, limiting the search range to a rectangular area of ​​±50 micrometers in its preset XY coordinates. Once recognition is successful, an iterative tracing process begins, with the search range narrowing to ±20 micrometers in each iteration. If no suitable next focal point is found for three consecutive frames, the lead is considered broken or severely bent. The lead trajectory tracing process first obtains the theoretical XY coordinates of the starting bond point (BondStart) and ending bond point (BondEnd) for each lead based on the CAD design data. The system searches downwards near the Z layer corresponding to the BondStart height group, setting the search image frames according to the lead's 3D morphology, and searching for an initial focal point that meets the conditions within a rectangular area of ​​±50 micrometers. Candidate points must simultaneously meet the following criteria: grayscale value greater than 180, Tenengrad gradient value greater than 15, and their local morphology conforming to the characteristics of an arc or straight line segment after Hough transform detection. Once an initial point is found, the system enters the iterative tracing stage. In each iteration, the system searches N=3 frames of images upwards and downwards within a circular neighborhood of ±20 micrometers, centered on the current focal point, to find the next focal point that meets the criteria. The selection of a new focal point is based on the highest comprehensive score, which is calculated by weighting grayscale intensity, gradient modulus, and morphological matching degree. The tracking process proceeds bidirectionally along the lead wire until the BondEnd height group is reached and the geometry of the termination bonding point is successfully identified within its ±30 micrometer range. If no valid candidate point is found, the lead wire is determined to have a breakage or severe bending defect, the tracking process terminates, and the abnormal location is recorded. All successfully tracked focal points are arranged in spatial order to form an ordered three-dimensional point sequence.

[0027] In the aforementioned high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, step 5 involves summarizing the XYZ coordinate data of all tracked wire focal points, using a cubic spline interpolation algorithm to smoothly fit the point sequence, generating a continuous 3D spatial curve for the wires. Simultaneously, the morphology of each bonding point is reconstructed based on the area of ​​the maximum bright region and the 2D projected contour, constructing a complete 3D geometric model containing all wires and bonding points. Specifically, in step 5, the 3D curve fitting uses a weighted least squares method, assigning higher weight coefficients to high-confidence focal points, controlling the fitting residual within ±2 micrometers. Simultaneously, a 2D panoramic image unfolding along the wire direction is synthesized based on local image patches surrounding each focal point in the image sequence, used to assist in manual verification and defect classification. The 3D model reconstruction first filters the ordered point sequence of each wire to remove outliers. Then, a smooth 3D spatial curve is generated using a cubic spline interpolation algorithm, with the interpolation nodes being the original focal points. To improve fitting accuracy, the system introduces a weighting mechanism, with the weighting coefficients proportional to the confidence score calculated in step 3. After fitting, the system calculates the curvature and torsion at each interpolation point for subsequent analysis of radius of curvature and spatial orientation. For bonding points, the system extracts the largest connected highlight region from the corresponding Z-layer image, calculates its area, centroid coordinates, and contour perimeter, and performs shape matching with a standard bonding point template to reconstruct its three-dimensional convex morphology. Finally, all lead curves and bonding point models are integrated into a complete three-dimensional geometric model and stored in a specified format as required. In addition, the system extracts 30-by-30 pixel local image patches around each focal point from the image sequence and stitches them together in the tracking order to form a two-dimensional panoramic image. This image visually displays the surface condition of the entire lead, facilitating manual verification by the operator.

[0028] In the aforementioned high-precision 3D inspection method for semiconductor wire bonding based on a monocular visible light camera, step 6 involves comparing the reconstructed 3D model with pre-stored CAD design data item by item, calculating the actual lead height, radius of curvature, tilt angle, coplanarity deviation, bonding point size, and whether there is breakage or offset. When any parameter exceeds the set tolerance range, it is marked as a defect and an inspection report is generated. Specifically, the defect judgment threshold in step 6 is dynamically configured according to product specifications: the allowable deviation for lead height is ±15 micrometers, the radius of curvature must not be less than 80 micrometers, the coplanarity error does not exceed 20 micrometers, and a bonding point area change rate greater than ±25% is considered abnormal. All inspection results are summarized to generate a unit-level comprehensive inspection report, which can be exported to a standardized data format for MES system access. The defect detection and analysis module first loads a CAD model matching the current LeadFrame model. This model contains the theoretical 3D paths and bonding point specifications of all leads. The system performs rigid body registration between the measured 3D model and the CAD model, using the ICP algorithm to minimize the distance from the point to the model. After registration, each parameter is calculated: lead height is the maximum deviation of the measured curve in the vertical direction; radius of curvature is calculated by fitting a local circular arc segment; coplanarity deviation is the range of Z-coordinates of all bonding points on the same device; bonding point size variation rate is the relative deviation between the measured area and the theoretical area. All calculation results are compared with preset tolerance thresholds, and those exceeding the range are marked as defects, with the defect type, location, and severity recorded. Finally, the system generates a unit-level comprehensive inspection report, including the inspection status of all leads, a defect list, 3D model screenshots, and statistical summaries, and exports it in XML or JSON format for integration into the Manufacturing Execution System (MES).

[0029] In this embodiment, a high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera is described, such as... Figure 3 , 4 As shown, the contours of each lead are obtained by synthesizing local images from multiple images with different focal planes. Therefore, it can be determined that the lead tracking algorithm described above can perform accurate 3D tracking for each lead. The algorithm has proven accurate and effective in practical verification.

[0030] To further improve inspection efficiency, this invention also includes a multi-module parallel inspection architecture. The entire LeadFrame is divided into N x M independent inspection units (Units), each Unit further subdivided into several field-of-view regions. Multiple identical inspection modules synchronously execute the above process, achieving batch automated 3D inspection of the entire leadframe. A single complete inspection cycle is controlled within 90 seconds, meeting the cycle time requirements of online high-speed production lines. Each inspection module includes an independent imaging unit and a local processing unit, sharing a central scheduling system. The central scheduling system is responsible for task allocation, data aggregation, and global coordinate unification. Through this parallel architecture, the system significantly reduces the overall inspection time while maintaining single-field-of-view inspection accuracy.

[0031] This invention is applicable to automotive power semiconductor packaging production lines. It can stably detect fine-pitch lead layouts with a line spacing of less than 50 micrometers. It has good anti-interference capabilities for cross-line routing, dense arrangement and mirror-reflective gold lines. The axial height resolution reaches the 5-micrometer level, and the spatial repeatability positioning accuracy is better than ±0.8 micrometers. The whole system does not require multi-camera synchronous calibration, and the maintenance cost is significantly lower than that of traditional multi-view vision solutions.

[0032] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0033] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0034] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0035] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera, characterized in that, include: Step 1 Image Acquisition: Position the lead frame to be tested directly below the microscope head using the X-axis and Y-axis motion platform. Control the Z-axis drive device to drive the camera assembly to continuously autofocus and capture images between the starting height ZStart and the ending height ZEnd at a step interval of 5 micrometers to 10 micrometers, acquiring an image sequence covering the entire lead height range. Each image is accompanied by its corresponding Z-axis coordinate information and stored in the data buffer. Step 2 Image alignment processing: Select the middle frame as the reference frame based on the image sequence, and establish alignment feature points based on the reference frame. Perform mapping processing based on the alignment feature points of all images to be corrected to obtain the transformation matrix of the corresponding reference frame coordinate system. Use the transformation matrix to perform graphic alignment processing. Step 3: Focusing analysis: For each target lead within the field of view, the XY coordinates are preset within a certain range. The brightness distribution and edge sharpness of the corresponding area in the image sequence are analyzed layer by layer along the Z-axis. The local bright area with the strongest reflection on the lead surface and the sharpest contour edge is identified as the best focal point of the corresponding area. Its X, Y and corresponding Z-axis coordinates are recorded to form an initial three-dimensional point set. Step 4 Lead Path Tracking: Starting from the height group where the initial bonding point BondStart is located, based on the identified focal positions, search for subsequent focal points on the same lead path in the adjacent N frames. Set the search neighborhood to within ±20 micrometers of the current point's XY coordinates. Determine if there are candidate points that meet the following criteria: high brightness intensity threshold greater than 180 gray units, edge gradient modulus greater than 15 gray levels / pixel, and shape conforming to cylindrical projection characteristics. If found, update the current position and continue tracking until the height group where the termination bonding point BondEnd is located is reached and its geometry is successfully identified. Step 5: 3D model reconstruction: Summarize the XYZ coordinate data of all tracked lead wire focal points, use cubic spline interpolation algorithm to smoothly fit the point series, generate a continuous lead wire 3D spatial direction curve, and reconstruct the bonding point morphology based on the maximum bright area of ​​each bonding point and the 2D projection contour, and construct a complete 3D geometric model containing all leads and bonding points. Step 6: Defect Detection and Analysis: Compare the reconstructed 3D model with the pre-stored CAD design data item by item, calculate the actual lead height, radius of curvature, tilt angle, coplanarity deviation, bonding point size, and whether there is breakage or offset. When any parameter exceeds the set tolerance range, it is marked as a defect and an inspection report is generated.

2. The high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera according to claim 1, characterized in that: The objective lens used in step 1 is an ultra-narrow depth-of-field objective lens with a depth of field of less than 10 micrometers. It is used with a 10x magnification for gold wire detection or a 5x magnification for aluminum wire detection. The diameter of the gold wire ranges from 10 micrometers to 25 micrometers, and the height ranges from 500 micrometers to 1500 micrometers. The diameter of the aluminum wire is greater than 50 micrometers, and the height ranges from 500 micrometers to 2000 micrometers. The light source system consists of a ring of white or blue LEDs fixed around the lens. The illumination angle is adjustable, and a coaxial point light source is configured to enhance the imaging contrast of the bonding point area.

3. The high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera according to claim 1, characterized in that: The focus determination criteria in step 3 include maximizing local contrast, detecting peak response of the Laplacian operator, and evaluating the Tenengrad gradient function. A corresponding sliding window is set according to the lead wire direction for sequential tracking. When a sphere exists, the center of the sphere is designated as the initial point for sequential tracking. Within each field of view, multiple analysis units are divided according to a preset grid. The optimal focusing layer of the lead wire segment in each region is tracked, and the optimal Z-value and its confidence score at the center point of each unit are output.

4. The high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera according to claim 1, characterized in that: In step 4, the lead tracing starting point selection strategy prioritizes searching from bond points with lower Z coordinates. The system initiates initial identification below the BondStart height group. The number of image frames for initial identification is set according to the 3D shape of the lead, and the search range is limited to a rectangular area of ​​±50 micrometers within its preset XY coordinates. Once identification is successful, the iterative tracking process is initiated, with the search range narrowing to ±20 micrometers in each iteration. If no next focal point matching the conditions is found for three consecutive frames, it is determined to be a broken or severely bent failure.

5. The high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera according to claim 1, characterized in that: In step 5, the three-dimensional curve fitting adopts the weighted least squares method, giving higher weight coefficients to high-confidence focal points, and the fitting residual is controlled within ±2 micrometers; at the same time, a two-dimensional panoramic image is synthesized based on the local image blocks around each focal point in the image sequence, which unfolds along the lead line to assist in manual review and defect classification.

6. The high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera according to claim 1, characterized in that: In step 6, the defect judgment threshold is dynamically configured according to the product specifications. The allowable deviation of the lead height is ±15 micrometers, the radius of curvature must not be less than 80 micrometers, the coplanarity error does not exceed the process threshold, and the bonding point area change rate is greater than ±25% as abnormal. All test results are summarized to generate a unit-level comprehensive test report, which can be exported to a standardized data format for MES system access. The process threshold is the chip feature threshold, which is 20-50 micrometers.

7. The high-precision 3D detection method for semiconductor wire bonding based on a monocular visible light camera according to claim 1, characterized in that: It also includes a multi-module parallel detection architecture, which divides the entire lead frame into N multiplied by M independent detection units. Each unit is further subdivided into several field of view areas, and multiple identical detection modules execute the detection process synchronously to achieve batch automated 3D detection of the entire lead frame. The single complete detection cycle is controlled to 10-15 seconds.

Images