Visual inspection method for gold wire bonding defects of optoelectronic devices

Abstract

The present application belongs to the technical field of photoelectric packaging, and particularly relates to a visual detection method for gold wire bonding defects of photoelectric devices, which comprises performing filter pretreatment and local contrast enhancement on the collected original image, extracting a skeleton center line to fit a wire arc trajectory, positioning a circle center coordinate and an equivalent radius to demarcate a neck region; in combination with a fixed illumination light source incident direction, a solder ball geometric self-shading prediction model is constructed to determine a self-shading coverage sub-region, a differential response graph is constructed by comparing gray values to mark a neck crack candidate area; a compensation illumination angle is calculated to collect a compensation illumination frame, and real neck defects are confirmed by comparing gray responses; the original fixed illumination frame and the compensation illumination frame are superimposed to construct a double-channel feature graph, and the defect type is excluded in combination with the wire arc trajectory and output. The present application effectively solves the problem of neck crack missed detection under the fixed angle illumination architecture, and significantly reduces the false positive rate.

Description

Technical Field

[0001] This invention belongs to the field of optoelectronic packaging technology, specifically relating to a visual inspection method for defects in gold wire bonding of optoelectronic devices. Background Technology

[0002] In the field of optical communication device packaging and manufacturing, gold wire bonding is a core process connecting chip pads to substrate pins. Optoelectronic packaging products, such as high-speed optical transceivers like BOSA and TOSA, rely on ultra-fine gold wires with diameters in the range of 18 to 25 micrometers for electrical interconnection between internal laser chips and active components such as transimpedance amplifiers and photodetectors. Bonding quality directly determines the high-frequency electrical performance and long-term reliability of the device, making it a crucial process for quality control in optoelectronic packaging.

[0003] Patent application CN116087209A discloses a method and system for visual inspection of gold wire bonding processes. This method acquires a reference image of the gold wire and multiple background images using different light source modes. In the reference image, a dividing line is set based on the contrast between the gold wire and the background object. The gold wire in the reference image is divided into multiple segments according to the dividing line. For each segment, the segment with the highest contrast is selected from the background and reference images. The selected segments are then spliced ​​together to obtain the gold wire to be tested that meets the contrast requirements. Defects such as spacing, bending, crossing, and breakage are detected in the gold wire to be tested according to the set inspection requirements. This invention achieves fully automated machine vision inspection of optical communication products through flexible light source configuration and inspection methods. It features high inspection adaptability and strong anti-interference capabilities, good inspection stability and product detection rate, and realizes automated inspection of gold wire bonding processes.

[0004] For defect detection in gold wire bonding, existing automated vision inspection solutions in the industry generally adopt a fixed-angle coaxial illumination combined with a high-resolution industrial camera acquisition architecture, and use threshold segmentation or convolutional feature extraction to judge the integrity of the wire arc contour. To further improve the recognition rate of macroscopic defects such as broken wires, missing wires, and poor solder joints, some systems have introduced multi-scale feature map extraction networks, which have achieved good results in the identification of fractures in the middle section of the wire.

[0005] Because the neck region between the gold wire and the first solder joint (i.e., the ball-on-chip solder joint on the chip pad side) is a zone of rapid deformation where the gold wire transitions from a spherical solder ball geometry to a cylindrical wire, the surface normal direction within this region rotates nearly 90° within a span of less than two wire diameters (approximately 40 micrometers). Under fixed-angle illumination conditions, the solder ball itself geometrically blocks the incident light, resulting in a dark crescent-shaped shadow band at the root of the neck that is strongly correlated with the illumination angle and whose position fluctuates with the wire arc angle. This self-shadowing area, determined by the solder ball's own geometry, cannot be eliminated under any single fixed illumination angle. Neck cracks, as the most critical defect with the highest failure probability in gold wire bonding of optoelectronic devices, have fracture surface image features that highly overlap with this self-shadowing band in terms of grayscale distribution, spatial location, and morphology. This phenomenon makes it impossible for existing detection systems to distinguish between the solder ball's geometric self-shadowing and the neck material fracture lines from a single frame image, leading to a systematic blind spot for detecting this critical defect. Summary of the Invention

[0006] This invention provides a visual inspection method for defects in gold wire bonding of optoelectronic devices, in order to solve the technical problem of blind spots in the neck region caused by fixed-angle illumination in the prior art.

[0007] This invention provides a visual inspection method for defects in gold wire bonding in optoelectronic devices, comprising the following steps: S1. Perform filtering preprocessing and local contrast enhancement on the original image of the gold wire bonding area, extract the skeleton center line of the gold wire to fit the wire arc trajectory, and locate the center coordinates and equivalent radius of the solder ball to obtain the enhanced image, thereby defining the spatial range of the neck region. S2, combined with the pre-calibrated fixed illumination source incident direction, construct a geometric self-shadowing prediction model for the solder ball in the neck region based on the center coordinates and equivalent radius to determine the self-shadowing coverage sub-region, and construct a differential response map by comparing the actual pixel grayscale value with the theoretical shadow grayscale reference value in the self-shadowing coverage sub-region, thereby marking the candidate area of ​​the neck crack; S3, in response to the neck crack candidate area being non-empty, calculate the compensation illumination angle to acquire the compensation illumination frame, and compare the grayscale response of the corresponding neck crack candidate area in the compensation illumination frame with the normal filament brightness reference to confirm whether the neck crack candidate area is a real neck defect. S4. For areas confirmed as true neck defects, the original fixed illumination frame and the compensated illumination frame are superimposed to construct a dual-channel feature map. The defect type is then checked on the dual-channel feature map according to priority and mutual exclusion based on the silk arc trajectory. The determination result of the defect type and its coordinate position are then output.

[0008] Its effect is that by constructing an analytical self-shadowing prediction model based on the spherical geometry of the solder ball, the self-shadowing coverage sub-region is accurately delineated, and the illumination geometric occlusion and material physical fracture are clearly distinguished at the imaging level, fundamentally eliminating the risk of missing neck cracks under a fixed-angle illumination architecture.

[0009] Furthermore, filtering preprocessing and local contrast enhancement are performed on the acquired original image of the gold wire bonding region, including: Anisotropic diffusion filtering preprocessing is performed on the original image; The enhancement coefficient is calculated based on the local gray-level dynamic range of the gold wire candidate area in the neighborhood of each pixel. Adaptive local contrast enhancement is performed on the filtered image using enhancement coefficients to obtain the final enhanced image.

[0010] Its effects are as follows: the original image is preprocessed using anisotropic diffusion filtering combined with adaptive local contrast enhancement. The filtering can smooth the uneven brightness fluctuations of the substrate background and completely preserve the gradient of the gold wire edge. The adaptive enhancement adjusts the enhancement coefficient according to the local grayscale dynamic range, avoiding grayscale saturation in the bright areas of the solder ball and loss of wire details, improving the accuracy of subsequent wire arc skeleton extraction and solder ball positioning, eliminating the chain detection errors caused by background interference, and providing a high-quality enhanced image base for subsequent shadow modeling and defect interpretation.

[0011] Furthermore, the centerline of the gold wire skeleton is extracted to fit the wire arc trajectory, and the center coordinates and equivalent radius of the solder ball are located, thereby defining the spatial range of the neck region, including: The ridge detection operator is used to extract the skeleton center line of the gold wire on the enhanced image in order to fit the complete wire arc trajectory; The center coordinates and equivalent radius of the positioning solder ball are determined by circular transformation; Using the center coordinates of the circle as the center and the equivalent radius to a preset multiple of the equivalent radius as the radial boundary, the spatial range of the neck region is finally defined.

[0012] Furthermore, combining the pre-calibrated incident direction of the fixed illumination source, a geometric self-shadowing prediction model for the solder ball geometry in the neck region is constructed based on the center coordinates and equivalent radius to determine the self-shadowing coverage sub-region, including: Using the center coordinates and equivalent radius of the circle, combined with the incident direction of a fixed lighting source, a geometric self-shadowing prediction model for solder balls is constructed analytically. The theoretical boundary curve of self-shading is determined based on the tangent condition of the spherical shading cone; The intersection of the theoretical boundary curve and the neck region is ultimately determined as the self-shadowed sub-region.

[0013] Furthermore, a differential response map is constructed by comparing the actual pixel grayscale values ​​with the theoretical shadow grayscale reference values ​​within the self-shadowed sub-region, thereby marking the candidate areas for neck cracks, including: Within the self-shadow-covered sub-region, the actual pixel grayscale value of the current frame image is subtracted from the theoretical shadow grayscale reference value to finally construct the differential response map; A threshold for filtering connected dark regions is set based on online statistical accumulation of defect-free samples; Connected regions in the differential response map that have negative values ​​and whose absolute values ​​exceed the threshold are ultimately marked as candidate regions for neck cracks.

[0014] Its effect is that by setting a threshold to filter connected dark areas, it effectively eliminates the interference of gray-scale disturbances introduced by non-structural factors such as random contamination on the solder ball surface and local reflectivity fluctuations in the oxide layer on the judgment results, making the construction of the differential response map more accurate and reliable.

[0015] Further, the compensated illumination angle is calculated to acquire compensated illumination frames, including: Starting from the center coordinates of the candidate area for neck cracks, determine the ray direction that satisfies the unobstructed condition; Among all the available angles that satisfy the unobstructed condition, the mirror direction of the incident direction of the fixed lighting source about the pad plane is selected as the compensation lighting angle; The corresponding quadrant light-emitting unit group is activated according to the compensation lighting angle to perform directional supplementary lighting, and finally the compensation lighting frame is acquired.

[0016] Furthermore, the grayscale response of the corresponding neck crack candidate area in the compensated illumination frame is compared with the normal filament brightness reference to confirm whether the neck crack candidate area is a real neck defect, including: The median pixel grayscale value of the main silk segment is extracted from historical defect-free frames and used as the normal silk brightness benchmark. If the grayscale response of the candidate area of ​​the neck crack in the compensated lighting frame is greater than or equal to the normal filament brightness reference, it is determined to be a false dark filament and removed. If the grayscale response is less than the normal filament brightness reference, then the candidate area for the neck crack is finally confirmed as a real neck defect.

[0017] Its effect is that by verifying the secondary brightness of the directional compensation lighting frame, the final determination of the neck region is based on the differential evidence of the two frames, avoiding the fragile judgment that relies on a single grayscale threshold, thereby significantly reducing the false alarm rate of neck defects.

[0018] Furthermore, the original fixed lighting frame and the compensated lighting frame are superimposed to construct a dual-channel feature map. Based on the silk arc trajectory, defect types are then identified on the dual-channel feature map according to priority and mutual exclusion, including: The original fixed illumination frame and the compensated illumination frame are superimposed pixel by pixel to finally construct a dual-channel feature map; On the dual-channel feature map, determine whether the real neck defect presents a transverse dark pattern along the tangent direction of the silk arc trajectory, and whether it continuously presents a dark response in both channels; If the conditions are met, the defect type is ultimately determined to be a neck crack.

[0019] Its effect is that by using the criteria for defects such as neck cracks to be checked in a mutually exclusive manner on the dual-channel feature map according to priority, and each type of defect corresponds to a clear geometric morphology criterion, the entire process of judging gold wire bonding defects in optoelectronic packaging production lines from candidate marking to type confirmation is realized.

[0020] Furthermore, combining the wire arc trajectory with the dual-channel feature map to identify defect types by priority and mutual exclusion, it also includes: When the criteria for determining a neck crack are not met, determine whether a diffuse dark area appears in the neck region of the dual-channel feature map. Detect whether the position of the silk arc trajectory changes at the neck endpoint and whether the endpoint coordinates deviate from the projection point of the center coordinates by more than a preset distance; If both of the above conditions are met, the final defect type is determined to be solder ball peeling.

[0021] Furthermore, combining the wire arc trajectory with the dual-channel feature map to identify defect types by priority and mutual exclusion, it also includes: When the criteria for determining neck cracks and solder ball peeling are not met, determine whether the brightness of the neck region in both channels of the dual-channel feature map is weaker than the brightness of the main wire region. Calculate the full width at half maximum (FWHM) of the neck filament and determine whether the FWHM is less than the threshold derived from the statistical mean and nominal filament diameter of the same batch. If the conditions are met, the final defect type is determined to be an excessively thin neck.

[0022] The beneficial effects are: This invention breaks through the limitations of traditional single fixed-angle illumination in detecting complex deformation areas. By utilizing the spherical geometry of solder balls to construct an analytical self-shadowing prediction model, it clearly distinguishes between illumination geometric occlusion and material physical fracture at the imaging level, eliminating the risk of missed detection of neck cracks under fixed-angle illumination architecture and significantly improving the reliability of optoelectronic device packaging quality control.

[0023] Furthermore, this invention introduces a compensated illumination verification process and a multi-angle feature map overlay interpretation method, establishing the final determination of the neck region based on dual-frame differential evidence. By combining mutually exclusive classification criteria and adaptive geometric parameters for the same batch, accurate classification and structured recording of various defects are achieved, providing complete and reproducible data support for subsequent process optimization and failure analysis, effectively ensuring the high-frequency electrical performance and long-term stability of optical communication devices. Attached Figure Description

[0024] Figure 1 This is a flowchart of a visual inspection method for defects in gold wire bonding in optoelectronic devices.

[0025] Figure 2 This is a close-up view of the detection effect in the gold wire bonding region. Detailed Implementation

[0026] 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, not all, of the embodiments of the present invention. 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.

[0027] An embodiment of the visual inspection method for gold wire bonding defects in optoelectronic devices provided by the present invention: like Figure 1 As shown, a visual inspection method for gold wire bonding defects in optoelectronic devices includes the following steps: S1, perform filtering preprocessing and delineate the neck region.

[0028] Anisotropic diffusion filtering preprocessing is performed on the acquired original image of the gold wire bonding area. Due to the significant difference in the reflection characteristics of incident light between the optoelectronic packaging substrate and the pad surface, the background brightness of the gold wire bonding area exhibits a significant non-uniform distribution in space. If edge extraction is directly performed on the original image, the background brightness gradient will overlap with the wire edge gradient, severely interfering with the positioning accuracy of the subsequent wire arc skeleton. Anisotropic diffusion filtering smooths the low-frequency brightness fluctuations in the background while selectively preserving the high-frequency gradient information at the edge of the gold wire, thereby effectively suppressing non-uniform background interference without blurring the wire arc contour.

[0029] Next, adaptive local contrast enhancement is performed on the filtered image. The enhancement coefficient is adaptively calculated based on the local grayscale dynamic range of the gold wire candidate area in the neighborhood of each pixel. Since the solder ball area itself has high-brightness reflections, if the same contrast enhancement coefficient is applied uniformly to the entire image, the bright areas of the solder balls will be overstretched, resulting in grayscale saturation at the edges of the solder balls and loss of detail information in pixels adjacent to the neck. By adopting a localized calculation method for the adaptive enhancement coefficient, the enhancement intensity is stronger in the low-brightness areas of the main wire and automatically converges in the bright areas of the solder balls. Thus, in a single enhancement operation, both the visibility of wire arc details and the preservation of grayscale levels in the solder ball area are taken into account, resulting in an enhanced image.

[0030] Furthermore, a ridge detection operator based on the principal curvature response of the Hessian matrix is ​​used to extract the skeleton centerline of the gold wire on the enhanced image, in order to fit the complete arc trajectory of the wire. ,in The arc length parameter is used. The ratio of the two principal curvatures of the Hessian matrix exhibits anisotropic characteristics at the wire-like structure, which are significantly different from the circular structure of the solder ball. Using this as the basis for the ridge response can effectively distinguish the wire arc skeleton from the solder ball contour in the same image, avoiding the solder ball arc segment being mistakenly included in the wire arc skeleton fitting. For each gold wire, the center coordinates of the solder ball are simultaneously located using the Hough circle transformation. and equivalent radius .

[0031] The spatial extent of the neck region is defined as follows: With the center, and To the outer boundary coefficient and The product of the products forms a ring-shaped region of interest with radial boundaries, denoted as . The outer boundary coefficient is determined based on the physical extension range of the neck transition region in the same batch of bonding processes. For example, the outer boundary coefficient can be 1.8, i.e. The radial outer boundary is Extracted Coordinates along with the centerline data of the arc This will serve as the geometric reference input for all subsequent steps.

[0032] It should be noted that there are many operators in this field capable of extracting the silk arc skeleton. Those skilled in the art can choose to replace the Hessian ridge operator based on gradient direction histograms or other curve fitting methods according to the actual hardware computing power and detection accuracy requirements, as long as the complete silk arc trajectory can be output with sub-pixel accuracy. Both the geometry of the solder balls and the parameters thereof are within the scope of protection of this invention.

[0033] By employing a combination of anisotropic diffusion filtering and adaptive local contrast enhancement preprocessing, along with Hessian ridge detection and Hough circle transform, wire arc skeleton fitting and solder ball geometric positioning were simultaneously achieved on a single frame of acquired image, driven by the physical extension range of the neck transition zone. The delineation method precisely locks the spatial range of all subsequent calculations within the neck deformation influence area, establishing a unique and definite geometric reference benchmark for subsequent self-shadowing modeling and defect interpretation, thus avoiding the chain error propagation caused by the neck ROI delineation error leading to the offset of shadow boundary calculation.

[0034] S2, construct a self-shading prediction model and label candidate areas for neck cracks.

[0035] Using the already anchored solder ball center coordinates With equivalent radius Combined with the pre-calibrated incident direction of the fixed lighting source Analytical construction targeting the neck region The self-shadowing prediction model for solder balls is described. The imaging coordinates of the self-shadowing boundary are determined by the tangent conditions of the spherical occlusion cone, i.e., the set of spherical points satisfying the following two constraints. The orthographic projection onto the imaging plane constitutes the theoretical boundary curve of the shadow. :

[0036] The geometric meaning of this condition is: the set of points where the incident ray is exactly tangent to the sphere constitutes the boundary of the illumination cone, and the projection of this boundary onto the imaging plane is the outer contour of the shadow. and The intersection of these regions is the self-shadowed sub-region. The rest The visible area of ​​the neck that can be properly exposed under fixed lighting.

[0037] Furthermore, in the self-shadowing sub-region The actual pixel grayscale value of the current frame image and the theoretical shadow grayscale reference value. Perform pixel-by-pixel comparison to construct a differential response map:

[0038] in For pixel coordinates, The actual grayscale value of the current frame to be detected at pixel position x, and the theoretical shadow grayscale reference value. Extracted from the calibration frames of the system on defect-free standard samples, specifically... The median pixel grayscale value of all defect-free historical frames within the region. Due to factors such as random contamination of the solder ball surface and localized fluctuations in the reflectivity of the oxide layer, [this value may vary]. Unstructured grayscale perturbations are introduced, and the spatial scale of these perturbations is typically much smaller than the pixel range covered by the neck crack fracture surface. Therefore, a threshold is set... Filtering is applied to connected dark areas to effectively eliminate the aforementioned noise interference. Negative values ​​and absolute values ​​exceeding Connected regions are marked as candidate areas for neck cracks. .

[0039] threshold The determination is based on the online statistical accumulation of defect-free samples. Let... The total number of pixels is ,but Take a defect-free sample The cumulative probability in the distribution reaches The corresponding quantile value. A larger value indicates a wider range of grayscale fluctuations in the defect-free sample, corresponding to more significant aging of the production line light source or environmental changes. The value is then adjusted upwards; a smaller value indicates that the system is in a stable lighting and clean environment, and the grayscale disturbance fluctuation is narrowed. This tightening further improves detection sensitivity. With the online rolling updates of the defect-free sample pool, It continuously adapts and corrects itself, always maintaining consistency with the actual lighting conditions on the production line.

[0040] By combining analytical shading cone calculation of solder ball spherical geometry with online statistics of defect-free samples, each pixel in the neck region of interest is clearly assigned to either a geometric self-shading coverage area or a normal exposure area at the imaging level. The construction method of the differential response map ensures that the labeling of the neck crack candidate area is entirely based on the numerical deviation from the defect-free baseline, eliminating the systematic blind zone caused by the overlap of neck self-shading and crack dark features under a fixed lighting architecture.

[0041] S3, acquire compensated lighting frames to confirm the actual neck defect.

[0042] Response to neck crack candidate region If the result is not empty, the system enters the compensation lighting acquisition process, calculates the compensation lighting angle to acquire compensation lighting frames, and then... Perform secondary directional lighting verification to physically distinguish between false alarms caused by solder ball geometry self-shadowing and genuine neck cracks.

[0043] Compensation lighting angle The calculation principle is: from the candidate area of ​​neck cracks center coordinates Depart, along Rays in the direction and with For the center of the ball, A sphere with radius r has no intersection points, thus satisfying the unobstructed condition:

[0044] in, The ray length is a scalar parameter; among all available angles that satisfy the unobstructed condition, the incident direction of the fixed illumination source is preferred. The mirror direction of the pad plane is used as a compensating lighting angle. ,Right now:

[0045] in The normal vector of the pad plane is selected based on the fact that the incident angles of the mirror illumination and the original illumination on the main body region of the filament are complementary, and the overall brightness distribution of the main body of the filament is the closest in the two frames. This eliminates the influence of the difference in reflectivity of the filament itself to the greatest extent during subsequent two-frame differential comparison, so that the brightness comparison is completely focused on the physical state of the candidate dark area of ​​the neck itself, avoiding being masked by the overall brightness deviation of the filament.

[0046] Next, the system's multi-angle programmable ring LED array is used to compensate for the lighting angle according to the calculated values. Activate the corresponding quadrant's light-emitting unit group to complete the operation without moving the device. Directional supplemental lighting for the area, and acquisition of compensation lighting frames. The determination of compensated illumination verification is based on the median pixel grayscale value of the filament backbone segment in historical defect-free frames. This standard serves as a reference for normal silk brightness and is extracted from the corresponding silk arc length parameter. The middle segment, located within a predetermined proportional range in the middle of the total arc length, is used to avoid interference from the neck's self-shaded area on the reference value. For example, this proportional range can be 20% to 80% of the total arc length.

[0047] Because the fracture surface of the neck crack does not physically contain a gold filament continuum, compensating illumination, even when incident from an unobstructed angle, cannot produce filament-scattered brightness at the fracture location. Therefore, this area is within the compensating illumination frame. The middle section will continue to exhibit a dark response; however, the pseudo-dark patterns caused by the solder ball's self-shadowing will, under compensated lighting, cause the shielding cone to no longer cover the area, allowing the neck wire continuum to normally scatter the incident light, and the grayscale response will return to the normal wire brightness level. Based on this, the grayscale response of the corresponding neck crack candidate area in the compensated lighting frame is compared with the normal wire brightness benchmark. If the compensated lighting frame... middle grayscale value of the region satisfy If the brightness is greater than or equal to the normal wire brightness reference, then the dark area is determined to be a pseudo-dark pattern caused by solder ball self-shading, and it is removed from the candidate defect list; if If the condition persists under compensated lighting (i.e., the grayscale response is less than the normal filament brightness reference), then it is confirmed that there is no physical gold wire continuum in that area, confirming the neck crack candidate area as a genuine neck defect, and proceeding to the subsequent morphological classification process. (Normal filament brightness reference) The sample pool of defect-free samples is continuously updated online, keeping in line with the actual lighting conditions on the production line.

[0048] By using directional compensation illumination driven by the mirror direction of the pad plane, the discrimination of neck dark areas is upgraded from single-frame grayscale threshold comparison to dual-frame physical brightness difference verification. Without moving the device or increasing the mechanical movement time, the geometric occlusion dark area and the material fracture dark area are distinguished by the complementary perspective of the compensation illumination frame and the fixed illumination frame. This effectively suppresses the false alarm rate of neck defects to a controllable level determined by the normal filament brightness benchmark.

[0049] S4, construct dual-channel feature maps to identify mutually exclusive defect types.

[0050] For areas confirmed as genuine neck defects, the extracted complete silk arc trajectory is used as a basis. This leads to the final defect morphology classification process. The original fixed lighting frame... With compensation lighting frame Pixel-by-pixel stacking is performed to construct a dual-channel feature map:

[0051] in, This represents a dual-channel pixel feature vector. Represents a fixed lighting frame The gray value at pixel x Indicates the compensation lighting frame The grayscale value at the same pixel x.

[0052] This feature map simultaneously carries texture and morphological information from two complementary perspectives, serving as input for defect classification. Defect types are then prioritized and mutually exclusive on the dual-channel feature map based on the silk arc trajectory. For the morphological differences of the three main types of neck defects, the criteria are used to prioritize and mutually exclusively eliminate defects sequentially.

[0053] The first type involves identifying neck cracks. The fracture surface of a neck crack appears on the imaging plane as a transverse dark line along the tangent of the wire arc trajectory. The extension direction of this dark line is perpendicular to the tangent of the wire arc trajectory. Because the crack fracture surface physically interrupts the gold wire continuum, there is no gold wire scattering brightness at this fracture location under both fixed and compensated illumination perspectives. and Two channels in All regions exhibit a persistent dark response. When both the transverse dark crack morphology and the dual-channel persistent dark response are simultaneously met, the defect type is determined to be a neck crack, and the defect type determination result and coordinate location are output.

[0054] The second category, solder ball peeling, is only considered when the neck crack criterion is not met. During solder ball peeling, the interface between the solder ball and the pad separates, and the gold wires in the neck region shift away from the solder ball as a whole, as shown in the dual-channel feature map. neck area The interior shows a diffuse dark area, and the detection wire arc trajectory is also observed. Does a positional jump occur at the neck endpoint, i.e., does the endpoint coordinate deviate from the center coordinate? The projected point exceeds a preset offset threshold determined based on the equivalent radius of the solder ball. For example, the preset offset threshold can be set to... If both of the above conditions are met, the final defect type is determined to be solder ball peeling.

[0055] The third category, the determination of an excessively thin neck, only applies when the criteria for determining neck cracks and solder ball peeling are not met. This criterion involves determining whether the brightness of the neck region in both channels of the dual-channel feature map is weaker than the brightness of the main filament region; and calculating the full width at half maximum (FWHM) of the neck filament. And determine whether the full width at half maximum (FWHM) is less than the threshold derived from the statistical mean and nominal wire diameter of the same batch, specifically satisfying:

[0056] in Pre-set proportional range for the middle of the arc length of the main section of the same batch of silk. For example, the statistical mean can be taken as 20% to 80% of the total arc length; The nominal wire diameter of the gold wire in the same batch and The derivation of the ratio, i.e. The smaller this ratio, the more significant the halo expansion effect of the imaging system, and the greater the magnification of the imaging width relative to the nominal filament diameter. Consequently, the threshold for determining an excessively thin neck is relaxed. Conversely, the closer this ratio is to 1, the more closely the imaging width matches the nominal filament diameter, thus tightening the threshold. The entire derivation process is based entirely on the geometric characteristics of products within the same batch. If the conditions are met, the final defect type is determined to be an excessively thin neck.

[0057] Furthermore, for non-neck defects in the main body segment of the wire arc, the original fixed lighting frame is directly used. Based on the silk arc trajectory The continuity of the skeleton is judged: if the length of the interruption exceeds a preset interruption threshold, it is judged as a broken wire. For example, the interruption threshold can be 3 wire diameters. If the distance between the nearest points of the skeleton curves of adjacent wire arcs in space is less than a preset spacing threshold, it is judged as a short circuit risk between wires. For example, the spacing threshold can be 1 wire diameter. The system records the judgment results of all defect types for each device and their image coordinate positions, providing complete data support for subsequent process optimization and failure analysis.

[0058] Thus, by combining mutually exclusive classification criteria on dual-channel feature maps with adaptive geometric parameters of the same batch, the determination of three types of defects—neck cracks, solder ball peeling, and excessively thin necks—is entirely based on reproducible morphological geometric evidence, realizing the full-process interpretation of gold wire bonding defects in optoelectronic packaging production lines from candidate marking to type confirmation.

[0059] like Figure 2 As shown, the macroscopic photography style presents the physical details of the ultra-fine gold wire arc on the device substrate. The annotation box marks two typical defect areas: neck crack and excessively thin neck, which intuitively demonstrates the defect localization effect of this method on real packaging production line images.

[0060] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A visual inspection method for gold wire bonding defects in optoelectronic devices, characterized in that, Includes the following steps: S1. Perform filtering preprocessing and local contrast enhancement on the original image of the gold wire bonding area, extract the skeleton center line of the gold wire to fit the wire arc trajectory, and locate the center coordinates and equivalent radius of the solder ball to obtain the enhanced image, thereby defining the spatial range of the neck region. S2, combined with the pre-calibrated fixed illumination source incident direction, construct a geometric self-shadowing prediction model for the solder ball in the neck region based on the center coordinates and equivalent radius to determine the self-shadowing coverage sub-region, and construct a differential response map by comparing the actual pixel grayscale value with the theoretical shadow grayscale reference value in the self-shadowing coverage sub-region, thereby marking the candidate area of ​​the neck crack; S3, in response to the neck crack candidate area being non-empty, calculate the compensation illumination angle to acquire the compensation illumination frame, and compare the grayscale response of the corresponding neck crack candidate area in the compensation illumination frame with the normal filament brightness reference to confirm whether the neck crack candidate area is a real neck defect. S4. For areas confirmed as true neck defects, the original fixed illumination frame and the compensated illumination frame are superimposed to construct a dual-channel feature map. The defect type is then checked on the dual-channel feature map according to priority and mutual exclusion based on the silk arc trajectory. The determination result of the defect type and its coordinate position are then output.

2. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 1, characterized in that, The acquired raw images of the gold wire bonding region were subjected to filtering preprocessing and local contrast enhancement, including: Anisotropic diffusion filtering preprocessing is performed on the original image; The enhancement coefficient is calculated based on the local gray-level dynamic range of the gold wire candidate area in the neighborhood of each pixel. Adaptive local contrast enhancement is performed on the filtered image using enhancement coefficients to obtain the final enhanced image.

3. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 1, characterized in that, The centerline of the gold wire skeleton is extracted to fit the wire arc trajectory, and the center coordinates and equivalent radius of the solder ball are located, thereby defining the spatial range of the neck region, including: The ridge detection operator is used to extract the skeleton center line of the gold wire on the enhanced image in order to fit the complete wire arc trajectory; The center coordinates and equivalent radius of the positioning solder ball are determined by circular transformation; Using the center coordinates of the circle as the center and the equivalent radius to a preset multiple of the equivalent radius as the radial boundary, the spatial range of the neck region is finally defined.

4. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 1, characterized in that, Based on the pre-calibrated incident direction of a fixed illumination source, a geometric self-shadowing prediction model for the solder ball geometry in the neck region is constructed using the center coordinates and equivalent radius to determine the self-shadowing coverage sub-region, including: Using the center coordinates and equivalent radius of the circle, combined with the incident direction of a fixed lighting source, a geometric self-shadowing prediction model for solder balls is constructed analytically. The theoretical boundary curve of self-shading is determined based on the tangent condition of the spherical shading cone; The intersection of the theoretical boundary curve and the neck region is ultimately determined as the self-shadowed sub-region.

5. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 1, characterized in that, A differential response map is constructed by comparing the actual pixel grayscale values ​​with the theoretical shadow grayscale reference values ​​within the self-shadowed sub-region, thereby marking the candidate areas for neck cracks, including: Within the self-shadow-covered sub-region, the actual pixel grayscale value of the current frame image is subtracted from the theoretical shadow grayscale reference value to finally construct the differential response map; A threshold for filtering connected dark regions is set based on online statistical accumulation of defect-free samples; Connected regions in the differential response map that have negative values ​​and whose absolute values ​​exceed the threshold are ultimately marked as candidate regions for neck cracks.

6. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 1, characterized in that, Calculate the compensated illumination angle to acquire compensated illumination frames, including: Starting from the center coordinates of the candidate area for neck cracks, determine the ray direction that satisfies the unobstructed condition; Among all the available angles that satisfy the unobstructed condition, the mirror direction of the incident direction of the fixed lighting source about the pad plane is selected as the compensation lighting angle; The corresponding quadrant light-emitting unit group is activated according to the compensation lighting angle to perform directional supplementary lighting, and finally the compensation lighting frame is acquired.

7. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 1, characterized in that, The grayscale response of the corresponding neck crack candidate area in the compensated illumination frame is compared with the normal filament brightness reference to confirm whether the neck crack candidate area is a real neck defect, including: The median pixel grayscale value of the main silk segment is extracted from historical defect-free frames and used as the normal silk brightness benchmark. If the grayscale response of the candidate area of ​​the neck crack in the compensated lighting frame is greater than or equal to the normal filament brightness reference, it is determined to be a false dark filament and removed. If the grayscale response is less than the normal filament brightness reference, then the candidate area for the neck crack is finally confirmed as a real neck defect.

8. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 1, characterized in that, The original fixed lighting frame and the compensated lighting frame are superimposed to construct a dual-channel feature map. Based on the wire arc trajectory, defect types are then identified on the dual-channel feature map according to priority and mutual exclusion, including: The original fixed illumination frame and the compensated illumination frame are superimposed pixel by pixel to finally construct a dual-channel feature map; On the dual-channel feature map, determine whether the real neck defect presents a transverse dark pattern along the tangent direction of the silk arc trajectory, and whether it continuously presents a dark response in both channels; If the conditions are met, the defect type is ultimately determined to be a neck crack.

9. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 8, characterized in that, Combining the wire arc trajectory with the dual-channel feature map to identify defect types by priority and mutual exclusion, it also includes: When the criteria for determining a neck crack are not met, determine whether a diffuse dark area appears in the neck region of the dual-channel feature map. Detect whether the position of the silk arc trajectory changes at the neck endpoint and whether the endpoint coordinates deviate from the projection point of the center coordinates by more than a preset distance; If both of the above conditions are met, the final defect type is determined to be solder ball peeling.

10. The visual inspection method for gold wire bonding defects in optoelectronic devices according to claim 9, characterized in that, Combining the wire arc trajectory with the dual-channel feature map to identify defect types by priority and mutual exclusion, it also includes: When the criteria for determining neck cracks and solder ball peeling are not met, determine whether the brightness of the neck region in both channels of the dual-channel feature map is weaker than the brightness of the main wire region. Calculate the full width at half maximum (FWHM) of the neck filament and determine whether the FWHM is less than the threshold derived from the statistical mean and nominal filament diameter of the same batch. If the conditions are met, the final defect type is determined to be an excessively thin neck.

Images