CN-121708017-B - Shape self-adaption and multidimensional feature fusion-based pad needle mark detection method

CN121708017BCN 121708017 BCN121708017 BCN 121708017BCN-121708017-B

Abstract

The invention discloses a shape self-adaption and multidimensional feature fusion-based pad needle mark detection method, which comprises the steps of obtaining pad image data, judging whether a pad type is strip-shaped, square or round according to an aspect ratio and a circularity factor of a pad in the pad image data, respectively carrying out edge partition and multi-scale self-adaption judgment on the pad in the pad image data according to the pad type, carrying out non-maximum suppression strategy on a plurality of candidate needle mark targets and texture clustering operation based on space collinearity, eliminating redundant needle mark targets to obtain a plurality of final needle mark targets, converting pixel coordinates of the plurality of final needle mark targets into normalized center coordinates through coordinate inverse mapping and format standardization, and drawing a mark frame in the pad image data according to the normalized center coordinates corresponding to the plurality of final needle mark targets to mark positions of the plurality of final needle mark targets. And (3) self-adapting to the shape of the bonding pad, inhibiting texture interference and separating edge noise, and outputting bonding pad image data with a needle mark frame.

Inventors

LI SONGBIN
XIAO YE
GAO MIN
CHEN JIACHENG
LIU HONGJUN

Assignees

无锡学院
苏州芯慧联半导体科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260212

Claims (9)

1. A method for detecting a needle mark of a bonding pad based on shape self-adaption and multidimensional feature fusion is characterized by comprising the following steps: S1, acquiring pad image data, constructing a geometric moment self-adaptive classifier, and judging whether the type of the pad is rectangular, square or round according to the aspect ratio and the circularity factor of the pad in the pad image data; S2, respectively carrying out edge partition and multi-scale self-adaptive judgment on the bonding pad in the bonding pad image data according to the bonding pad type, when the bonding pad type is long-strip-shaped, equalizing local illumination of the long-strip-shaped bonding pad by adopting a partition detection strategy of an overlapped sliding window, carrying out needle mark detection by a built three-channel candidate extraction model based on local statistical characteristics, and extracting a plurality of initial candidate areas; When the type of the bonding pad is square, the local self-adaptive threshold value and gradient direction verification are combined to perform needle mark detection on the square bonding pad, and a plurality of candidate needle mark targets are screened out; S3, performing non-maximum suppression strategy and texture clustering operation based on spatial collinearity on a plurality of candidate needle mark targets, removing redundant needle mark targets, and obtaining a plurality of final needle mark targets; S4, converting pixel coordinates of the plurality of final needle mark targets into normalized center coordinates through coordinate inverse mapping and format standardization, drawing a marking frame in the pad image data according to the normalized center coordinates corresponding to the plurality of final needle mark targets, and marking positions of the plurality of final needle mark targets.
2. The method for detecting the pin marks of the bonding pads based on the shape self-adaption and multi-dimensional feature fusion according to claim 1, wherein in the step S1, bonding pad type of the bonding pads is judged by inputting bonding pad image data into a geometric moment self-adaption classifier, and the method comprises the following steps: S1-1, calculating the length-width ratio of a pad in pad image data, comparing the length-width ratio of the pad with a set length-width threshold, and judging the type of the pad to be a strip-shaped pad when the length-width ratio of the pad is larger than or equal to the length-width threshold, and performing dark corner pixel detection in four-corner pixel detection when the length-width ratio of the pad is smaller than Yu Chang width threshold; S1-2, comparing the detected dark angle pixel value with a dark angle threshold value, judging that the type of the bonding pad is a round bonding pad when the dark angle pixel value is smaller than the dark angle threshold value, and detecting the bright angle pixel when the dark angle pixel value is larger than or equal to the dark angle threshold value; S1-3, comparing the detected bright angle pixel value with a bright angle threshold value, judging that the type of the bonding pad is a square bonding pad when the bright angle pixel value is larger than or equal to the bright angle threshold value, and calculating the contour circularity of the bonding pad in the bonding pad image data when the bright angle pixel value is smaller than the bright angle threshold value; S1-4, comparing the calculated contour circularity with a set circularity threshold value, judging that the type of the bonding pad is a circular bonding pad when the contour circularity is larger than or equal to the circularity threshold value, and judging that the type of the bonding pad is a square bonding pad when the contour circularity is smaller than the circularity threshold value.
3. The method for detecting the needle marks of the bonding pads based on the shape self-adaption and multi-dimensional feature fusion of claim 1 is characterized in that in the step S2, when the bonding pad type is long, a partition detection strategy of an overlapped sliding window is adopted to balance local illumination of the long bonding pad, the long bonding pad is divided into a plurality of subareas along the long side direction, a certain proportion of overlapping is kept between the adjacent subareas, bilateral filtering operation is adopted for each subarea in the plurality of subareas, and the plurality of subareas subjected to bilateral filtering operation are sequentially screened through a high-contrast channel, an extremely dark pixel channel and a mixed channel which are parallel in a three-channel candidate extraction model based on local statistical characteristics, so that a plurality of initial candidate areas are obtained.
4. The method for detecting the needle mark of the bonding pad based on the shape self-adaption and multi-dimensional feature fusion according to claim 3, wherein four physical dimensions including darkness score, circularity score, contrast score and morphology score are comprehensively considered in a quality scoring system of the multi-dimensional feature fusion, and the confidence score of each initial candidate region is calculated, wherein the calculation formula is as follows: Wherein S dark is represented as darkness score, S circ is represented as circularity score, S cont is represented as contrast score, S morph is represented as morphology score, and W1, W2, W3 and W4 are represented as weight coefficients of darkness score, circularity score, contrast score and morphology score, respectively.
5. The method for detecting the needle mark of the bonding pad based on the shape self-adaption and multi-dimensional feature fusion is characterized in that the edge distance is partitioned, the minimum Euclidean distance from the center of each initial candidate area to the boundary of the strip bonding pad is calculated, when the minimum Euclidean distance of any initial candidate area is smaller than a set distance threshold value, the initial candidate area marked as the edge area is judged by adopting a high confidence threshold value, when the confidence score of the initial candidate area marked as the edge area is larger than the high confidence threshold value, the initial candidate area marked as the edge area is judged as a needle mark target of the bonding pad, and when the confidence score of the initial candidate area marked as the edge area is smaller than or equal to the high confidence threshold value, the initial candidate area marked as the edge area is judged as a non-needle mark target; When the minimum Euclidean distance of any initial candidate region is greater than or equal to a set distance threshold, the initial candidate region is judged to be an internal region, the initial candidate region marked as the internal region is judged by adopting a low confidence coefficient threshold, when the confidence coefficient score of the initial candidate region marked as the internal region is greater than the low confidence coefficient threshold, the initial candidate region marked as the internal region is judged to be a candidate needle mark target, and when the confidence coefficient score of the initial candidate region marked as the internal region is less than or equal to the low confidence coefficient threshold, the initial candidate region marked as the internal region is judged to be a non-needle mark target.
6. The method for detecting the needle mark of the bonding pad based on the shape self-adaption and multidimensional feature fusion according to claim 1 is characterized in that when the bonding pad type is square, the needle mark detection is carried out on the square bonding pad by combining a local self-adaption threshold value with gradient direction verification, a plurality of candidate needle mark targets are screened out, and the method comprises the following steps: S2-1, calculating pixel statistics of an internal area of a square bonding pad to obtain a mean value and a standard deviation, calculating a dynamic binarization threshold value through the mean value and the standard deviation, and dividing bonding pad image data after Gaussian blur in parallel based on the dynamic binarization threshold value and a maximum inter-class variance method based on Otsu to obtain a plurality of dark candidate areas; S2-2, calculating a Sobel gradient field of a dark candidate region close to the boundary of the square bonding pad, and judging that the dark candidate region is an edge gradient artifact and removing when the gradient amplitude Mag of any dark candidate region exceeds the P 60 percentile of the global edge intensity and the gradient direction theta shows high consistency, so as to obtain a reserved dark candidate region; s2-3, setting multidimensional feature constraints including luminosity constraints, morphological constraints and position constraints on the reserved dark candidate areas, and screening the reserved dark candidate areas by combining geometry and luminosity to screen out a plurality of candidate needle mark targets.
7. The method for detecting the needle marks of the bonding pad based on the shape self-adaption and multi-dimensional feature fusion of claim 1, wherein when the bonding pad type is circular, the needle marks of the circular bonding pad are detected by a detection strategy based on region growth and concentric partition statistical verification, a plurality of candidate needle mark targets are screened out, and the method comprises the following steps: s3-1, extracting a pure detection area by adopting an edge reverse area growth method to obtain a plurality of non-communicated needle mark areas; S3-2, decomposing the connected region of the adhesion needle mark into a plurality of independent needle mark regions by a segmentation algorithm based on distance transformation; S3-3, dividing the circular bonding pad into three concentric areas of a core area, a middle area and an edge area according to Euclidean distance from a pixel point in the circular bonding pad to the geometric center of the circular bonding pad; S3-4, taking the needle mark region in the core region as a reference sample, calculating weighted standardized distances between the needle mark regions in the central region and the edge region and the distribution of the reference sample in the core region, comparing the weighted standardized distances with a loose threshold when the needle mark region is positioned in the central region, judging the needle mark region as a candidate needle mark target when the weighted standardized distance is smaller than or equal to the loose threshold, and comparing the weighted standardized distances with a strict threshold when the needle mark region is positioned in the edge region, and judging the needle mark region as a candidate needle mark target when the weighted standardized distance is smaller than or equal to the strict threshold.
8. The method for detecting the needle mark of the bonding pad based on the shape self-adaption and multidimensional feature fusion is characterized in that in the step S3, the non-maximum value inhibition strategy is adopted, a plurality of candidate needle mark targets are arranged in descending order according to confidence scores of the candidate needle mark targets, a plurality of candidate needle mark targets with highest confidence scores are selected as needle mark targets of reserved candidates, the Euclidean distance between the needle mark target with the highest confidence score in the remaining candidate needle mark targets and the center point of the needle mark targets of reserved candidates is calculated, the Euclidean distance of the center point is compared with a dynamic overlapping threshold, and if the Euclidean distance of the center point is lower than the dynamic overlapping threshold, the needle mark targets with the highest confidence scores in the remaining candidate needle mark targets are removed.
9. The method for detecting the needle marks of the bonding pad based on the shape self-adaption and multi-dimensional feature fusion is characterized in that in the step S3, a transverse clustering device and a longitudinal clustering device are constructed based on the texture clustering operation based on the space co-linearity, the transverse clustering device carries out horizontal grouping on the candidate needle mark targets based on the central ordinate of the candidate needle mark targets, the longitudinal clustering device carries out vertical grouping on the candidate needle mark targets based on the central abscissa of the candidate needle mark targets, for any non-empty group, the distance between adjacent needle mark targets is calculated after sorting according to the main axis coordinates, when the maximum distance between the adjacent needle mark targets is larger than a tight condition threshold value, all the needle mark targets of the non-empty group are judged to be non-needle mark targets, the non-empty group is eliminated, and when the number of the needle mark targets in the non-empty group is larger than or equal to 2, all the needle mark targets of the non-empty group are judged to be non-needle mark targets, and the non-empty group is eliminated.

Description

Shape self-adaption and multidimensional feature fusion-based pad needle mark detection method Technical Field The invention relates to the technical field of pad needle mark detection, in particular to a pad needle mark detection method based on shape self-adaption and multidimensional feature fusion. Background In the integrated circuit manufacturing process, the wafer probe CP is a core link connecting the wafer manufacturing and packaging test, the probe card contacts with the chip pad through the micro probe and applies an electrical signal, and the trace left after the test contains key process information. The depth and the form of the needle mark directly reflect the contact quality, namely, the excessive shallow depth can cause false killing of good products, and the excessive deep depth can damage the passivation layer and embed the hidden danger of reliability, so that the high-precision automatic detection of the needle mark is an important means for monitoring the state of the probe card and ensuring the yield of chips. In the face of advanced process, the traditional manual visual inspection is low in efficiency and high in subjectivity, and the existing automatic detection technology based on template matching or fixed threshold is difficult to meet industrial requirements in terms of adaptability and precision, and particularly, when complex texture background and special-shaped bonding pads are processed, two main bottlenecks of feature confusion and uneven light field are faced. The prior detection technology mainly faces serious challenges in practical application, and mainly comprises four mutually related layers, namely firstly, the generalization problem caused by the diversification of the shapes of the bonding pads, the problem that the traditional single operator has difficulty in considering different light and shadow distributions of bonding pads with different shapes to cause missed detection or false alarm, secondly, the problem that the strip bonding pads have obvious surface brightness gradient due to large span, the problem that a fixed threshold algorithm cannot consider the whole situation, the partition detection strategy of self-adaptive local illumination is needed, on the basis, the strong interference of the surface textures of the bonding pads further aggravates the detection difficulty, the grain microstructure and the surface morphology formed in the aluminum metal sputtering deposition process, the edge burrs left by etching openings of the passivation layer and the slight scratches accumulated on the surface of the bonding pads due to the contact of probes are extremely similar to real needle marks in gray scale and gradient characteristics, and are regularly distributed along a specific direction, the judgment of single-point characteristics is extremely easy to cause false defect judgment, and the problem that multi-dimensional characteristic fusion and macroscopic space distribution analysis are necessary to be introduced to remove internal noise of the bonding pads, and finally, the micrometer needle marks have extremely similar to have high contrast and shallow contrast and high sensitivity on the wafer. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a method for detecting the needle marks of the bonding pad based on the fusion of shape self-adaption and multidimensional characteristics, which can adapt to the shape of the bonding pad, inhibit the internal texture interference of the bonding pad, accurately separate edge noise, accurately detect the needle marks on the bonding pad and output bonding pad image data with a needle mark marking frame. In order to achieve the purpose, the method for detecting the needle mark of the bonding pad based on the shape self-adaption and multidimensional feature fusion comprises the following steps: S1, acquiring pad image data, constructing a geometric moment self-adaptive classifier, and judging whether the type of the pad is rectangular, square or round according to the aspect ratio and the circularity factor of the pad in the pad image data; S2, respectively carrying out edge partition and multi-scale self-adaptive judgment on the bonding pad in the bonding pad image data according to the bonding pad type, when the bonding pad type is long-strip-shaped, equalizing local illumination of the long-strip-shaped bonding pad by adopting a partition detection strategy of an overlapped sliding window, carrying out needle mark detection by a built three-channel candidate extraction model based on local statistical characteristics, and extracting a plurality of initial candidate areas; When the type of the bonding pad is square, the local self-adaptive threshold value and gradient direction verification are combined to perform needle mark detection on the square bonding pad, and a plurality of candidate needle mark targets are screened out; S3, performing non-maximum suppressio