Integrated circuit surface defect classification method based on differential manifold

Abstract

The invention discloses an integrated circuit surface defect classification method based on a micro manifold, which comprises a microscopic imaging acquisition system, an image preprocessing system and a defect image classification system, and comprises the following steps: firstly, mapping a preprocessed integrated circuit image to a Riemannian manifold space; secondly, feature descriptors of the images are constructed, feature vectors are obtained, and a covariance matrix can be constructed and corrected into a positive definite symmetric matrix of the Lie group manifold; then, further mapping the matrix to a low-dimensional Lie group space, and designing a Lie group kernel function; and finally, realizing accurate classification of the diversity defect images on the surface of the integrated circuit. The method solves the problem of rapid classification of diversity defect images on the surface of the integrated circuit, improves the reliability in the manufacturing process of the integrated circuit, and guarantees the quality of the downstream application industry of the integrated circuit.

Description

Technical Field

[0001] The invention belongs to the technical field of integrated circuit manufacturing, and in particular relates to an integrated circuit surface defect classification method based on differential manifolds. Background Art

[0002] With the rapid arrival of the global smart economy era, a new round of technological innovation and industrial transformation and upgrading is rapidly evolving. The semiconductor industry, one of the five emerging manufacturing industries, is located at the upstream port, which determines its priority development status, and the industrial competition among countries is becoming increasingly fierce. As a core component of modern electronic technology, integrated circuits play an important role in applications in various fields. Its high integration, miniaturization and high performance make it have broad application prospects in information technology, communications, medical and industrial control. Summary of the invention

[0003] The object of the present invention is to provide an integrated circuit surface defect classification method based on differential manifolds to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for classifying surface defects of integrated circuits based on differential manifolds, specifically comprising the following steps:

[0005] Step (1) collecting digital images of various parts of the surface of the integrated circuit by microscopic imaging;

[0006] Step (2) preprocessing the acquired image;

[0007] Step (3) extracts and classifies the preprocessed integrated circuit surface image to facilitate subsequent determination of defect types.

[0008] Furthermore, the acquisition process of step (1) requires the aid of a high-power microscope. Usually, a complete integrated circuit needs to be divided into hundreds of regions for image acquisition, and there is always a small overlap between the images of each region. Therefore, it is often difficult to obtain a standard template in this process.

[0009] Furthermore, in step (2), since a high-power microscope is used to collect industrial images online, the complexity of the environment often leads to poor image quality, such as the introduction of multiple unknown noises, low image contrast, and jitter blur in the complex industrial environment. Regardless of the reasons for the image quality degradation, it will affect the accuracy of subsequent defect classification. Preprocessing the collected images first will help improve the accuracy of defect classification. The specific steps include:

[0010] Step (2.1) analyzing factors affecting the quality of surface images collected online by integrated circuits;

[0011] Step (2.2) performs noise reduction processing on the integrated circuit image collected online;

[0012] Step (2.3) improves the overall brightness of the denoised integrated circuit surface image and performs contrast enhancement processing.

[0013] Step (2.4) completes the preprocessing operation of the integrated circuit image.

[0014] Furthermore, in step (2.2), there are many types of noise introduced during the online acquisition process of the industrial environment, such as salt and pepper noise, Gaussian noise, and mixed noise. This patent performs noise reduction processing on a variety of unknown noises, specifically including:

[0015] Step (2.2.1) uses topological correlation theory to analyze the surface defect image of the integrated circuit;

[0016] Step (2.2.2) selects a homotopy mapping suitable for the integrated circuit surface defect image and constructs a new topological space;

[0017] Step (2.2.3) combines Gaussian-Laplace transform to design a filter template suitable for a variety of unknown noises;

[0018] Step (2.2.4) implements noise reduction processing for various unknown noises.

[0019] Furthermore, in the step (2.3), the surface image of the integrated circuit is first subjected to contrast enhancement processing, and then a zero-crossing processing operation is performed to ensure that the intrinsic features of the defects are extracted, thereby improving the accuracy of defect classification.

[0020] Furthermore, the step (3) specifically includes:

[0021] Step (3.1) selects a suitable mapping method for the preprocessed integrated circuit image and maps the image to the Riemann manifold space.

[0022] Step (3.2) constructs a feature descriptor of the image and obtains a feature vector L;

[0023] Step (3.3) constructs the covariance matrix J through the eigenvectors;

[0024] Step (3.4) modifies the covariance matrix J to be the positive definite symmetric matrix X of the Lie group manifold;

[0025] Step (3.5) calculates the distance D between positive definite symmetric matrices;

[0026] Step (3.6) maps the positive definite symmetric matrix image descriptor to the low-dimensional Lie group space and designs the Lie group kernel function;

[0027] Step (3.7) realizes accurate classification of integrated circuit surface defect images and forms a "defect type library" of the system.

[0028] Furthermore, in the step (3.1), since the metric of Euclidean space is difficult to use for nonlinear data, the preprocessed integrated circuit image is mapped to the Riemannian manifold space so as to more significantly represent various features of the image, including color, edge, and texture.

[0029] Furthermore, in the step (3.2), due to the diversity of surface defects of integrated circuits, a single feature descriptor cannot effectively represent the diverse defect features, so the feature descriptor of this patent is composed of curvature, function gradient, divergence of vector domain, position spectrum, spectrum value, spectrum gradient, arc length and angle features. A feature vector L(x, y) can be defined for any position (x, y) of the image. There are many types of curvature of the surface, including normal curvature, Gauss curvature, cross-sectional curvature, mean curvature, and geodesic curvature. The commonly used curvature and cross-sectional curvature formulas in this patent are as follows:

[0030] Assume (M, g) is a Riemann manifold, ▽ is a Riemann connection, then the curvature tensor R is:

[0031] R(X,Y,Z,W)=<X,Z><Y,W> -<Y,Z><X,W> , (1)

[0032] Let Π be the tangent space T p The two-dimensional subspace of M, taking its basis as X, Y, defines the cross-sectional curvature of Π as:

[0033]

[0034] Among them, |X∧Y| 2 =<X,X><Y,Y> -<X,Y> 2 , the definition of the cross-sectional curvature is independent of the choice of the basis of Π.

[0035] The defect area of ​​integrated circuits produced during the manufacturing process is often smaller than the background area. Therefore, this patent first divides the image into blocks, constructs feature descriptors for the defective areas, and regards the remaining non-defective areas as invalid areas. This can not only reduce time consumption, but also effectively represent various types of defect characteristics.

[0036] Furthermore, in step (3.3), the feature vector L(x, y) is constructed by using multiple feature descriptors, and then the covariance between each feature vector and other feature vectors is calculated to construct a covariance matrix, which is designed as:

[0037]

[0038] in, The covariance matrix J reflects the statistical characteristics and spatial structure characteristics of the image, and the order of the feature descriptors is not determined when constructing the covariance matrix, so it has good robustness to the rotation and scaling of the image. Finally, the constructed covariance matrix J is normalized.

[0039] Furthermore, in the step (3.4), due to the small sample size of the integrated circuit defect image; on the other hand, although the online collected image has been subjected to noise reduction processing, there may still be a small amount of noise, which will cause the covariance matrix to be non-positive definite, thereby causing the learning algorithm to have unstable prediction results, so the non-positive definiteness of the covariance matrix is ​​very important. That is, after constructing the covariance matrix, it is necessary to first determine whether the covariance matrix is ​​non-positive definite, if so, analyze the cause, and then use the corresponding method of adding small positive values, regularization, and principal component analysis (PCA) to effectively correct the covariance matrix to a positive definite symmetric matrix.

[0040] Furthermore, in the step (3.5), the two images I i and I j The weighted distance between them is:

[0041]

[0042] Where K represents the total number of image regions, D(X i,k ,X j,k ) is image I i and I j The distance of the corresponding k-th region represents the weight of the k-th region.

[0043] Furthermore, in step (3.6), in the low-dimensional Lie group space, based on the geodesic distance, the Lie group kernel function between the two images can be designed as:

[0044] K(I i ,I j )=exp(-α(D(I i ,I j ))) , (5)

[0045] Here, the parameter α is the scaling factor of the direct space.

[0046] Preferably, in the step (3.7), the methods in which the Lie group kernel can be used for classification include k-nearest neighbor algorithm and support vector machine.

[0047] Compared with the prior art, the technical effects and advantages of the present invention are as follows: the integrated circuit surface defect classification method based on differential manifolds,

[0048] 1. The present invention pre-processes the integrated circuit image before classification, which helps to improve the accuracy of classification;

[0049] 2. The present invention maps the integrated circuit surface image to the Riemann manifold space, and the feature descriptor is composed of curvature, function gradient, divergence of the vector domain, position spectrum, spectrum value, spectrum gradient, arc length and angle features, so as to better represent the intrinsic characteristics of the diversity defects;

[0050] 3. The present invention adopts the idea of ​​first dividing into blocks and then constructing feature descriptors, which can effectively remove background interference;

[0051] 4. The present invention effectively classifies various types of defects on the surface of integrated circuits, which helps to more accurately determine whether such defects can be put into subsequent production;

[0052] 5. The present invention has an important breakthrough in the quality control of integrated circuits, improves the reliability of the integrated circuit manufacturing process, and ensures the quality of the downstream application industry of integrated circuits. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the integrated circuit surface defect rapid classification system in the detection method of the present invention;

[0054] Figure 2 is a flow chart of an image preprocessing system in the detection method of the present invention;

[0055] Figure 3 is a flow chart of diversity defect classification in the detection method of the present invention;

[0056] Figure 4 The present invention is a flowchart of the integrated circuit surface defect classification method based on differential manifolds. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] See also Figure 1-4 The present invention provides a technical solution: a method for classifying surface defects of integrated circuits based on differential manifolds, which specifically includes the following steps:

[0059] Step (1) collecting digital images of various parts of the surface of the integrated circuit by microscopic imaging;

[0060] Step (2) preprocessing the acquired image;

[0061] Step (3) extracts and classifies the preprocessed integrated circuit surface image to facilitate subsequent determination of defect types.

[0062] The acquisition process of step (1) requires the use of a high-power microscope. Usually, a complete integrated circuit needs to be divided into hundreds of regions for image acquisition. There is always a small overlap between the images of each region. Therefore, it is often difficult to obtain a standard template in this process.

[0063] In step (2), since a high-power microscope is used to collect industrial images online, the complexity of the environment often leads to poor image quality, such as the introduction of multiple unknown noises, low image contrast, and jitter blur in the complex industrial environment. Regardless of the reason for the image quality degradation, it will affect the accuracy of subsequent defect classification. Preprocessing the collected images first will help improve the accuracy of defect classification. The specific steps include:

[0064] Step (2.1) analyzing factors affecting the quality of surface images collected online by integrated circuits;

[0065] Step (2.2) performs noise reduction processing on the integrated circuit image collected online;

[0066] Step (2.3) improves the overall brightness of the denoised integrated circuit surface image and performs contrast enhancement processing.

[0067] Step (2.4) completes the preprocessing operation of the integrated circuit image.

[0068] In the step (2.2), there are many types of noise introduced during the online acquisition process of the industrial environment, such as salt and pepper noise, Gaussian noise, and mixed noise. This patent performs noise reduction processing on a variety of unknown noises, specifically including:

[0069] Step (2.2.1) uses topological correlation theory to analyze the surface defect image of the integrated circuit;

[0070] Step (2.2.2) selects a homotopy mapping suitable for the integrated circuit surface defect image and constructs a new topological space;

[0071] Step (2.2.3) combines Gaussian-Laplace transform to design a filter template suitable for a variety of unknown noises;

[0072] Step (2.2.4) implements noise reduction processing for various unknown noises.

[0073] In the step (2.3), the surface image of the integrated circuit is first subjected to contrast enhancement processing, and then a zero-crossing processing operation is performed to ensure that the intrinsic features of the defects are extracted, thereby improving the accuracy of defect classification.

[0074] The step (3) specifically comprises:

[0075] Step (3.1) selects a suitable mapping method for the preprocessed integrated circuit image and maps the image to the Riemann manifold space.

[0076] Step (3.2) constructs a feature descriptor of the image and obtains a feature vector L;

[0077] Step (3.3) constructs the covariance matrix J through the eigenvectors;

[0078] Step (3.4) modifies the covariance matrix J to be the positive definite symmetric matrix X of the Lie group manifold;

[0079] Step (3.5) calculates the distance D between positive definite symmetric matrices;

[0080] Step (3.6) maps the positive definite symmetric matrix image descriptor to the low-dimensional Lie group space and designs the Lie group kernel function;

[0081] Step (3.7) realizes accurate classification of integrated circuit surface defect images and forms a "defect type library" of the system.

[0082] In the step (3.1), since the metric of Euclidean space is difficult to use for nonlinear data, the preprocessed integrated circuit image is mapped to the Riemannian manifold space in order to more significantly represent various features of the image, including color, edge, and texture.

[0083] In the step (3.2), due to the diversity of surface defects of integrated circuits, a single feature descriptor cannot effectively represent the diverse defect features, so the feature descriptor is composed of curvature, function gradient, divergence of vector domain, position spectrum, spectrum value, spectrum gradient, arc length and angle features. A feature vector L(x, y) can be defined for any position (x, y) of the image. There are many types of surface curvature, including normal curvature, Gauss curvature, cross-sectional curvature, mean curvature, and geodesic curvature. Commonly used curvature and cross-sectional curvature formulas are as follows:

[0084] Assume (M, g) is a Riemann manifold, ▽ is a Riemann connection, then the curvature tensor R is:

[0085] R(X,Y,Z,W)=<X,Z><Y,W> -<Y,Z><X,W> , (1)

[0086] Let Π be the tangent space T pThe two-dimensional subspace of M, taking its basis as X, Y, defines the cross-sectional curvature of ∏ as:

[0087]

[0088] Among them, |X∧Y| 2 =<X,X><Y,Y> -<X,Y> 2 , the definition of the cross-sectional curvature is independent of the choice of the basis of Π.

[0089] The defect area of ​​integrated circuits produced during the manufacturing process is often smaller than the background area. Therefore, this patent first divides the image into blocks, constructs feature descriptors for the defective areas, and regards the remaining non-defective areas as invalid areas. This can not only reduce time consumption, but also effectively represent various types of defect characteristics.

[0090] In the step (3.3), the feature vector L(x, y) is constructed by using multiple feature descriptors, and then the covariance between each feature vector and other feature vectors is calculated to construct a covariance matrix, which is designed as:

[0091]

[0092] in, The covariance matrix J reflects the statistical characteristics and spatial structure characteristics of the image, and the order of the feature descriptors is not determined when constructing the covariance matrix, so it has good robustness to the rotation and scaling of the image. Finally, the constructed covariance matrix J is normalized.

[0093] In the step (3.4), due to the small sample size of the integrated circuit defect image; on the other hand, although the online collected image has been subjected to noise reduction processing, there may still be a small amount of noise, which will cause the covariance matrix to be non-positive definite, thereby causing the learning algorithm to have unstable prediction results, so the non-positive definiteness of the covariance matrix is ​​very important. That is, after constructing the covariance matrix, it is necessary to first determine whether the covariance matrix is ​​non-positive definite. If so, analyze the cause, and then use the corresponding method of adding small positive values, regularization, and principal component analysis (PCA) to effectively correct the covariance matrix to a positive definite symmetric matrix.

[0094] In the step (3.5), the two images I i and I j The weighted distance between them is:

[0095]

[0096] Where K represents the total number of image regions, D(X i,k ,X j,k ) is image I i and Ij The distance of the corresponding k-th region represents the weight of the k-th region.

[0097] In the step (3.6), in the low-dimensional Lie group space, based on the geodesic distance, the Lie group kernel function between the two images can be designed as:

[0098] K(I i ,I j )=exp(-α(D(I i ,I j ))) , (5)

[0099] Here, the parameter α is the scaling factor of the direct space.

[0100] In the step (3.7), the Lie group kernel can be used for classification by k-nearest neighbor algorithm and support vector machine.

[0101] The above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modification, substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for classifying surface defects of integrated circuits based on differential manifolds, characterized in that: The specific steps include: Step (1). Collecting digital images of various parts of the surface of the integrated circuit by microscopic imaging; Step (2) preprocessing the acquired image; Step (3). Extract and classify the preprocessed integrated circuit surface image.

2. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 1, characterized in that: The acquisition process of step (1) requires the use of a high-power microscope, and a complete integrated circuit needs to be divided into hundreds of areas for image acquisition. There is always a small overlap between the images of each area.

3. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 1, characterized in that: The preprocessing process of step (2) mainly solves the various unknown noises and low image contrast introduced in the online acquisition process of step (1), which helps to improve the accuracy of defect classification. The specific steps include: Step (2.1) analyzing factors affecting the quality of surface images collected online by integrated circuits; Step (2.2) performs noise reduction processing on the integrated circuit image collected online; Step (2.3) improves the overall brightness of the denoised integrated circuit surface image and performs contrast enhancement processing; Step (2.4) completes the preprocessing operation of the integrated circuit image.

4. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 3, characterized in that: In the step (2.2), the types of noise introduced during the online acquisition process of the industrial environment include: Step (2.2.1) uses topological correlation theory to analyze the surface defect image of the integrated circuit; Step (2.2.2) selects a homotopy mapping suitable for the integrated circuit surface defect image and constructs a new topological space; Step (2.2.3) combines Gaussian-Laplace transform to design a filter template suitable for a variety of unknown noises; Step (2.2.4) realizes noise reduction processing for various unknown noises.

5. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 3, characterized in that: In the step (2.3), the surface image of the integrated circuit is first subjected to contrast enhancement processing and then to zero-crossing processing, so as to extract the intrinsic features of the defect.

6. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 1, characterized in that: The specific steps of step (3) include: Step (3.1) selects a suitable mapping method for the preprocessed integrated circuit image and maps the image to the Riemann manifold space; Step (3.2) constructs a feature descriptor of the image and obtains a feature vector L; Step (3.3) constructs the covariance matrix J through the eigenvectors; Step (3.4) modifies the covariance matrix J to be the positive definite symmetric matrix X of the Lie group manifold; Step (3.5) calculates the distance D between positive definite symmetric matrices; Step (3.6) maps the positive definite symmetric matrix image descriptor to the low-dimensional Lie group space and designs the Lie group kernel function; Step (3.7) realizes accurate classification of integrated circuit surface defect images and forms a "defect type library" of the system.

7. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 6, characterized in that: In the step (3.1), the preprocessed integrated circuit image is mapped to the Riemann manifold space. The feature descriptor of the step (3.2) is composed of curvature, function gradient, divergence of the vector domain, position spectrum, spectrum value, spectrum gradient, arc length and angle features. In the step (3.2), the image is first divided into blocks, and the feature descriptor is constructed for the area containing defects, and the remaining parts without defects are regarded as invalid areas, indicating various types of defect characteristics.

8. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 6, characterized in that: The step (3.3) constructs a feature vector through multiple feature descriptors, and then calculates the covariance between each feature vector and other feature vectors, that is, constructs a covariance matrix. In the step (3.4), if the covariance matrix is ​​not positive definite, first analyze the reasons that cause the covariance matrix to be not positive definite, and then use the corresponding methods of adding small positive values, regularization, and principal component analysis (PCA) to correct the covariance matrix to a positive definite symmetric matrix of the Lie group manifold.

9. The method for classifying integrated circuit surface defects based on differential manifolds according to claim 6, characterized in that: In the step (3.5), the weighted distance between the two images adopts the Riemann metric. In the step (3.6), the Lie group kernel function between the two images is designed based on the geodesic distance in the low-dimensional Lie group space. In the step (3.7), the Lie group kernel is used for classification by the k-nearest neighbor algorithm and the support vector machine.

Images