A method for extracting the contour of photolithography images

Abstract

A method for extracting a lithography image contour comprises acquiring a lithography image, performing a rough extraction process on the lithography image contour by using a gradient amplitude to perform non-maximum suppression, obtaining a first lithography image contour C, and performing a fine extraction process on the roughly extracted first lithography image contour according to prior information of the lithography image, obtaining a second lithography image contour C. eff Therefore, the present invention designs relevant algorithms based on the inherent characteristics of the lithography pattern to suppress image defects, thereby obtaining the correct image contour, and completes the preparation of the contour extraction neural network training data set based on the above method, ensuring the quality of the neural network training data set and thus improving the accuracy of network inference.

Description

Technical Field

[0001] The invention belongs to the field of digital image processing and proposes a method for extracting the contour of a photolithography image. Background Art

[0002] Optical Proximity Calibration (OPC) technology is used to correct the deviation of lithography patterns caused by diffraction. OPC is a lithography enhancement technology mainly used in the semiconductor production process. With the continuous advancement of Moore's Law, the line width size is getting smaller and smaller, and is far lower than the wavelength of the lithography light source. The projection has a significant diffraction effect, resulting in image shape distortion, such as line width wider or narrower than the design, rounded corners, etc. If these distortions are not corrected, the electrical performance of the produced circuit will be greatly changed.

[0003] OPC technology corrects the above-mentioned graphic distortion by moving the edge or position of the graphic on the mask or adding additional polygonal patterns. OPC technology is mainly divided into two categories, one is the optical proximity effect correction based on experience, and the other is the optical proximity effect correction based on models. The former is more obviously affected by experience and lighting conditions. When the process conditions change, the rules must be revised again, which is less efficient. The latter has gradually begun to be widely used with the evolution of process generations (90nm technology node and later). This type of method mainly implements the OPC function based on optical models and photoresist models. The accuracy is mainly affected by experimental data. The more and more comprehensive the data, the higher the fitting accuracy.

[0004] OPC modeling requires CD measurement data of critical dimension scanning electron microscope (CDSEM) images. In the early days, 1D image measurement CD data was mostly used. As the complexity of processes and patterns increased, such data gradually failed to meet the needs of modeling. Therefore, the industry proposed modeling based on pattern contour data.

[0005] Those skilled in the art know that the edge of an image refers to a portion of a local area of ​​the image where the brightness changes significantly. The grayscale profile of this area can generally be regarded as a step, that is, a sharp change from one grayscale value in a small buffer area to another grayscale value with a large grayscale difference. The edge of an image concentrates most of the information of the image. The determination and extraction of the edge of the image is very important for the recognition and understanding of the entire image scene. It is also an important feature that image segmentation relies on. Edge detection is mainly the measurement, detection and positioning of the grayscale change of the image.

[0006] That is to say, from the perspective of image information, the amount of information contained in the contour data is much higher than that in the CD data, so the contour-based modeling method also achieves a better fitting effect.

[0007] However, due to the limited shooting conditions, the image of the lithography pattern is significantly disturbed by noise, the pattern contour is significantly distorted, and its contour extraction is significantly affected by noise and shape defects. The traditional image contour extraction method has two main problems:

[0008] ① The image contour accuracy is low, and the contour has serious pseudo-edge and disconnection phenomena;

[0009] ② The image contour extraction speed is slow, and the contour pattern is usually obtained through multiple iterations.

[0010] Specifically, the key to OPC technology based on pattern contour data lies in the acquisition of contour data. Due to its test environment conditions and other reasons, CDSEM images have defects such as significant noise and shape distortion, so it is very challenging to obtain accurate contour data. Traditional contour extraction techniques include various edge operators: sobel, prewit, log and other simple operators, but these operators have simple features and are usually suitable for less complex pattern contour extraction. For complex images, the effect is often poor.

[0011] The most widely used non-machine learning algorithm at present is the Canny algorithm, which is based on gradient calculation and non-maximum suppression algorithm. It can obtain the contours of complex patterns more accurately, so it is widely used. However, the accuracy of obtaining the contours of complex patterns is still not ideal.

[0012] There are relatively few contour extraction methods based on machine learning, most of which are derived from image segmentation technology. However, actual efficient image segmentation technology is often supported by complex network models and hardware with powerful computing power, and is not suitable for the extraction of lithography image contours. Summary of the invention

[0013] In view of the problems of the above contour extraction technology, the present invention proposes an accurate and efficient contour extraction method, which mainly includes the following two parts:

[0014] ① A high-quality image contour extraction method based on prior information of lithography images is proposed, and based on this method, the preparation of datasets for contour network training is realized;

[0015] ② A deep neural network is designed to extract the contour information of lithography images, which has a good repair effect on defects such as pseudo edges and edge missing.

[0016] To achieve the above object, the technical solution of the present invention is as follows:

[0017] A method for extracting a lithography image contour comprises the following steps:

[0018] Step S11: acquiring a photolithography image, and performing a rough extraction process on the photolithography image contour by using a gradient amplitude non-maximum suppression method to obtain a first photolithography image contour C;

[0019] Step S12: performing a refined extraction process on the roughly extracted first photolithography image contour according to the prior information of the photolithography image to obtain the second photolithography image contour C eff .

[0020] Furthermore, the prior information of the lithography image is the first prior information, the second prior information and the third prior information; wherein, the first prior information: the pixel value of the background area is low, and the edge measurement is usually the background and the shadow; the second prior information: if it is a shadow, the connection between each part of the image is background-shadow-pattern, and a pseudo-edge appears between the shadow and the background; if it is an underlying pattern, the connection is background-shadow-background; the third prior information: the pseudo-edge length is limited, much lower than the pattern edge length, and is not connected to the edge.

[0021] Furthermore, step S12 in the photolithography image contour extraction method specifically includes:

[0022] Step S121: Calculate the neighborhood pixel mean M in the first lithography image profile C by combining the first prior information and the second prior information;

[0023] Step S122: Calculate the mean M of the neighborhood pixels of each pixel in the first lithography image profile C, and compare the neighborhood pixel mean M with the threshold T. If the neighborhood pixel mean M is greater than or equal to the threshold T, the pixel is a valid edge pixel and its pixel value is retained. Otherwise, the pixel value of the pixel is set to zero. Finally, a first intermediate lithography image profile C is obtained according to each pixel of the first lithography image profile C. eff12 ;

[0024] Step S123: Combine the third prior information to calculate the first intermediate lithography image profile C eff12 The edge points in the image are processed by 8-link component length threshold to obtain the second intermediate lithography image contour C eff13 Then, the second intermediate lithography image profile C eff13 is the seed, based on the second intermediate lithography image profile C eff13 Perform 8 linking to obtain the second lithography image profile C eff .

[0025] Furthermore, the photolithography image contour extraction method further comprises:

[0026] Step 124: Detect the second photolithography image profile C eff The second photolithography image profile C is repaired eff The fracture.

[0027] To achieve the above object, another technical solution of the present invention is as follows:

[0028] A method for extracting a lithography image contour comprises the following steps:

[0029] Step S21: acquiring a plurality of photolithography images;

[0030] Step S22: using the above-mentioned method for extracting the contour of the lithography image, extracting the contours of the plurality of acquired lithography images one by one to form a training contour data set;

[0031] Step S23: Based on the training contour data set, use a full-resolution convolutional deep neural network to learn image contour features.

[0032] 6. The method according to claim 5, characterized in that step S23 specifically comprises the following steps:

[0033] Step S231: splitting the original contour image in the training contour data set into multiple sub-images using a reversible downsampling method, and using the multiple contour sub-images as neural network feature learning inputs;

[0034] Step S232: inputting the multiple contour sub-graphs into the full-resolution nonlinear mapping layer of the full-resolution convolutional deep neural network to learn image contour features, and outputting the contour sub-graphs;

[0035] Step S233: fusing the multiple contour sub-images into a full-resolution contour image through downsampling inverse transformation.

[0036] Furthermore, the full-resolution nonlinear mapping layer includes three types of operations; the first type of operation is convolution Conv, the second type of operation is batch normalization BN, and the third type of operation is the rectified linear activation function ReLU.

[0037] Furthermore, the full-resolution nonlinear mapping layer includes K cyclic units, and the cyclic units include three layers. The first layer adopts the Conv+ReLU mode, the second layer adopts the Conv+BN+ReLU mode, and the third layer adopts the pure Conv mode to realize the mapping function from image to contour and realize the output of contour sub-image, where K is a positive integer.

[0038] Furthermore, the reversible downsampling adopts a downsampling rate of 1 / 2, the size of the original contour image in the training contour data set is (M, N), then the number of the contour sub-images is 4, and the size of the contour sub-image is (M / 2)*(N / 2).

[0039] Furthermore, a zero-padding method is used to compensate for the resolution loss caused by the first type of operation.

[0040] It can be seen from the above technical scheme that compared with the existing contour extraction technology, the contour extraction technology in the present invention has the remarkable characteristics of high accuracy and high speed, and is very suitable for integrated circuit manufacturing design and production. The specific beneficial effects are as follows:

[0041] On the one hand, the present invention uses the powerful feature characterization capability of deep neural networks to extract contour features of lithography images and applies them to contour extraction of other patterns, obtaining accurate results. In addition, actual testing of the neural network can be accelerated by hardware (GPU, etc.), achieving real-time processing effects and significantly improving the efficiency of current lithography contour extraction.

[0042] On the other hand, the present invention proposes a high-accuracy contour extraction method based on prior information of lithography images. By designing relevant algorithms based on the inherent characteristics and alignability of the lithography patterns, image defects can be suppressed to obtain the correct image contour. Based on this method, the preparation of the contour extraction neural network training data set is completed, which ensures the quality of the neural network training data set and thus improves the accuracy of network inference. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 FIG. 1 is a flow chart of a contour extraction method based on a priori lithography images in an embodiment of the present invention.

[0044] Figure 2 Schematic diagram showing the result of contour extraction algorithm based on lithography image prior in an embodiment of the present invention

[0045] Figure 3 The figure shows a schematic diagram of the structure of a deep neural network in an embodiment of the present invention.

[0046] Figure 4 FIG. 1 is a flow chart of a contour extraction method based on a deep neural network in an embodiment of the present invention.

[0047] Figure 5 Schematic diagram showing the contour extraction result based on a deep neural network in an embodiment of the present invention DETAILED DESCRIPTION

[0048] The following is combined with Figure 1-5 , the specific implementation modes of the present invention are further described in detail.

[0049] It should be noted that the present invention discloses a method for extracting lithography image contours, which adopts a machine learning method to train a deep neural network model suitable for lithography pattern contour extraction through a self-made contour data set, and makes use of the powerful characterization ability of deep neural networks for image features and the advantages of neural network hardware acceleration to achieve accurate and rapid extraction of lithography pattern contours.

[0050] The main innovations of the present invention mainly include three aspects:

[0051] ①. Use prior information of lithography images to prepare high-quality contour training datasets;

[0052] ② Use full-resolution convolutional deep neural network to learn image contour features;

[0053] ③. Use downsampling technology to increase the field of view width, enhance feature learning capabilities, and improve processing speed.

[0054] For the first major innovation of the present invention, please refer to Figure 1 , Figure 1 The figure shows a flow chart of the contour extraction method based on lithography image prior information in an embodiment of the present invention. It should be noted that the high-accuracy contour extraction method based on lithography image prior information is based on the inherent characteristics of the lithography pattern, and the relevant algorithm is designed to suppress image defects, thereby obtaining the correct image contour, and based on this method, the contour extraction neural network training data set preparation is completed, which ensures the quality of the neural network training data set and thus improves the accuracy of network inference.

[0055] like Figure 1 As shown, the lithography image contour extraction method includes the following steps:

[0056] Step S11: acquiring a photolithography image, and performing a rough extraction process on the photolithography image contour by using a gradient amplitude non-maximum suppression method to obtain a first photolithography image contour C;

[0057] Step S12: performing a refined extraction process on the roughly extracted first photolithography image contour according to the prior information of the photolithography image to obtain the second photolithography image contour C eff .

[0058] In an embodiment of the invention, the prior information of the lithography image may be first prior information, second prior information and third prior information; wherein,

[0059] First prior information: the background area has low pixel values, and edge measurements are usually background and shadows;

[0060] Second prior information: if it is a shadow, the connection between each part of the image is background-shadow-pattern, and a pseudo edge appears between the shadow and the background; if it is an underlying pattern, the connection is background-shadow-background;

[0061] The third prior information: the pseudo edge length is limited, much lower than the pattern edge length, and is not connected to the edge.

[0062] Please see again Figure 1 In a preferred embodiment of the present invention, the step S12 may specifically include:

[0063] Step S121: Calculate the neighborhood pixel mean M in the first lithography image profile C by combining the first prior information and the second prior information;

[0064] Step S122: Calculate the mean M of the neighborhood pixels of each pixel in the first lithography image profile C, and compare the neighborhood pixel mean M with the threshold T. If the neighborhood pixel mean M is greater than or equal to the threshold T, the pixel is a valid edge pixel and its pixel value is retained. Otherwise, the pixel value of the pixel is set to zero. Finally, a first intermediate lithography image profile C is obtained according to each pixel of the first lithography image profile C. eff12 ;

[0065] Step S123: Combine the third prior information to calculate the first intermediate lithography image profile C eff12 The edge points in the image are processed by 8-link component length threshold to obtain the second intermediate lithography image contour C eff13 Then, the second intermediate lithography image profile C eff13 is the seed, based on the second intermediate lithography image profile C eff13 Perform 8 linking to obtain the second lithography image profile C eff .

[0066] In the embodiment of the present invention, the second photolithography image profile C can be further repaired. eff Specifically, the step S12 may further include:

[0067] Step 124: Detect the second photolithography image profile C eff The second photolithography image profile C is repaired eff The fracture.

[0068] It should be noted that, in the embodiment of the present invention, the edge points in the first lithography image profile C may be subjected to 8-link component length threshold processing in combination with the third prior information to obtain the second intermediate lithography image profile C. eff13 Then, by combining the first prior information and the second prior information, the second intermediate lithography image profile C is calculated. eff13The mean value M of the pixels in the neighborhood; the second intermediate lithography image profile C eff13 The neighborhood pixels of each pixel in the intermediate lithography image are averaged and the neighborhood pixel average M is compared with the threshold T. If the neighborhood pixel average M is greater than or equal to the threshold T, the pixel is a valid edge pixel and its pixel value is retained. Otherwise, the pixel value of the pixel is set to zero. Finally, according to the second intermediate lithography image profile C eff13 Each pixel of the first intermediate lithography image profile C eff12 Then, with the first intermediate lithography image profile C eff12 is the seed, based on the second intermediate lithography image profile C eff13 Perform 8 linking to obtain the second lithography image profile C eff .

[0069] In practical applications, after completing the above-mentioned lithography image contour extraction method steps, a relatively clear contour image can also be obtained. Figure 2 , Figure 2 The figure shows the result of the contour extraction algorithm based on the prior of the lithography image in the embodiment of the present invention. From the result, it can be seen that the pseudo edges of the lithography image are well consistent, and the pseudo edges caused by the shadow of the underlying pattern are also effectively eliminated.

[0070] On the basis of the completion of the above-mentioned lithography image contour extraction method steps, the second and third innovations of the present invention are well embodied. In other words, the second innovation of the present invention is to use a full-resolution convolutional deep neural network to learn image contour features, and the training set used comes from the above-mentioned lithography image contour extraction method.

[0071] In an embodiment of the present invention, a purely discriminative convolutional learning network is used to learn a full-resolution mapping model from a lithography image to a contour. Figure 3 , Figure 3 FIG. 1 is a schematic diagram of a deep neural network structure in an embodiment of the present invention. Figure 3 As shown, the upper part of the figure is a schematic diagram of the neural network framework, and the lower part is a schematic diagram of the full-resolution nonlinear mapping structure of the intermediate layer.

[0072] See also Figure 4 , Figure 4 FIG. 2 is a flow chart of a contour extraction method based on a deep neural network in an embodiment of the present invention. Figure 4 As shown, the contour extraction method based on deep neural network includes the following steps:

[0073] Step S21: acquiring a plurality of photolithography images; each photolithography image profile will be used as a set of data in constructing a training data set;

[0074] Step S22: using the above-mentioned method for extracting the contour of the lithography image, extracting the contours of the plurality of acquired lithography images one by one to form a training contour data set;

[0075] Step S23: Based on the training contour data set, use a full-resolution convolutional deep neural network to learn image contour features.

[0076] The downsampling technology is used to increase the field of view width, and the step S23 specifically includes the following steps:

[0077] Step S231: using a reversible down-sampling method to split the original contour image in the training contour data set into multiple sub-images, and using the multiple contour sub-images as neural network feature learning inputs.

[0078] Specifically, in an embodiment of the present invention, a reversible downsampling method is used to split the original contour image into multiple contour sub-images. Figure 4 When , a 1 / 2 downsampling rate can be used, so the original contour image is split into four contour sub-images (such as Figure 3 As shown). Assuming that the size of the original contour image is (M, N), there are 4 contour sub-images, and the size of the contour sub-image is (M / 2)*(N / 2). Then, the sub-image is used as the input of the neural network feature learning.

[0079] The role of sub-sampling is two-fold:

[0080] ① Expand the receptive field of the neural network;

[0081] ②. Accelerate the processing speed of neural networks (because the input image size is reduced by 75%).

[0082] With the above outline subgraph, you can perform the following steps:

[0083] Step S232: inputting the multiple contour sub-graphs into the full-resolution nonlinear mapping layer of the full-resolution convolutional deep neural network to learn image contour features, and outputting the contour sub-graphs;

[0084] Step S233: fusing the multiple contour sub-images into a full-resolution contour image through downsampling inverse transformation.

[0085] like Figure 3As shown, the contour sub-image is input into the full-resolution nonlinear mapping layer for learning image contour features. The full-resolution nonlinear mapping layer includes three types of operations; the first type of operation is convolution Conv, the second type of operation is batch normalization BN, and the third type of operation is the rectified linear activation function ReLU. The full-resolution nonlinear mapping layer includes K cyclic units, and the cyclic unit includes three layers. The first layer adopts the Conv+ReLU mode, the second layer adopts the Conv+BN+ReLU mode, and the third layer adopts the pure Conv mode to realize the mapping function from image to contour and realize the output of contour sub-image, where K is a positive integer. The reversible downsampling adopts a downsampling rate of 1 / 2. The size of the original contour image in the training contour data set is (M, N), then the contour sub-images are 4, and the size of the contour sub-image is (M / 2)*(N / 2).

[0086] Finally, the full-resolution contour image (M, N) is fused through the downsampling inverse transform. Preferably, in order to maintain the image resolution, the intermediate layer uses the zero padding method to compensate for the resolution loss caused by the convolution operation. In this example, the convolution kernel size is set to 3*3, the number of intermediate layers is 15, the number of channels is 64, the activation function uses the root mean square error, and the learning rate is set to 5*10 -5 .

[0087] See also Figure 5 , Figure 5 FIG. 1 is a schematic diagram of contour extraction results based on a deep neural network in an embodiment of the present invention. Figure 5 As shown, the lithography pattern contour result obtained by the neural network processing of the present invention has two outstanding effects. First, the broken part of the lithography pattern is successfully repaired, and second, no pseudo edge appears, indicating that the model obtained by the deep neural network through the high-accuracy contour data set training can well "characterize" the correct contour features and obtain correct test results.

[0088] In summary, neural network training requires a high-quality data set to obtain better contour extraction effects. Due to the shape distortion of the lithography image, the contour of the real image cannot be obtained naturally. If the contour data of the distorted image is directly used as the training data set, the model effect will inevitably be poor. If the pattern contour data set is prepared based on empirical information, it will inevitably take too long and the accuracy and consistency will be difficult to guarantee. The high-quality contour extraction method based on image prior information provided in the present invention can significantly improve the quality of the training data set.

[0089] The above-described embodiments are only preferred embodiments of the present invention, and the embodiments are not intended to limit the patent protection scope of the present invention. Therefore, any equivalent structural changes made using the description and drawings of the present invention should also be included in the protection scope of the present invention.

Claims

1. A method for extracting a lithography image contour, characterized in that: The method comprises the following steps: Step S11: acquiring a photolithography image, and performing a rough extraction process on the photolithography image contour by using a gradient amplitude non-maximum suppression method to obtain a first photolithography image contour C; Step S12: performing a refined extraction process on the roughly extracted first photolithography image contour according to the prior information of the photolithography image to obtain a second photolithography image contour C eff ; The prior information of the lithography image is first prior information, second prior information and third prior information; The step S12 specifically includes: Step S121: Calculate the neighborhood pixel mean M in the first lithography image profile C by combining the first prior information and the second prior information; Step S122: Calculate the mean M of the neighborhood pixels of each pixel in the first lithography image profile C, and compare the neighborhood pixel mean M with the threshold T. If the neighborhood pixel mean M is greater than or equal to the threshold T, the pixel is a valid edge pixel and its pixel value is retained. Otherwise, the pixel value of the pixel is set to zero. Finally, a first intermediate lithography image profile C is obtained according to each pixel of the first lithography image profile C. eff12 ; Step S123: Combine the third prior information to calculate the first intermediate lithography image profile C eff12 The edge points in the image are processed by 8-link component length threshold to obtain the second intermediate lithography image contour C eff13 Then, the second intermediate lithography image profile C eff13 is the seed, based on the second intermediate lithography image profile C eff13 Perform 8 linking to obtain the second lithography image profile C eff .

2. The method according to claim 1, characterized in that First prior information: the background area has low pixel values, and edge measurements are usually background and shadows; Second prior information: if it is a shadow, the connection between each part of the image is background-shadow-pattern, and a pseudo edge appears between the shadow and the background; if it is an underlying pattern, the connection is background-shadow-background; The third prior information: the pseudo edge length is limited, much lower than the pattern edge length, and is not connected to the edge.

3. The method according to claim 1, characterized in that Also includes: Step 124: Detect the second photolithography image profile C eff The second photolithography image profile C is repaired eff The fracture.

4. A method for extracting a lithography image contour, characterized in that: include: Step S21: acquiring a plurality of photolithography images; Step S22: using the method for extracting the contour of a lithography image according to any one of claims 1 to 3, extracting the contours of the plurality of acquired lithography images one by one to form a training contour data set; Step S23: Based on the training contour data set, use a full-resolution convolutional deep neural network to learn image contour features.

5. The method according to claim 4, characterized in that The downsampling technology is used to increase the field of view width, and the step S23 specifically includes the following steps: Step S231: splitting the original contour image in the training contour data set into multiple sub-images using a reversible downsampling method, and using the multiple contour sub-images as neural network feature learning inputs; Step S232: inputting the multiple contour sub-graphs into the full-resolution nonlinear mapping layer of the full-resolution convolutional deep neural network to learn image contour features, and outputting the contour sub-graphs; Step S233: fusing the multiple contour sub-images into a full-resolution contour image through downsampling inverse transformation.

6. The method according to claim 5, characterized in that The full-resolution nonlinear mapping layer includes three types of operations; the first type of operation is convolution Conv, the second type of operation is batch normalization BN, and the third type of operation is the rectified linear activation function ReLU.

7. The method according to claim 6, characterized in that The full-resolution nonlinear mapping layer includes K cyclic units, and the cyclic units include three layers. The first layer adopts the Conv+ReLU mode, the second layer adopts the Conv+BN+ReLU mode, and the third layer adopts the pure Conv mode to realize the mapping function from image to contour and realize the output of contour sub-image, where K is a positive integer.

8. The method according to claim 5, characterized in that The reversible downsampling adopts a downsampling rate of 1 / 2, the size of the original contour image in the training contour data set is (M, N), then the number of the contour sub-images is 4, and the size of the contour sub-image is (M / 2)*(N / 2).

9. The method according to claim 6, characterized in that The zero-padding method is used to compensate for the resolution loss caused by the first type of operation.

Images