Semiconductor defect detection method and system based on deep learning
By collecting, preprocessing data in the test of semiconductor chip defects, and using deep learning network models for defect feature extraction and image segmentation, the data labeling time labeling, defect imbad data, and small fuzzy defects in the existing technology are difficult to identify. It has achieved high -efficiency and accurate semiconductor chip defect detection.
Patent Information
- Application Number
- CN202411406341.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2025-01-28
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the detection of defects in semiconductor chips, there are problems such as time-consuming and labor-intensive data imbalance, and difficult to identify small fuzzy defects in the prior art, resulting in missed or missed detection.
Make a semiconductor chip sample picture with a number of high -definition cameras, and perform noise, cleaning, screening and enhanced operations to generate high -quality data sets. Use the first and second defect feature algorithms to extract geometric and gray defect characteristics, and use the defect segmentation algorithm to perform image segmentation, and build a deep learning network model for training and prediction.
The efficient semiconductor chip defect detection has been achieved, the accuracy and efficiency of the detection is improved, the cost and time cost of labor are reduced, and the quality control in the process of semiconductor manufacturing provides reliable tools.
Smart Images
Figure CN119379621A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of semiconductor defect detection, and specifically to a semiconductor defect detection method and system based on deep learning. Background Art
[0002] Semiconductor chips are tiny electronic components made of semiconductor materials, usually in the form of thin sheets. They are used to build integrated circuits, which are one of the core components of modern electronic devices. Semiconductor chip defect detection is used to detect possible defects in the semiconductor chip manufacturing process to ensure the quality and reliability of the final product. As the core component of electronic devices, semiconductor chips may have various defects during their manufacturing process, such as crystal defects, metal contamination, oxides, poor etching, etc. If these defects are not discovered and repaired in time, they may lead to chip performance degradation, increased failure rate, and even product failure.
[0003] Since deep learning models require a large amount of labeled data for training, labeling a large number of semiconductor chip images for semiconductor defect detection may require a lot of human resources and time. In the actual semiconductor manufacturing process, different types of defect data may present an unbalanced distribution, which may lead to the deep learning model's insufficient ability to recognize defects of a few categories. In addition, some particularly small or vague defects may be difficult to be accurately identified by traditional deep learning methods, which may lead to missed detections or false detections. Summary of the invention
[0004] The purpose of the present invention is to provide a semiconductor defect detection method and system based on deep learning to solve the problems raised in the above background technology.
[0005] To achieve the above object, the present invention provides the following technical solution: a semiconductor defect detection method based on deep learning, comprising the following steps:
[0006] S1. Collect semiconductor chip sample images through multiple high-definition cameras to generate high-precision semiconductor chip images, and perform denoising operations on the collected images to obtain the original semiconductor chip data set;
[0007] S2, cleaning and screening the original semiconductor chip data set obtained, removing duplication, blur and shadow operations from the cleaned and screened semiconductor chip data set, and enhancing the sample images in the semiconductor chip data set after removing duplication, blur and shadow;
[0008] S3. Based on the semiconductor chip dataset after cleaning and screening, the enhanced semiconductor chip dataset is divided into four categories according to defect types: ink dot defect, foreign object defect, scratch defect and contamination defect, and a corresponding label is assigned to each defect type. The labeled semiconductor chip dataset is divided into a training set, a validation set and a test set according to the division ratio;
[0009] S4, extracting geometric defect features of sample images in the training set, the validation set, and the test set by using a first defect feature algorithm, and extracting grayscale defect features of sample images in the training set, the validation set, and the test set by using a second defect feature algorithm;
[0010] S5, performing image segmentation on the sample images in the training set, validation set and test set after feature extraction by using a defect segmentation algorithm;
[0011] S6, constructing an initial deep learning network model, and inputting the sample images in the training set after image segmentation into the input end of the initial deep learning network model to train the initial deep learning network model, and obtaining a trained deep learning network model;
[0012] S7. Input the semiconductor chip images in the validation set into the deep learning network model, and the deep learning network model outputs the predicted defect location, defect type and confidence. Input the semiconductor chip images in the test set into the deep learning network model to evaluate the trained learning network model.
[0013] Preferably, the denoising operation on the collected image comprises the following steps:
[0014] A1. Divide the pixels in the sample image into multiple pixel blocks, and determine the flat pixel blocks according to the gradient mean of the pixel values in the pixel blocks. The gradient mean of the pixel values is:
[0015]
[0016] Among them, G represents the gradient mean of pixel values, u0 represents the sample image, represents the gradient magnitude of the sample image, represents the direction of the gradient field, represents the transformed image after equalization, Indicates ladder point;
[0017] A2. Calculate the pixel value variance of pixels in each flat pixel block, and estimate the noise intensity of the image according to the mean of the pixel value variances of all flat pixel blocks and the mean of the pixel values of the pixels in the image;
[0018] A3. Perform pixel conversion compensation on the current pixel value according to the estimated image noise intensity.
[0019] Preferably, the enhancing operation on the sample image in the semiconductor chip data set after removing duplication, blur and shadow comprises the following steps:
[0020] B1. Convert the sample image into a vector image consisting of three components: hue, saturation and brightness;
[0021] B2. Perform Fourier transform on the brightness component to convert the brightness to the frequency domain. The vector image is converted from the time domain to the frequency domain through Fourier transform, and enhancement processing is performed by analyzing the spectrum information of the image.
[0022] B3, iteratively enhancing the brightness component in the frequency domain, and performing inverse Fourier transform on the result obtained in each iteration to convert it back to the time domain;
[0023] B4. Combine the inversely transformed time domain image with the original hue and saturation components, convert it back to a sample image, and calculate various image parameters of the sample image.
[0024] Preferably, the first defect feature algorithm includes a pixel point extraction algorithm, an edge pixel extraction algorithm, a compactness extraction algorithm, a center of gravity coordinate extraction algorithm, a rectangularity extraction algorithm and a duty cycle extraction algorithm.
[0025] Among them, the pixel extraction algorithm is:
[0026]
[0027] Among them, S represents the number of pixels in the defect area, (x, y) represents the coordinates of the pixel, and R d Represents the pixels inside the region;
[0028] The edge pixel extraction algorithm is:
[0029]
[0030] Among them, L represents the number of edge pixels, R b Indicates the region boundary;
[0031] The compactness extraction algorithm is:
[0032]
[0033] Among them, C represents the degree of closeness to a circle;
[0034] The barycenter coordinate extraction algorithm is:
[0035]
[0036] Among them, X c The horizontal coordinate of the center of gravity, Y cThe ordinate represents the center of gravity;
[0037] The rectangular degree extraction algorithm is:
[0038]
[0039] Among them, Rt represents the rectangularity, Ls represents the short side of the minimum circumscribed rectangle, and Lt represents the side length;
[0040] The duty cycle extraction algorithm is:
[0041]
[0042] Among them, R q Indicates the duty cycle.
[0043] Preferably, the second defect feature algorithm includes a grayscale mean extraction algorithm, a grayscale variance extraction algorithm and a grayscale entropy extraction algorithm.
[0044] Among them, the gray mean extraction algorithm is:
[0045]
[0046] Among them, ur represents the grayscale mean, i represents the pixel value, L represents the grayscale level of the grayscale image, and p(i) represents the grayscale value distribution probability;
[0047] The grayscale variance extraction algorithm is:
[0048]
[0049] Among them, V r represents the grayscale variance;
[0050] The grayscale entropy extraction algorithm is:
[0051]
[0052] Among them, G e Represents grayscale entropy.
[0053] Preferably, the defect segmentation algorithm is:
[0054]
[0055] Among them, (i, j) represents the coordinates of the sample image, g(i, j) represents the processed sample image, and f(i, j) represents the original sample image. represents the sample image after image enhancement, T max and T min represents the maximum and minimum thresholds, T represents the dynamic threshold constant, and k represents the image enhancement factor.
[0056] The present invention also provides the following technical solutions: a semiconductor defect detection system based on deep learning, which is applicable to a semiconductor defect detection method based on deep learning, including a data acquisition and preprocessing module, a data annotation and division module, a feature extraction module, an image segmentation module, a deep learning network model training module and a model prediction and evaluation module;
[0057] The data acquisition and preprocessing module is used to collect semiconductor chip sample images through multiple high-definition cameras, and perform denoising, cleaning, screening and enhancement operations to generate a high-quality semiconductor chip image data set;
[0058] The data annotation and division module is used to classify the cleaned and screened high-quality semiconductor chip data set into four categories according to defect types: ink dot defects, foreign matter defects, scratch defects and contamination defects, and assign corresponding labels. Then, the data set is divided into a training set, a validation set and a test set according to the division ratio;
[0059] The feature extraction module is used to extract geometric defect features and grayscale defect features from sample images in a training set, a validation set, and a test set through a first defect feature algorithm and a second defect feature algorithm.
[0060] Preferably, the image segmentation module is used to segment the defect areas of the sample images in the training set, validation set and test set after feature extraction using a defect segmentation algorithm, so that the deep learning network model can identify and classify each defect area;
[0061] The deep learning network model training module is used to construct an initial deep learning network model, and use the sample images in the training set to input the model for training, so as to learn the characteristics of semiconductor defects and optimize the model parameters;
[0062] The model prediction and evaluation module is used to input the semiconductor chip images in the validation set into the trained deep learning network model, and to predict the defect location, defect type and confidence. At the same time, the images in the test set are used to evaluate the performance and generalization ability of the trained deep learning network model.
[0063] Compared with the prior art, the present invention has the following beneficial effects:
[0064] The present invention collects images with multiple high-definition cameras and performs denoising and enhancement operations on the images, thereby obtaining a high-quality and high-precision semiconductor chip image dataset, which is beneficial to the accuracy of subsequent defect identification. By cleaning and screening the dataset and annotating the defect types of sample images, the present invention realizes the automatic processing of large-scale datasets, thus saving labor costs and time costs.
[0065] The present invention extracts geometric defect features and grayscale defect features through the first defect feature algorithm and the second defect feature algorithm, and uses the defect segmentation algorithm to segment the sample image, which helps the deep learning network model to more accurately identify and classify defects. The deep learning network model is used to train the training set so that it learns the characteristics of semiconductor defects and can predict and evaluate the defect location, type and confidence on the verification set and the test set, providing a reliable tool for quality control in the semiconductor manufacturing process. BRIEF DESCRIPTION OF THE DRAWINGS
[0066] Figure 1 A schematic diagram of the overall method flow provided by an embodiment of the present invention;
[0067] Figure 2 A schematic diagram of a denoising operation process provided by an embodiment of the present invention;
[0068] Figure 3 A schematic diagram of a flow chart of an enhanced operation method provided by an embodiment of the present invention;
[0069] Figure 4 A schematic diagram of the overall system flow provided by an embodiment of the present invention.
[0070] In the figure: 1. Data acquisition and preprocessing module; 2. Data labeling and division module; 3. Feature extraction module; 4. Image segmentation module; 5. Deep learning network model training module; 6. Model prediction and evaluation module. DETAILED DESCRIPTION
[0071] 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.
[0072] Embodiment 1
[0073] See also Figure 1-3 The present invention provides a technical solution: a semiconductor defect detection method based on deep learning, comprising the following steps:
[0074] S1. Collect semiconductor chip sample images through multiple high-definition cameras to generate high-precision semiconductor chip images, and perform denoising operations on the collected images to obtain the original semiconductor chip data set;
[0075] S2, cleaning and screening the original semiconductor chip data set obtained, removing duplication, blur and shadow operations from the cleaned and screened semiconductor chip data set, and enhancing the sample images in the semiconductor chip data set after removing duplication, blur and shadow;
[0076] S3. Based on the semiconductor chip dataset after cleaning and screening, the enhanced semiconductor chip dataset is divided into four categories according to defect types: ink dot defect, foreign object defect, scratch defect and contamination defect, and a corresponding label is assigned to each defect type. The labeled semiconductor chip dataset is divided into a training set, a validation set and a test set according to the division ratio;
[0077] S4, extracting geometric defect features of sample images in the training set, the validation set, and the test set by using a first defect feature algorithm, and extracting grayscale defect features of sample images in the training set, the validation set, and the test set by using a second defect feature algorithm;
[0078] S5, performing image segmentation on the sample images in the training set, validation set and test set after feature extraction by using a defect segmentation algorithm;
[0079] S6, constructing an initial deep learning network model, and inputting the sample images in the training set after image segmentation into the input end of the initial deep learning network model to train the initial deep learning network model, and obtaining a trained deep learning network model;
[0080] S7. Input the semiconductor chip images in the validation set into the deep learning network model, and the deep learning network model outputs the predicted defect location, defect type and confidence. Input the semiconductor chip images in the test set into the deep learning network model to evaluate the trained learning network model.
[0081] The denoising operation is performed on the collected image, including the following steps:
[0082] A1. Divide the pixels in the sample image into multiple pixel blocks, and determine the flat pixel blocks according to the gradient mean of the pixel values in the pixel blocks. The gradient mean of the pixel values is:
[0083]
[0084] Among them, G represents the gradient mean of pixel values, u0 represents the sample image, represents the gradient magnitude of the sample image, represents the direction of the gradient field, represents the transformed image after equalization, Indicates ladder point;
[0085] A2. Calculate the pixel value variance of pixels in each flat pixel block, and estimate the noise intensity of the image according to the mean of the pixel value variances of all flat pixel blocks and the mean of the pixel values of the pixels in the image;
[0086] A3, performing pixel conversion compensation on the current pixel value according to the estimated image noise intensity;
[0087] The sample images in the semiconductor chip dataset are enhanced, including the following steps:
[0088] B1. Convert the sample image into a vector image consisting of three components: hue, saturation and brightness;
[0089] B2. Perform Fourier transform on the brightness component to convert the brightness to the frequency domain. The vector image is converted from the time domain to the frequency domain through Fourier transform, and enhancement processing is performed by analyzing the spectrum information of the image.
[0090] B3, iteratively enhancing the brightness component in the frequency domain, and performing inverse Fourier transform on the result obtained in each iteration to convert it back to the time domain;
[0091] B4, combining the inverse transformed time domain image with the original hue and saturation components, converting it back to a sample image, and calculating various image parameters of the sample image;
[0092] The first defect feature algorithm includes a pixel point extraction algorithm, an edge pixel extraction algorithm, a compactness extraction algorithm, a center of gravity coordinate extraction algorithm, a rectangularity extraction algorithm and a duty cycle extraction algorithm.
[0093] Among them, the pixel extraction algorithm is:
[0094]
[0095] Among them, S represents the number of pixels in the defect area, (x, y) represents the coordinates of the pixel, and R d Represents the pixels inside the region;
[0096] The edge pixel extraction algorithm is:
[0097]
[0098] Among them, L represents the number of edge pixels, R b Indicates the region boundary;
[0099] The compactness extraction algorithm is:
[0100]
[0101] Among them, C represents the degree of closeness to a circle;
[0102] The barycenter coordinate extraction algorithm is:
[0103]
[0104] Among them, X c The horizontal coordinate of the center of gravity, Y c The ordinate represents the center of gravity;
[0105] The rectangular degree extraction algorithm is:
[0106]
[0107] Among them, Rt represents the rectangularity, Ls represents the short side of the minimum circumscribed rectangle, and Lt represents the side length;
[0108] The duty cycle extraction algorithm is:
[0109]
[0110] Among them, R q Indicates duty cycle;
[0111] The second defect feature algorithm includes grayscale mean extraction algorithm, grayscale variance extraction algorithm and grayscale entropy extraction algorithm.
[0112] Among them, the gray mean extraction algorithm is:
[0113]
[0114] Among them, ur represents the grayscale mean, i represents the pixel value, L represents the grayscale level of the grayscale image, and p(i) represents the grayscale value distribution probability;
[0115] The grayscale variance extraction algorithm is:
[0116]
[0117] Among them, V r represents the grayscale variance;
[0118] The grayscale entropy extraction algorithm is:
[0119]
[0120] Among them, G e represents grayscale entropy;
[0121] The defect segmentation algorithm is:
[0122]
[0123] Among them, (i, j) represents the coordinates of the sample image, g(i, j) represents the processed sample image, and f(i, j) represents the original sample image. represents the sample image after image enhancement, T max and T min represents the maximum and minimum thresholds, T represents the dynamic threshold constant, and k represents the image enhancement factor.
[0124] Embodiment 2
[0125] See also Figure 4 , the present invention also provides the following technical solutions: a semiconductor defect detection system based on deep learning, which is applicable to a semiconductor defect detection method based on deep learning, including a data acquisition and preprocessing module 1, a data annotation and division module 2, a feature extraction module 3, an image segmentation module 4, a deep learning network model training module 5 and a model prediction and evaluation module 6;
[0126] The data acquisition and preprocessing module 1 is used to collect semiconductor chip sample images through multiple high-definition cameras, and perform denoising, cleaning, screening and enhancement operations to generate a high-quality semiconductor chip image data set;
[0127] The data labeling and partitioning module 2 is used to classify the cleaned and screened high-quality semiconductor chip data set into four categories according to defect types: ink dot defects, foreign matter defects, scratch defects, and contamination defects, and assign corresponding labels, and then divide the data set into a training set, a validation set, and a test set according to the partitioning ratio;
[0128] The feature extraction module 3 is used to extract geometric defect features and grayscale defect features from sample images in the training set, the validation set and the test set through a first defect feature algorithm and a second defect feature algorithm;
[0129] The image segmentation module 4 is used to segment the defect areas of the sample images in the training set, the validation set and the test set after feature extraction using a defect segmentation algorithm, so that the deep learning network model can identify and classify each defect area;
[0130] The deep learning network model training module 5 is used to construct an initial deep learning network model and use the sample images in the training set to input the model for training, so as to learn the characteristics of semiconductor defects and optimize the model parameters;
[0131] The model prediction and evaluation module 6 is used to input the semiconductor chip images in the validation set into the trained deep learning network model, and to predict the defect location, defect type and confidence. At the same time, the images in the test set are used to evaluate the performance and generalization ability of the trained deep learning network model.
[0132] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0133] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A semiconductor defect detection method based on deep learning, characterized in that: The following steps are involved: S1. Collect semiconductor chip sample images through multiple high-definition cameras to generate high-precision semiconductor chip images, and perform denoising operations on the collected images to obtain the original semiconductor chip data set; S2, cleaning and screening the original semiconductor chip data set obtained, removing duplication, blur and shadow operations from the cleaned and screened semiconductor chip data set, and enhancing the sample images in the semiconductor chip data set after removing duplication, blur and shadow; S3. Based on the semiconductor chip dataset after cleaning and screening, the enhanced semiconductor chip dataset is divided into four categories according to defect types: ink dot defect, foreign object defect, scratch defect and contamination defect, and a corresponding label is assigned to each defect type. The labeled semiconductor chip dataset is divided into a training set, a validation set and a test set according to the division ratio; S4, extracting geometric defect features of sample images in the training set, the validation set, and the test set by using a first defect feature algorithm, and extracting grayscale defect features of sample images in the training set, the validation set, and the test set by using a second defect feature algorithm; S5, performing image segmentation on the sample images in the training set, validation set and test set after feature extraction by using a defect segmentation algorithm; S6, constructing an initial deep learning network model, and inputting the sample images in the training set after image segmentation into the input end of the initial deep learning network model to train the initial deep learning network model, and obtaining a trained deep learning network model; S7. Input the semiconductor chip images in the validation set into the deep learning network model, and the deep learning network model outputs the predicted defect location, defect type and confidence. Input the semiconductor chip images in the test set into the deep learning network model to evaluate the trained learning network model.
2. The semiconductor defect detection method based on deep learning according to claim 1, characterized in that: The denoising operation on the collected image comprises the following steps: A1. Divide the pixels in the sample image into multiple pixel blocks, and determine the flat pixel blocks according to the gradient mean of the pixel values in the pixel blocks. The gradient mean of the pixel values is: Among them, G represents the gradient mean of the pixel value, u0 represents the sample image, |▽u0| represents the gradient modulus of the sample image, represents the direction of the gradient field, L(|▽u0|) represents the transformed image after equalization, and ▽ represents the gradient point; A2. Calculate the pixel value variance of pixels in each flat pixel block, and estimate the noise intensity of the image according to the mean of the pixel value variances of all flat pixel blocks and the mean of the pixel values of the pixels in the image; A3. Perform pixel conversion compensation on the current pixel value according to the estimated image noise intensity.
3. The semiconductor defect detection method based on deep learning according to claim 1, characterized in that: The method of performing an enhancement operation on the sample image in the semiconductor chip data set after removing duplication, blur and shadow comprises the following steps: B1. Convert the sample image into a vector image consisting of three components: hue, saturation and brightness; B2. Perform Fourier transform on the brightness component to convert the brightness to the frequency domain. The vector image is converted from the time domain to the frequency domain through Fourier transform, and enhancement processing is performed by analyzing the spectrum information of the image. B3, iteratively enhancing the brightness component in the frequency domain, and performing inverse Fourier transform on the result obtained in each iteration to convert it back to the time domain; B4. Combine the inversely transformed time domain image with the original hue and saturation components, convert it back to a sample image, and calculate various image parameters of the sample image.
4. The semiconductor defect detection method based on deep learning according to claim 1, characterized in that: The first defect feature algorithm includes a pixel point extraction algorithm, an edge pixel extraction algorithm, a compactness extraction algorithm, a center of gravity coordinate extraction algorithm, a rectangularity extraction algorithm and a duty cycle extraction algorithm. Among them, the pixel extraction algorithm is: Among them, S represents the number of pixels in the defect area, (x, y) represents the coordinates of the pixel, and R d Represents the pixels inside the region; The edge pixel extraction algorithm is: Among them, L represents the number of edge pixels, R b Indicates the region boundary; The compactness extraction algorithm is: Among them, C represents the degree of closeness to a circle; The barycentric coordinate extraction algorithm is: Among them, X c The horizontal coordinate of the center of gravity, Y c The ordinate represents the center of gravity; The rectangular degree extraction algorithm is: Among them, Rt represents the rectangularity, Ls represents the short side of the minimum circumscribed rectangle, and Lt represents the side length; The duty cycle extraction algorithm is: Among them, R q Indicates duty cycle.
5. The semiconductor defect detection method based on deep learning according to claim 1, characterized in that: The second defect feature algorithm includes a grayscale mean extraction algorithm, a grayscale variance extraction algorithm and a grayscale entropy extraction algorithm. Among them, the gray mean extraction algorithm is: Among them, ur represents the grayscale mean, i represents the pixel value, L represents the grayscale level of the grayscale image, and p(i) represents the grayscale value distribution probability; The grayscale variance extraction algorithm is: Among them, V r represents the grayscale variance; The grayscale entropy extraction algorithm is: Among them, G e Represents grayscale entropy.
6. The semiconductor defect detection method based on deep learning according to claim 1, characterized in that: The defect segmentation algorithm is: Among them, (i, j) represents the coordinates of the sample image, g(i, j) represents the processed sample image, and f(i, j) represents the original sample image. represents the sample image after image enhancement, T max and T min represents the maximum and minimum thresholds, T represents the dynamic threshold constant, and k represents the image enhancement factor.
7. A semiconductor defect detection system based on deep learning, which is applicable to a semiconductor defect detection method based on deep learning as described in any one of claims 1 to 6, comprising a data acquisition and preprocessing module (1), a data annotation and division module (2), a feature extraction module (3), an image segmentation module (4), a deep learning network model training module (5) and a model prediction and evaluation module (6), characterized in that: The data acquisition and preprocessing module (1) is used to acquire semiconductor chip sample images through multiple high-definition cameras, and perform denoising, cleaning, screening and enhancement operations to generate a high-quality semiconductor chip image data set; The data labeling and partitioning module (2) is used to classify the cleaned and screened high-quality semiconductor chip data set into four categories according to defect types: ink dot defects, foreign matter defects, scratch defects and contamination defects, and assign corresponding labels, and then divide the data set into a training set, a validation set and a test set according to the partitioning ratio; The feature extraction module (3) is used to extract geometric defect features and grayscale defect features from sample images in a training set, a validation set and a test set by using a first defect feature algorithm and a second defect feature algorithm.
8. The semiconductor defect detection system based on deep learning according to claim 7, characterized in that: The image segmentation module (4) is used to use a defect segmentation algorithm to segment defect areas of sample images in the training set, validation set and test set after feature extraction, so that the deep learning network model can identify and classify each defect area; The deep learning network model training module (5) is used to construct an initial deep learning network model, and use the sample images in the training set to input the model for training, so as to learn the characteristics of semiconductor defects and optimize the model parameters; The model prediction and evaluation module (6) is used to input the semiconductor chip images in the verification set into the trained deep learning network model, and to predict the defect location, defect type and confidence. At the same time, the images in the test set are used to evaluate the performance and generalization ability of the trained deep learning network model.