Search

CN-122023310-A - Wafer defect detection method and system based on attention guide network

CN122023310ACN 122023310 ACN122023310 ACN 122023310ACN-122023310-A

Abstract

The invention discloses a wafer defect detection method and a wafer defect detection system based on an attention guide network, which belong to the technical field of wafer defect detection and comprise the following steps of firstly, adjusting light supplementing according to detection requirements, and then acquiring wafer images; the method comprises the steps of collecting wafer images, carrying out spectrum enhancement pretreatment on the collected wafer images, extracting the characteristics of the enhanced images through a multi-scale distributed characteristic extraction backbone network, and fusing the extracted characteristics through a attention-guided characteristic pyramid network. The method effectively inhibits the complex background interference of the wafer, remarkably enhances the defect characteristics, constructs a characteristic extraction mechanism capable of capturing local details and long-range context information at the same time, realizes balanced and accurate detection of multi-scale defects, embeds the physical prior of the defects into a network in a learnable manner, improves learning efficiency and generalization, maintains lower model complexity while guaranteeing ultrahigh precision, and meets the real-time requirement of a production line.

Inventors

  • ZHOU CAIJIAN
  • XU QINGYANG
  • YU JIAHUI
  • ZHAO SHUWEN
  • ZENG DONGSHENG
  • CHEN AN
  • CHEN ENZAN

Assignees

  • 杭州汇萃智能科技有限公司
  • 广东广源智能科技有限公司

Dates

Publication Date
20260512
Application Date
20260124

Claims (10)

  1. 1. The wafer defect detection method based on the attention guidance network is characterized by comprising the following steps of: the method comprises the steps of firstly, adjusting light supplementing according to detection requirements, and then collecting wafer images; performing spectrum enhancement pretreatment on the collected wafer image; And thirdly, extracting the enhanced image features through a multi-scale distributed feature extraction backbone network, fusing the extracted features through a feature pyramid network guided by attention, calculating total loss, judging whether the total loss is larger than a defect threshold, if so, judging that the wafer surface has defects, and if not, judging that the wafer surface has no defects.
  2. 2. The method according to claim 1, wherein the second specific operation is as follows: step 21, gray scale processing is carried out on the collected wafer image to obtain a gray scale image ; Step 22. For gray scale image Performing logarithmic Fourier transform to obtain converted data ; Step 23, converting the diagram Input to a learnable spectrum suppression filter for spectrum suppression to obtain suppression data ; Step 24, suppressing data Obtaining an enhanced image after background inhibition through inverse Fourier transform and exponential operation 。
  3. 3. The method according to claim 2, wherein the third step is performed as follows: Step 31, extracting the enhanced image features through a multi-scale distributed feature extraction backbone network to obtain shallow high-resolution features And deep low resolution features ; Step 32 shallow high resolution features And deep low resolution features Inputting the features to a feature pyramid network for directing attention to perform feature fusion; And 33, calculating total loss through the fused characteristics, judging whether the total loss is larger than a defect threshold, if yes, determining that the surface of the wafer has defects, and if no, determining that the surface of the wafer has no defects.
  4. 4. The detection method according to claim 3, wherein the multi-scale distributed feature extraction backbone network is provided with four groups of network layers Stage1, stage2, stage3 and Stage4, and feature map sizes of Stage1, stage2, stage3 and Stage4 are respectively And The void fraction of Stage1 was 2, And 4, the void ratios of stage2 were 2, 4, and 6, the void ratios of stage3 were 4, 6, and 8, and the void ratios of stage4 were 6, 8, and 12.
  5. 5. The method according to claim 4, wherein the step 31 specifically comprises the following steps: Step 311: enhancing the image Downsampling the convolution kernel with step length of 2 input to the first convolution layer to obtain Feature map of dimensions ; Step 312, re-combining Feature map of dimensions Is input into the network layer Stage1, and is subjected to 3×3 convolution processing by using void ratios 1, 2 and 4 to obtain Shallow high resolution feature of size ; Step 313, convolution check with step 2 Shallow high resolution feature of size Downsampling to obtain Is characterized by (a) Feature map then Input to network layer Stage2, and 3×3 convolution processing with void ratios 2, 4, and 6 to obtain Feature map of dimensions ; Step 314, convolution check with step 2 Is characterized by (a) Downsampling to obtain Is characterized by (a) Then input to the network layer Stage3, and the 3×3 convolution processing with the void ratios of 4, 6 and 8 is used to obtain Feature map of dimensions ; Step 315, using a convolution kernel with step size 2 to obtain Feature map of dimensions Downsampling to obtain Is characterized by (a) Feature map Input to network layer Stage4, network layer Stage4 uses 3×3 convolution processing of void ratios 6, 8, and 12 to obtain Deep low resolution features of size 。
  6. 6. The method according to claim 5, wherein the step 32 specifically comprises the following steps: step 321, according to shallow high resolution features And deep low resolution features Computing a spatial attention map ; Step 322, separately characterizing from shallow high resolution features over three parallel lightweight subnetworks Extracting channel attention vectors related to linearity, circularity, and irregularity Then carrying out feature fusion to obtain fusion features 。
  7. 7. The method of claim 6, wherein the features are fused The calculation is as follows: ; ; Wherein, the As an intermediate feature map, a plurality of feature images are displayed, For the concatenation and fusion of the channel dimensions, As a result of the scale factor being a learnable, The initial value of (2) is 1.0, and the value range is 0.5-1.5.
  8. 8. The inspection method of claim 7, wherein the lightweight subnetworks are a linear defect extraction network, a circularity defect extraction network, and an irregularity defect extraction network, respectively.
  9. 9. The method of claim 8, wherein the specific operation of step 33 is as follows: step 331, fusing features Inputting the probability distribution of crack defects, bubble defects, sundry defects and backgrounds in the decoupling detection head structure, and predicting the position offset of a boundary frame of the regression branch of the decoupling detection head structure by the classification branch of the decoupling detection head structure; Step 332, calculating the classified branch loss and the regression branch loss, carrying out weighted summation on the classified branch loss and the regression branch loss to obtain total loss, judging whether the total loss is larger than a defect threshold, if yes, determining that the surface of the wafer has defects, and if no, determining that the surface of the wafer has no defects.
  10. 10. An attention-based guidance network wafer defect detection system, comprising: the image acquisition module is used for acquiring images; the multi-mode mixed lighting module is used for adjusting light supplementing according to detection requirements; the preprocessing module is used for carrying out spectrum enhancement preprocessing on the acquired wafer image; And the fusion calculation module is used for extracting the enhanced image features through the multi-scale distributed feature extraction backbone network, fusing the extracted features through the attention-guided feature pyramid network, calculating total loss, judging whether the total loss is larger than a defect threshold, if so, determining that the wafer surface has defects, and if not, determining that the wafer surface has no defects.

Description

Wafer defect detection method and system based on attention guide network Technical Field The invention relates to the technical field of wafer defect detection, in particular to a wafer defect detection method and system based on an attention guide network. Background The wafer is used as a base stone in the semiconductor industry, and its surface quality directly determines the performance, yield and reliability of the chip. In the manufacturing process, the wafer surface is extremely easy to generate micro-scale or nano-scale defects, and mainly comprises geometric cracks, bubbles with specific optical scattering characteristics and impurities with different forms. These defects can damage the integrity of the circuit structure, leading to device shorts, leakage, or complete failure. Therefore, 100% of total inspection is performed on the wafer in the production line, and the defects are accurately identified and classified, so that the method is a key link for improving the product yield and reducing the production cost; the wafer inspection adopts the following method: Detection method based on frequency domain filtering and manual feature extraction. The method has the advantages of clear principle and certain effect on the wafer with strong periodic background. The method has the advantages that the design of the filter depends on priori knowledge, the wafer patterns of different production lines and different process nodes are difficult to adapt, the effect of the filter on the non-periodic or weak periodic background area is rapidly reduced, the manual characteristic has insufficient defect characterization capability on tiny, fuzzy or low contrast with the background, and the generalization performance is poor. Detection methods based on standard CNN target detectors. The method can directly predict the boundary box containing cracks, bubbles and sundries and the category confidence of the boundary box. The method avoids complex manual feature design and realizes end-to-end detection to a certain extent. However, its direct application to wafer defect inspection scenarios faces the serious challenge of, first, the local receptive field limitation of standard convolution. Wafer defects, especially micro-cracks and low contrast bubbles, may have significant features spread over a large area, and standard convolution kernels (e.g., 3x3, 5x 5) have difficulty capturing this long-range dependence, resulting in inadequate feature extraction. Second, complex background and scale differences. The circuit pattern on the wafer surface constitutes an extremely complex background noise, while the dimensions of defects are extremely wide, ranging from tiny dust at several pixels to cracks across hundreds of pixels. Standard CNNs tend to lose detailed information of small defects during pooling, while capturing insufficient morphological context information of large defects. Finally, the computational complexity contradicts the real-time performance. To achieve high accuracy, deeper networks and higher image resolution are often required, which is contrary to the real-time detection requirements of semiconductor production lines for high Throughput (Throughput). Detection methods based on specific optical imaging. The method improves the visibility of the defects on a physical level through a hardware means, and then combines the algorithm to detect. The scheme can effectively improve the detection rate of specific types of defects, but has the defects of high system cost and complex configuration, and generally one optical configuration is only sensitive to certain types of defects, so that the defects with different physical characteristics, namely cracks, bubbles and sundries, are difficult to detect optimally at the same time by using a uniform imaging mode. Based on the above, the present invention designs a wafer defect detection method and system based on an attention-guided network to solve the above-mentioned problems. Disclosure of Invention In view of the above drawbacks of the prior art, the present invention provides a wafer defect detection method and system based on an attention-directed network. In order to achieve the above purpose, the invention is realized by the following technical scheme: The wafer defect detection method based on the attention guidance network comprises the following steps: the method comprises the steps of firstly, adjusting light supplementing according to detection requirements, and then collecting wafer images; performing spectrum enhancement pretreatment on the collected wafer image; And thirdly, extracting the enhanced image features through a multi-scale distributed feature extraction backbone network, fusing the extracted features through a feature pyramid network guided by attention, calculating total loss, judging whether the total loss is larger than a defect threshold, if so, judging that the wafer surface has defects, and if not, judging that the wafer surface