A multi-channel image fusion defect enhancement method and related products

Abstract

The application discloses a kind of multi-channel image fusion defect enhancement method and related products, which can be applied to the field of semiconductor technology, and the method comprises the following steps: obtaining the target chip defect map corresponding to each physical imaging channel respectively, the target left adjacent chip reference map and the target right adjacent chip reference map of target defect point position;Image intensity stretching processing is carried out on the above obtained image, and the target composite tensor is constructed based on the image obtained after imaging intensity stretching processing;The target composite tensor is input into the pre-trained target UNET model, and the target defect difference map for wafer defect detection is obtained.Such, the target chip defect map after imaging intensity stretching processing, the target left adjacent chip reference map and the target right adjacent chip reference map are used to construct the target composite tensor, and the tensor is input into the pre-trained target UNET model to obtain the target defect difference map for wafer defect detection, which improves the detection accuracy and signal-to-noise ratio of wafer defect.

Description

Technical Field

[0001] This application relates to the field of semiconductor technology, and in particular to a method for enhancing defects in multi-channel image fusion and related products. Background Technology

[0002] In the design of image-based dark-field defect detection algorithms, the die-to-die (Die2Die) imaging-based defect detection algorithm is currently the mainstream design scheme. This scheme simultaneously acquires defect images of the target chip and reference images of its left and right adjacent chips, and identifies abnormal defects on the wafer surface by using differential operations between the defect image of the target chip and the reference images of the adjacent chips.

[0003] However, in actual Die-to-Die imaging, subtle differences in process technology and optical imaging conditions exist between adjacent chips, making background residue highly susceptible to occur after differential operations. Existing Die-to-Die defect detection methods only generate differential maps within a single channel and employ fixed threshold segmentation, failing to fully utilize the feature differences between multi-channel images. When the grayscale of the differential residue is close to that of the defect signal, threshold segmentation cannot distinguish between background residue and real defects, and minute defects are easily falsely filtered out, ultimately leading to a decrease in the detection accuracy and signal-to-noise ratio of wafer defects.

[0004] Therefore, improving the detection accuracy and signal-to-noise ratio of wafer defects is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a multi-channel image fusion defect enhancement method and related products. The method constructs a target composite tensor using a target chip defect map after imaging intensity stretching, a reference map of the target's left adjacent chip, and a reference map of the target's right adjacent chip. This tensor is then input into a pre-trained target U-shaped convolutional neural network (UNET) model to obtain a target defect difference map for wafer defect detection, thereby improving the detection accuracy and signal-to-noise ratio of wafer defects.

[0006] In a first aspect, embodiments of this application provide a multi-channel image fusion defect enhancement method for wafer die-to-die mode defect detection, comprising: Obtain the target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target corresponding to the target defect location in each physical imaging channel; The target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target are subjected to imaging intensity stretching processing; A target composite tensor is constructed based on the target chip defect map after imaging intensity stretching, the target left adjacent chip reference map, and the target right adjacent chip reference map; the target composite tensor includes: the target chip defect map of each physical imaging channel, the differential fusion map obtained by operating the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map inside each physical imaging channel, and the target reference template map obtained by taking the median value of the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels; The target composite tensor is input into the pre-trained target UNET model to obtain the target defect difference map for wafer defect detection.

[0007] Optionally, the differential fusion graph can be generated in the following ways: Within the same physical imaging channel, the target chip defect map after image intensity stretching is compared with the reference map of the adjacent chip on the left side of the target to obtain the first difference map; A second difference image is obtained by performing a difference operation between the target chip defect image after the imaging intensity stretching process and the reference image of the adjacent chip to the right of the target; The first difference map and the second difference map are superimposed and fused to obtain the difference fused map corresponding to each physical imaging channel.

[0008] Optionally, the target reference template diagram can be generated in the following ways: Obtain reference images of the target's left-side adjacent chips and the target's right-side adjacent chips corresponding to all physical imaging channels; Pixel-level median calculation is performed on all the reference images of the target's left adjacent chip and the target's right adjacent chip to obtain the target reference template image.

[0009] Optionally, the number of physical imaging channels is at least two.

[0010] Optionally, the target UNET model is trained using the following method: Obtain the training chip defect map, the training chip reference map to the left of the adjacent chip, and the training chip reference map to the right of the adjacent chip in each physical imaging channel; the training defect map is marked with a truth value label. The defect map of the training chip, the reference map of the left adjacent chip, and the reference map of the right adjacent chip are subjected to imaging intensity stretching processing. A training composite tensor is constructed based on the training chip defect map after imaging intensity stretching, the training left adjacent chip reference map, and the training right adjacent chip reference map. By combining the aforementioned training composite tensor, the initial UNET model is trained using the target loss function to obtain the target UNET model.

[0011] Optionally, the target loss function includes: a background zero loss term, a defect fidelity loss term, and a total variational loss term.

[0012] Optionally, the step of performing imaging intensity stretching processing on the target chip defect map, the reference map of the chip adjacent to the left of the target, and the reference map of the chip adjacent to the right of the target includes: The imaging intensity of the target chip defect map, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip is stretched from a first gray level to a second gray level. The second gray level is higher than the first gray level.

[0013] Secondly, embodiments of this application provide a multi-channel image fusion defect enhancement device for wafer die-to-die mode defect detection, comprising: The acquisition module is used to acquire the target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target corresponding to the target defect point in each physical imaging channel. The stretching module is used to perform imaging intensity stretching processing on the target chip defect map, the reference map of the target adjacent chip to the left and the reference map of the target adjacent chip to the right. A construction module is used to construct a target composite tensor based on the target chip defect map after imaging intensity stretching, the target left adjacent chip reference map, and the target right adjacent chip reference map; the target composite tensor includes: the target chip defect map of each physical imaging channel, the differential fusion map obtained by operating the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map inside each physical imaging channel, and the target reference template map obtained by taking the median value of the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels; The processing module is used to input the target composite tensor into the pre-trained target UNET model and obtain the target defect difference map for wafer defect detection.

[0014] Thirdly, embodiments of this application provide a multi-channel image fusion defect enhancement device, comprising: Memory, used to store computer programs; A processor is configured to implement the steps of the multi-channel image fusion defect enhancement method described above when executing the computer program.

[0015] Fourthly, embodiments of this application provide a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the multi-channel image fusion defect enhancement method described above.

[0016] As can be seen from the above technical solutions, compared with the prior art, this application has the following advantages: This application provides a multi-channel image fusion defect enhancement method for wafer Die2Die mode defect detection. First, it acquires the target chip defect image, the target left adjacent chip reference image, and the target right adjacent chip reference image corresponding to the target defect location in each physical imaging channel. Then, it performs image intensity stretching processing on the target chip defect image, the target left adjacent chip reference image, and the target right adjacent chip reference image, and constructs a target composite tensor based on the stretched image. This target composite tensor includes: the target chip defect image of each physical imaging channel; a differential fusion image obtained within each physical imaging channel by operations on the target chip defect image and the corresponding target left adjacent chip reference image and target right adjacent chip reference image; and a target reference template image obtained by taking the median value from the target left adjacent chip reference image and target right adjacent chip reference image of all physical imaging channels. Finally, the target composite tensor is input into a pre-trained target UNET model to obtain a target defect differential image for wafer defect detection. Thus, a target composite tensor is constructed using the target chip defect map after imaging intensity stretching, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip. This tensor is then input into a pre-trained target UNET model. By relying on the characteristics of strong cross-channel correlation of defects and weak cross-channel correlation of differential residues, background texture interference is suppressed and weak defect details are fully preserved. Finally, a target defect differential map for wafer defect detection is obtained, which improves the detection accuracy and signal-to-noise ratio of wafer defects. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a multi-channel image fusion defect enhancement method provided in this application embodiment; Figure 2 This is a schematic diagram illustrating the effect comparison of an embodiment of this application; Figure 3 A schematic diagram of a tensor matrix provided for an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a UNET model provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a multi-channel image fusion defect enhancement device provided in an embodiment of this application. Detailed Implementation

[0018] As mentioned earlier, existing Die-to-Die defect detection methods suffer from insufficient detection accuracy and signal-to-noise ratio (SNR) for wafer defects. Specifically, in actual Die-to-Die imaging, adjacent chips exhibit subtle differences in process technology and optical imaging conditions, easily leading to background residue after differential operations. Existing Die-to-Die defect detection methods only generate differential maps within a single channel and employ fixed threshold segmentation, failing to fully utilize the feature differences between multi-channel images. When the grayscale of the differential residue and the defect signal are close, threshold segmentation cannot distinguish between background residue and true defects, and minute defects are easily misfiltered, ultimately resulting in a decrease in the detection accuracy and SNR of wafer defects.

[0019] To address the aforementioned issues, this application provides a multi-channel image fusion defect enhancement method. First, it acquires the target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map corresponding to the target defect location in each physical imaging channel. Then, it performs imaging intensity stretching processing on the target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map, and constructs a target composite tensor based on the stretched target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map. This target composite tensor includes: the target chip defect map of each physical imaging channel; a differential fusion map obtained within each physical imaging channel by operations on the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map; and a target reference template map obtained by taking the median value from the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels. Finally, the target composite tensor is input into a pre-trained target UNET model to obtain a target defect differential map for wafer defect detection.

[0020] Thus, a target composite tensor is constructed using the target chip defect map after imaging intensity stretching, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip. This tensor is then input into a pre-trained target UNET model. By relying on the characteristics of strong cross-channel correlation of defects and weak cross-channel correlation of differential residues, background texture interference is suppressed and weak defect details are fully preserved. Finally, a target defect differential map for wafer defect detection is obtained, which improves the detection accuracy and signal-to-noise ratio of wafer defects.

[0021] It should be noted that the multi-channel image fusion defect enhancement method and related products provided in this application can be applied to the field of semiconductor technology. The above are merely examples and do not limit the application field of the multi-channel image fusion defect enhancement method and related products provided in this application.

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0023] Figure 1 This is a flowchart illustrating a multi-channel image fusion defect enhancement method provided in an embodiment of this application. (Combined with...) Figure 1 As shown in the embodiment of this application, a multi-channel image fusion defect enhancement method is provided for wafer die2die mode defect detection, which may include: S101: Obtain the target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target corresponding to the target defect point in each physical imaging channel.

[0024] The number of physical imaging channels is at least two.

[0025] In practical applications, wafer inspection equipment has the capability to detect defects in multiple physical imaging channels. To ensure the final defect enhancement effect, this solution requires at least two physical imaging channels, with different optical imaging conditions used in different channels. The true defect signal will appear synchronously in all channels, while background difference residues caused by process deviations and uneven illumination will only appear sporadically in a few channels. The wafer inspection equipment can determine multiple defect locations on the target wafer through a pre-processing algorithm. A defect location refers to the center coordinates of the suspected defect location on the target wafer; the corresponding defect may be a true detection or a false detection. Furthermore, for each target defect location in its respective physical imaging channel, according to the Die2Die algorithm, there exists a corresponding target chip defect image, a reference image of the target's left adjacent chip, and a reference image of the target's right adjacent chip. To construct the target composite tensor, this solution first needs to obtain the target chip defect image, the target's left adjacent chip reference image, and the target's right adjacent chip reference image corresponding to each physical imaging channel. Thus, by setting up at least two physical imaging channels and simultaneously acquiring defect images and reference images corresponding to the target defect points in each physical imaging channel, multi-dimensional imaging information can be obtained. This fully utilizes the imaging principle that "defects have strong correlations in multiple channels, but weak correlations between differential residual channels," providing a complete data source for subsequent construction of the target composite tensor. Compared to traditional methods that only use single-channel images, this approach can better distinguish between minute defects and background interference, improving the recognition effect of small, low-contrast defects.

[0026] S102: Perform imaging intensity stretching processing on the target chip defect map, the reference map of the target adjacent chip to the left, and the reference map of the target adjacent chip to the right.

[0027] In practical applications, due to the varying optical conditions of different physical imaging channels, the brightness and grayscale dynamic range of the target chip defect images, the reference images of the adjacent chips to the left and right of the target acquired under different physical imaging channels will differ significantly. Images captured by some channels will be generally dark, suppressing defect signals. If these images with inconsistent brightness are directly used for difference operations and tensor construction, additional errors will be introduced due to the brightness deviation between the various physical imaging channels, weakening defect features and affecting the subsequent model's ability to distinguish between real defects and background residue. Therefore, this solution performs image intensity stretching processing on the above three types of images, adjusting the grayscale dynamic range of each image to eliminate the brightness differences caused by different physical imaging channels. This provides all images with a unified signal benchmark, offering balanced image input for subsequent difference operations, target composite tensor construction, and model defect identification.

[0028] Furthermore, since there are different ways to perform image intensity stretching processing, this application embodiment can describe one possible image intensity stretching processing method.

[0029] In one case, the imaging intensity stretching processing of the target chip defect map, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip includes: The imaging intensity of the target chip defect map, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip is stretched from a first gray level to a second gray level. The second gray level is higher than the first gray level.

[0030] In practical applications, the first gray level is the original gray level of the target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target, actually acquired by each physical imaging channel before image stretching. Due to the influence of channel optical conditions, the first gray level is generally low, the acquired image is dark, and the intensity of the defect signal is weak. In this scheme, the imaging intensity stretching process stretches the imaging intensity of each image (the target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target corresponding to the target defect point in each physical imaging channel) from the first gray level before processing to the second gray level, and the second gray level is higher than the first gray level, thereby raising the overall gray level reference of a single image. This stretching process preserves the original relative brightness between pixels within the image, while widening the usable gray dynamic range of the image, increasing the gray level difference between defect pixels and background pixels, thereby preventing the defect signal from being masked by noise under weak imaging conditions, which is beneficial for subsequent image difference calculation and reduces the loss of weak defect information during processing.

[0031] S103: Construct a target composite tensor based on the target chip defect map after imaging intensity stretching, the target left adjacent chip reference map, and the target right adjacent chip reference map; the target composite tensor includes: the target chip defect map of each physical imaging channel, the differential fusion map obtained by operating the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map inside each physical imaging channel, and the target reference template map obtained by taking the median value of the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels.

[0032] In practical applications, after image intensity stretching, the target chip defect image, the reference image of the target's left adjacent chip, and the reference image of the target's right adjacent chip are used to construct the target composite tensor. The target composite tensor, as the input to the subsequent target UNET model, mainly contains the following three parts of image data: The first part is the stretched target chip defect image corresponding to each physical imaging channel, which retains the original acquired defect pixel details; the second part is the difference fusion image obtained by operating the stretched target chip defect image and the corresponding target's left and right adjacent chip reference images within each physical imaging channel, corresponding to the difference features within each channel; the third part is the target reference template image obtained by taking the median value from the stretched target's left and right adjacent chip reference images corresponding to all physical imaging channels, which can reduce the interference of single-point channel noise. In this way, a target composite tensor is constructed that simultaneously gathers three types of information: original defect pixel details, single-channel differential features, and target reference template image obtained by multi-channel fusion. Instead of relying solely on a single differential image for judgment, it fully utilizes the imaging characteristics of strong cross-channel correlation of defects and weak inter-channel correlation of differential residues, thereby improving the model's ability to perceive weak low-contrast defects and improving the quality of the final output target defect differential image.

[0033] Furthermore, since there are different ways to generate differential fusion maps, this application embodiment can describe one possible generation method.

[0034] In one case, the differential fusion graph is generated in the following ways: Within the same physical imaging channel, the target chip defect map after image intensity stretching is compared with the reference map of the adjacent chip on the left side of the target to obtain the first difference map; A second difference image is obtained by performing a difference operation between the target chip defect image after the imaging intensity stretching process and the reference image of the adjacent chip to the right of the target; The first difference map and the second difference map are superimposed and fused to obtain the difference fused map corresponding to each physical imaging channel.

[0035] In practical applications, the target defect location corresponds to a set of target chip defect images, a reference image of the adjacent chip to the left of the target, and a reference image of the adjacent chip to the right of the target within each physical imaging channel. When generating the differential fusion image, each physical imaging channel needs to be processed independently. First, within the same physical imaging channel, the target chip defect image after intensity stretching is differentially processed with the reference image of the adjacent chip to the left of the target to obtain the first differential image. Then, within the same physical imaging channel, the target chip defect image after intensity stretching is differentially processed with the reference image of the adjacent chip to the right of the target to obtain the second differential image. Finally, the first and second differential images corresponding to the same physical imaging channel are superimposed and fused to obtain the differential fusion image corresponding to each physical imaging channel. Differential operations can highlight areas where there are pixel differences between the target chip and adjacent chips, and initially locate potential defect locations. However, when using any single differential image, it is easy to introduce a lot of differential residue due to the texture and noise of the reference image itself. Therefore, this scheme further superimposes and fuses the two differential images obtained in the same channel to obtain the differential fusion image corresponding to that physical imaging channel. Thus, by performing differential analysis on the reference images of adjacent chips on the left and right sides of the target and then superimposing and fusing them, the information from the two differential results can be combined to a certain extent to suppress some random residual interference. Compared with a single differential image, the defect signal is stronger and the background noise is less, providing higher quality differential features for the target composite tensor, which facilitates the subsequent target UNET model to further distinguish between real defects and spurious residuals.

[0036] Furthermore, since there are different ways to generate target reference template diagrams, this application embodiment can describe one possible generation method.

[0037] In one case, the target reference template diagram is generated in the following manner: Obtain reference images of the target's left-side adjacent chips and the target's right-side adjacent chips corresponding to all physical imaging channels; Pixel-level median calculation is performed on all the reference images of the target's left adjacent chip and the target's right adjacent chip to obtain the target reference template image.

[0038] In practical applications, generating the target reference template image requires first obtaining all reference images of the target's left-side adjacent chips and right-side adjacent chips, after image intensity stretching processing, corresponding to all physical imaging channels. Then, pixel-level median calculation is performed on all left-side and right-side adjacent chip reference images to obtain the target reference template image. Pixel-level median calculation refers to taking the median of all pixel grayscale values ​​at the same location in multiple reference images and using this as the pixel grayscale value for the corresponding location in the target reference template image. It's understandable that using only a single reference image from a single channel is susceptible to interference from the unique grain noise and transient imaging disturbances of that channel. By performing median calculation on multiple reference images from multiple channels, accidental noise and local abnormal textures appearing in individual channels are filtered out. This suppresses random noise interference introduced by a single channel, resulting in a more reliable target reference template image. Incorporating this into the target composite tensor provides a stable background comparison basis for the subsequent target UNET model, facilitating the model's differentiation between real process defects and false residues caused by imaging, further improving the reliability of defect detection.

[0039] S104: Input the target composite tensor into the pre-trained target UNET model to obtain the target defect difference map for wafer defect detection.

[0040] In practical applications, the pre-trained target UNET model has learned the imaging principle of "strong cross-channel correlation of defects and weak inter-channel correlation of differential residues," enabling it to identify real process defect features in images while distinguishing various background residue interferences caused by channel noise and process deviations. By inputting a target composite tensor—which simultaneously gathers three dimensions of information—including original defect pixel details, single-channel differential features, and a target reference template image obtained through multi-channel fusion, into this target UNET model, a target defect differential map for wafer defect detection can be obtained. Figure 2 This is a schematic diagram illustrating the effect comparison of an embodiment of this application. (In conjunction with...) Figure 2 As shown, the target chip defect map obtained under the background of dual physical imaging channels for the target defect location (the target chip defect corresponding to one of the physical imaging channels). Figure 1 Target chip defects corresponding to another physical imaging channel Figure 2The target defect difference map obtained through the target UNET model significantly suppresses background residue compared to the original defect difference map obtained without any background processing. Thus, this scheme relies on a target composite tensor containing multi-dimensional information as input. The pre-trained target UNET model no longer depends solely on single-channel difference results for judgment, effectively suppressing various types of residual noise in the difference model and highlighting subtle process defect features within the image. In the output target defect difference map, the grayscale contrast between the real defect and the background is further enhanced, reducing the false detection probability and improving the identification accuracy of real defects among suspected defect points on the wafer. Of course, in other embodiments, the wafer inspection equipment can also have three physical imaging channels, four physical imaging channels, etc. Correspondingly, it is necessary to use data from 3 channels, 4 channels, or even more channels to construct more complex input tensors.

[0041] Furthermore, since the methods for training the target UNET model are not entirely the same, this application can describe one possible training method.

[0042] In one case, the target UNET model is trained using the following method: Obtain the training chip defect map, the training chip reference map to the left of the adjacent chip, and the training chip reference map to the right of the adjacent chip in each physical imaging channel; the training defect map is marked with a truth value label. The defect map of the training chip, the reference map of the left adjacent chip, and the reference map of the right adjacent chip are subjected to imaging intensity stretching processing. A training composite tensor is constructed based on the training chip defect map after imaging intensity stretching, the training left adjacent chip reference map, and the training right adjacent chip reference map. By combining the aforementioned training composite tensor, the initial UNET model is trained using the target loss function to obtain the target UNET model.

[0043] The target loss function includes: a background zero loss term, a defect fidelity loss term, and a total variational loss term.

[0044] In practical applications, training the UNET model first requires dataset preparation. The dataset can be constructed based on a large number of defect locations on the target wafer identified by the wafer inspection equipment through pre-processing algorithms. These numerous defect locations detected by the equipment are used as training defect locations. Each training defect location is labeled with a ground truth label, and its corresponding training chip defect image, left adjacent chip reference image, and right adjacent chip reference image for each physical imaging channel are stored in the dataset for subsequent model training. The number of training samples is on the order of 10. 4Taking a dual-channel physical imaging system as an example, the data in the dataset comes from chip defect maps of both channels, a reference map of the adjacent chip on the left, and a reference map of the adjacent chip on the right, totaling 6... There are N images, where N is the number of defect points used for training. Then, the training chip defect images, the left adjacent chip reference images, and the right adjacent chip reference images in the dataset are subjected to image intensity stretching. A training composite tensor is then constructed based on these stretched images. The image intensity stretching method for the training images can be the same as that for the target images, and the construction method for the training composite tensor can also be the same as that for the target composite tensor. Figure 3 This is a schematic diagram of a tensor matrix provided in an embodiment of this application. (In conjunction with...) Figure 3 As shown, taking a dual-channel physical imaging system as an example, any training defect point will be constructed into a 5-channel image. 32 A 32-dimensional image matrix tensor (a composite tensor for training) is used as input to the initial UNET model for training. 32 A 32-dimensional image matrix tensor, including: Channel 1 defect image 32 32 (Defect image of the training chip after stretching corresponding to one of the physical imaging channels), Defect image of channel 2 32 32 (Defect map of the training chip after stretching corresponding to another physical imaging channel), Channel 1 difference map 32 32 (a differential fusion map obtained by calculating the stretched training chip defect map inside one of the physical imaging channels with the corresponding training chip reference map on the left and the training chip reference map on the right), Channel 2 differential map 32 32 (A differential fusion image obtained by operating on the stretched training chip defect image and the corresponding training chip reference image on the left and right sides of the training chip reference image inside another physical imaging channel) and template (Golden) image 32 32 (The training reference template image is obtained by taking the median of the training reference image on the left and the training reference image on the right after stretching the images corresponding to these two physical imaging channels). Figure 4 This is a schematic diagram of a UNET model provided in an embodiment of this application. (Combined with...) Figure 4 As shown, the UNET model structure used in this scheme is a standard U-shaped symmetrical architecture. Continuing with the example of a dual-channel physical imaging system, for any training defect location, the input 5... 32 The 32-dimensional image matrix tensor will be encoded through 4 layers of encoders, from 5 32 32-bit encoding to 512-bit encoding 2 2. Each encoder layer sequentially performs convolutional network feature extraction and pooling operations, extracting 512-dimensional high-level semantic information at the final encoder layer. Following the encoder are four decoder layers, each performing deconvolution and concatenation operations. The deconvolution operation progressively restores the high-level semantic features to pixel-level spatial details. The concatenation operation concatenates the intermediate results from the encoder layers pixel-wise, allowing the decoder to process both high-level semantic information and low-level image spatial details simultaneously. The final decoder layer outputs a dimension of 128. 32 A 32-dimensional feature map, after being processed by the Sigmoid function, outputs a dimension of 1. 32 The residual suppression difference map output is 32. This UNET architecture integrates high-level semantics and low-level image spatial details, which can preserve the features of the input image as much as possible during the iterative optimization of the loss function. Thus, combined with the training set constructed from a large number of training defect points labeled with ground truth, the initial UNET model is iteratively trained using the target loss function to finally obtain the target UNET model. Furthermore, the expression for this target loss function is as follows: ; In the formula, Represents the loss function. Zero loss for background To compensate for defects and losses, For the total variational loss, , , These represent the weights corresponding to the background zero loss, defect fidelity loss, and total variational loss, respectively. For the background zero loss... In this regard, the corresponding expression is as follows: ; In the formula, This is a soft mask for the background region after Gaussian weight decay in the central region. The image data represents the output of the model. This represents the 1-norm of x, i.e., its absolute value. The background zero loss forces the network output to tend to 0 in non-central regions (areas with high weights), thereby eliminating all background residue, noise, and unreduced circuit textures. In our pre-processor algorithm, detected defects are displayed in the center by default; therefore, the core issue is suppressing background in non-central regions while preserving relevant details. For defect fidelity loss... In this regard, the corresponding expression is as follows: ; In the formula, The image data represents the output of the model. To fuse differential signals, This represents finding the 1-norm of x. A Gaussian weight mask is used for the central region. Defect fidelity loss is used to make the network output as close as possible to the fused differential signal in the central region (where the Gaussian weights are high). (Sum of the two input channel difference plots), which preserves and enhances the deficiency in cross-channel consistency. For the total variational loss... In this regard, the corresponding expression is as follows: ; In the formula, The image data represents the output of the model. Represents the gradient in the x-direction of the image. This represents the gradient in the y-direction of the image. The total variational loss reduces local gradients in the image, suppresses isolated noise and small fluctuations in the output, further smooths background regions (which may still have slight residuals even after being pushed to 0 by zero loss), and simultaneously protects the sharp edges of central defects (because the gradient at the defect edges is still necessary under the fidelity loss). Additionally, in the weighting, > As an example, = 8, = 5. This ensures that the defect fidelity in the central region dominates the background with zero loss in the transition zone, guaranteeing that the defect signal is not weakened. A smaller value (e.g., 0.3) is used only as an auxiliary smoothing term to prevent excessive blurring defects. This weighting ratio has been experimentally verified to effectively extract weak defects. Of course, in other embodiments, the UNET model structure can use 5 or even 6 layers of encoders / encoders for deeper image feature extraction, and this application does not limit this.

[0045] In summary, this application first obtains the target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map corresponding to the target defect location in each physical imaging channel. Then, it performs imaging intensity stretching processing on the target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map, and constructs a target composite tensor based on the stretched target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map. This target composite tensor includes: the target chip defect map of each physical imaging channel; a differential fusion map obtained within each physical imaging channel by operations on the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map; and a target reference template map obtained by taking the median value from the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels. Finally, the target composite tensor is input into a pre-trained target UNET model to obtain a target defect differential map for wafer defect detection. Thus, a target composite tensor is constructed using the target chip defect map after imaging intensity stretching, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip. This tensor is then input into a pre-trained target UNET model. By relying on the characteristics of strong cross-channel correlation of defects and weak cross-channel correlation of differential residues, background texture interference is suppressed and weak defect details are fully preserved. Finally, a target defect differential map for wafer defect detection is obtained, which improves the detection accuracy and signal-to-noise ratio of wafer defects.

[0046] Figure 5 This is a schematic diagram of a multi-channel image fusion defect enhancement device provided in an embodiment of this application. (Combined with...) Figure 5 As shown, the multi-channel image fusion defect enhancement device 500 is used for wafer die-to-die mode defect detection, and includes: The acquisition module 501 is used to acquire the target chip defect map, the reference map of the target adjacent chip to the left of the target, and the reference map of the target adjacent chip to the right of the target corresponding to the target defect point in each physical imaging channel. The stretching module 502 is used to perform imaging intensity stretching processing on the target chip defect map, the reference map of the target adjacent chip to the left and the reference map of the target adjacent chip to the right. The construction module 503 is used to construct a target composite tensor based on the target chip defect map after imaging intensity stretching, the target left adjacent chip reference map, and the target right adjacent chip reference map; the target composite tensor includes: the target chip defect map of each physical imaging channel, the differential fusion map obtained by operating the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map inside each physical imaging channel, and the target reference template map obtained by taking the median value of the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels; The processing module 504 is used to input the target composite tensor into the pre-trained target UNET model and obtain the target defect difference map for wafer defect detection.

[0047] The number of physical imaging channels is at least two.

[0048] As one implementation method, regarding how to generate the differential fusion graph, the aforementioned construction module 503 includes: a first generation module; the first generation module is used for: Within the same physical imaging channel, the target chip defect map after image intensity stretching is compared with the reference map of the adjacent chip on the left side of the target to obtain the first difference map; A second difference image is obtained by performing a difference operation between the target chip defect image after the imaging intensity stretching process and the reference image of the adjacent chip to the right of the target; The first difference map and the second difference map are superimposed and fused to obtain the difference fused map corresponding to each physical imaging channel.

[0049] As one implementation method, regarding how to generate the target reference template diagram, the aforementioned construction module 503 includes: a second generation module; the second generation module is used for: Obtain reference images of the target's left-side adjacent chips and the target's right-side adjacent chips corresponding to all physical imaging channels; Pixel-level median calculation is performed on all the reference images of the target's left adjacent chip and the target's right adjacent chip to obtain the target reference template image.

[0050] As one implementation, the multi-channel image fusion defect enhancement device 500 further includes a training module for training the target UNET model; the training module is used for: Obtain the training chip defect map, the training chip reference map to the left of the adjacent chip, and the training chip reference map to the right of the adjacent chip in each physical imaging channel; the training defect map is marked with a truth value label. The defect map of the training chip, the reference map of the left adjacent chip, and the reference map of the right adjacent chip are subjected to imaging intensity stretching processing. A training composite tensor is constructed based on the training chip defect map after imaging intensity stretching, the training left adjacent chip reference map, and the training right adjacent chip reference map. By combining the aforementioned training composite tensor, the initial UNET model is trained using the target loss function to obtain the target UNET model.

[0051] The target loss function includes: a background zero loss term, a defect fidelity loss term, and a total variational loss term.

[0052] As one implementation method, regarding how to perform image intensity stretching processing, the stretching module 502 is specifically used for: The imaging intensity of the target chip defect map, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip is stretched from a first gray level to a second gray level. The second gray level is higher than the first gray level.

[0053] In summary, this application first obtains the target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map corresponding to the target defect location in each physical imaging channel. Then, it performs imaging intensity stretching processing on the target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map, and constructs a target composite tensor based on the stretched target chip defect map, the target left adjacent chip reference map, and the target right adjacent chip reference map. This target composite tensor includes: the target chip defect map of each physical imaging channel; a differential fusion map obtained within each physical imaging channel by operations on the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map; and a target reference template map obtained by taking the median value from the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels. Finally, the target composite tensor is input into a pre-trained target UNET model to obtain a target defect differential map for wafer defect detection. Thus, a target composite tensor is constructed using the target chip defect map after imaging intensity stretching, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip. This tensor is then input into a pre-trained target UNET model. By relying on the characteristics of strong cross-channel correlation of defects and weak cross-channel correlation of differential residues, background texture interference is suppressed and weak defect details are fully preserved. Finally, a target defect differential map for wafer defect detection is obtained, which improves the detection accuracy and signal-to-noise ratio of wafer defects.

[0054] In addition, this application also provides a multi-channel image fusion defect enhancement device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the multi-channel image fusion defect enhancement method as described above.

[0055] In addition, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the multi-channel image fusion defect enhancement method described above.

[0056] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for enhancing defects in multi-channel image fusion, characterized in that, The method is used for wafer die-to-die mode defect detection, and the method includes: Obtain the target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target corresponding to the target defect location in each physical imaging channel; The target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target are subjected to imaging intensity stretching processing; A target composite tensor is constructed based on the target chip defect map after imaging intensity stretching, the target left adjacent chip reference map, and the target right adjacent chip reference map; the target composite tensor includes: the target chip defect map of each physical imaging channel, the differential fusion map obtained by operating the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map inside each physical imaging channel, and the target reference template map obtained by taking the median value of the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels; The target composite tensor is input into the pre-trained target UNET model to obtain the target defect difference map for wafer defect detection.

2. The method according to claim 1, characterized in that, The generation methods of the differential fusion map include: Within the same physical imaging channel, the target chip defect map after image intensity stretching is compared with the reference map of the adjacent chip on the left side of the target to obtain the first difference map; A second difference image is obtained by performing a difference operation between the target chip defect image after the imaging intensity stretching process and the reference image of the adjacent chip to the right of the target; The first difference map and the second difference map are superimposed and fused to obtain the difference fused map corresponding to each physical imaging channel.

3. The method according to claim 1, characterized in that, The methods for generating the target reference template diagram include: Obtain reference images of the target's left-side adjacent chips and the target's right-side adjacent chips corresponding to all physical imaging channels; Pixel-level median calculation is performed on all the reference images of the target's left adjacent chip and the target's right adjacent chip to obtain the target reference template image.

4. The method according to claim 1, characterized in that, The number of physical imaging channels is at least two.

5. The method according to claim 1, characterized in that, The target UNET model was trained using the following method: Obtain the training chip defect map, the training chip reference map to the left of the adjacent chip, and the training chip reference map to the right of the adjacent chip in each physical imaging channel; the training defect map is marked with a truth value label. The defect map of the training chip, the reference map of the left adjacent chip, and the reference map of the right adjacent chip are subjected to imaging intensity stretching processing. A training composite tensor is constructed based on the training chip defect map after imaging intensity stretching, the training left adjacent chip reference map, and the training right adjacent chip reference map. By combining the aforementioned training composite tensor, the initial UNET model is trained using the target loss function to obtain the target UNET model.

6. The method according to claim 5, characterized in that, The target loss function includes: background zero loss term, defect fidelity loss term, and total variational loss term.

7. The method according to claim 1, characterized in that, The imaging intensity stretching process performed on the defect map of the target chip, the reference map of the chip adjacent to the left of the target, and the reference map of the chip adjacent to the right of the target includes: The imaging intensity of the target chip defect map, the reference map of the target's left adjacent chip, and the reference map of the target's right adjacent chip is stretched from a first gray level to a second gray level. The second gray level is higher than the first gray level.

8. A multi-channel image fusion defect enhancement device, characterized in that, The device is used for wafer die-to-die mode defect detection, including: The acquisition module is used to acquire the target chip defect image, the reference image of the chip adjacent to the left of the target, and the reference image of the chip adjacent to the right of the target corresponding to the target defect point in each physical imaging channel. The stretching module is used to perform imaging intensity stretching processing on the target chip defect map, the reference map of the target adjacent chip to the left and the reference map of the target adjacent chip to the right. A construction module is used to construct a target composite tensor based on the target chip defect map after imaging intensity stretching, the target left adjacent chip reference map, and the target right adjacent chip reference map; the target composite tensor includes: the target chip defect map of each physical imaging channel, the differential fusion map obtained by operating the target chip defect map and the corresponding target left adjacent chip reference map and target right adjacent chip reference map inside each physical imaging channel, and the target reference template map obtained by taking the median value of the target left adjacent chip reference map and target right adjacent chip reference map of all physical imaging channels; The processing module is used to input the target composite tensor into the pre-trained target UNET model and obtain the target defect difference map for wafer defect detection.

9. A multi-channel image fusion defect enhancement device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the multi-channel image fusion defect enhancement method as described in any one of claims 1 to 7 when executing the computer program.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the steps of the multi-channel image fusion defect enhancement method as described in any one of claims 1 to 7.

Images