An image exposure correction method and system

Abstract

The application provides an image exposure correction method and system, the method comprising: obtaining an input image to be corrected; performing multi-scale decomposition on the input image and its inverse image to obtain corresponding low-frequency components and high-frequency components respectively; performing exposure correction processing on the low-frequency components and detail enhancement processing on the high-frequency components by using a neural network; reconstructing the processed low-frequency components and high-frequency components to obtain at least one intermediate correction image; adaptively fusing the input image and at least one intermediate correction image to output an exposure correction image, effectively restoring the complete appearance and clear edges of the mark, and providing a high-quality and high-credibility input image for a subsequent high-precision positioning algorithm, and fundamentally improving the performance basis of an alignment system.

Description

Technical Field

[0001] This invention belongs to the field of exposure technology, and specifically relates to an image exposure correction method and system. Background Technology

[0002] In photolithography, a core process in semiconductor integrated circuit manufacturing, the performance of the stepper (or scanner) directly determines the chip's linewidth and yield. The stepper precisely copies the pattern on a mask / reticle onto a silicon wafer coated with photoresist by using a projection lens. This process requires an extremely high-precision alignment system to ensure that the current layer is precisely aligned with the previous layer.

[0003] Modern alignment systems typically employ machine vision-based image sensors to achieve precise positioning by acquiring images of specific alignment marks on masks and silicon wafers and calculating their positional deviations.

[0004] However, in actual imaging, the reflectivity of silicon wafers varies significantly due to differences in thin film materials (such as silicon nitride and polycrystalline silicon), the thickness and composition of photoresist, and the structure of the markings themselves (such as tendches and bars). High-reflectivity areas (such as metal markings) are prone to overexposure (saturation) in the image, resulting in blurred marking edges and loss of shape details; while low-reflectivity areas may be underexposed, reducing the signal-to-noise ratio and making it difficult to extract marking features. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an image exposure correction method and system to resolve the technical issues in the prior art.

[0006] On one hand, the invention provides the following technical solution: an image exposure correction method, the method comprising: The input image to be corrected is obtained, and the input image and its inverse image are decomposed into multiple scales to obtain the corresponding low-frequency components and high-frequency components respectively. The low-frequency components are subjected to exposure correction using a neural network, and the high-frequency components are subjected to detail enhancement processing. The processed low-frequency components and high-frequency components are reconstructed to obtain at least one intermediate corrected image; The input image is adaptively fused with at least one of the intermediate correction images to output an exposure-corrected image.

[0007] Compared to existing technologies, the advantages of this application are as follows: through multi-scale decomposition and frequency division processing, this method can simultaneously solve the two major problems most challenging in exposure machine images: overexposure leading to marker edge saturation and underexposure leading to low marker signal-to-noise ratio. It effectively restores the complete shape and clear edges of the marker, providing a high-quality, high-reliability input image for subsequent high-precision positioning algorithms, fundamentally improving the performance basis of the alignment system.

[0008] Furthermore, prior to the adaptive fusion step, the method further includes: Obtain the raw image containing alignment marks; Identify the region of interest where the alignment markers are located in the original image; Based on the identified region of interest, a region importance weight map with the same size as the input image is generated, wherein pixels located within the region of interest are assigned a first weight value, pixels located outside the region of interest are assigned a second weight value, and the first weight value is greater than the second weight value. The input image, the at least one intermediate corrected image, and the region importance weight map are input into a fusion network, and the corresponding weight vector is output. The input image and the at least one intermediate correction image are weighted and fused using the weight vector to output the final exposure-corrected image.

[0009] Furthermore, the step of identifying the region of interest includes: Based on the preset or historical alignment information of the mask and silicon wafer of the original image, the predicted regions in which the first alignment mark and the second alignment mark of the original image appear in the current original image are calculated using the alignment information calculation formula, wherein the original image contains the first alignment mark located on the mask and the second alignment mark located on the silicon wafer. Define the predicted region as the region of interest; The formula for calculating the alignment information is as follows:

[0010] in, and The coordinates of two feature points on the first alignment mark are obtained from the pre-stored reference image. and Let d be the coordinates of two feature points on the second alignment mark obtained from the pre-stored reference image, and let d be the distance between the two feature points on the second alignment mark. This represents the offset between the mask and the silicon wafer in the X direction. This represents the offset between the mask and the silicon wafer in the Y direction. The relative rotation angle between the mask and the silicon wafer.

[0011] Furthermore, the step of identifying the region of interest includes: Preliminary edge detection or feature extraction is performed on the original image to directly locate the rough outline of the alignment mark; The rough outline is expanded outward by a preset number of pixels to form a bounding box, and the area within the bounding box is defined as the region of interest.

[0012] Furthermore, the adaptive fusion step is implemented through the following formula:

[0013] in, For the final exposure-corrected image, The input image, and The intermediate corrected image, , , This represents the weight vector related to the image content generated by the fusion network.

[0014] Furthermore, the method also includes: Determine whether the color balance of the exposure-corrected image meets the requirements; If the color balance of the exposure-corrected image does not meet the requirements, the exposure-corrected image is input into the prediction neural network for inference to predict the required red channel gain value and blue channel gain value for the current image. The exposure-corrected image is corrected using a correction formula based on the parameters of the red channel gain value and the blue channel gain value to obtain a re-corrected exposure-corrected image.

[0015] Furthermore, the prediction neural network is an improved network based on the ResNet34 architecture, and its improvements include: Replace the activation function from ReLU with Leaky ReLU; Replace the global average pooling operation at the end of the network with a local pooling operation; The final output layer of the network is a fully connected layer with an output dimension of 2, corresponding to the gain values ​​of the red channel and the blue channel, respectively. The predictive neural network is trained using a combined loss function; The combined loss function includes:

[0016] in, The loss function is Log-Cosh. For CosineEmbedding loss function, This is the combined loss function; The correction formula includes

[0017] in , , These represent the red, green, and blue channel components of the exposure-corrected image, respectively. The image is shown as the exposure-corrected image after further correction.

[0018] Secondly, the invention provides the following technical solution: an image exposure correction system, the system comprising: The decomposition module is used to acquire the input image to be corrected, and to decompose the input image and its inverse image into multiple scales to obtain the corresponding low-frequency components and high-frequency components respectively. The processing module is used to perform exposure correction processing on the low-frequency components using a neural network and to perform detail enhancement processing on the high-frequency components. The reconstruction module is used to reconstruct the processed low-frequency components and high-frequency components to obtain at least one intermediate corrected image; The fusion module is used to adaptively fuse the input image with at least one of the intermediate correction images to output an exposure correction image.

[0019] Thirdly, the invention provides the following technical solution: a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the image exposure correction method as described above.

[0020] Fourthly, the invention provides the following technical solution: a storage medium storing a computer program, which, when executed by a processor, implements the image exposure correction method described above. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of the image exposure correction method provided in the first embodiment of the present invention; Figure 2 This is a structural block diagram of the image exposure correction system provided in the second embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of a computer provided in the third embodiment of the present invention.

[0023] The embodiments of the present invention will be further described below with reference to the accompanying drawings. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0025] In the description of the embodiments of the present invention, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0026] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0027] Example 1 In the first embodiment of the present invention, please refer to Figure 1 As shown, an image exposure correction method includes the following steps S01 to S04: S01, acquire the input image to be corrected, and perform multi-scale decomposition on the input image and its inverse image to obtain the corresponding low-frequency components and high-frequency components respectively; In this embodiment, taking the alignment of a semiconductor wafer with an image as an example, the specific processing procedure is as follows: Obtain the alignment image of the semiconductor wafer to be corrected (i.e., the input image X). This image is usually an underexposed or overexposed RGB image with a size of H×W×3 (including the three RGB channels).

[0028] To build the foundation for dual-path processing, the inverse image of the input image X is generated. The calculation formula is:

[0029] This operation generates the inverted image of the underexposed image X. This indicates an overexposed state; for overexposed images Its reverse image This results in an underexposed state, thus transforming the exposure correction problem of a single image into a collaborative processing of a pair of complementary exposure problems.

[0030] Subsequently, the input image X and its inverse image are processed. Gaussian pyramid decomposition is performed separately. Taking the input image X as an example, L-level decomposition is performed (L=3 in this embodiment) to obtain the corresponding low-frequency component set. and high frequency component set Among them, the lowest level The lowest frequency component (presenting a highly blurred image state). High-frequency details at the original image resolution. For the inverted image. Perform the exact same decomposition process as the input image X to obtain its corresponding low-frequency component set. …and high-frequency component sets …

[0031] S02, use a neural network to perform exposure correction processing on the low-frequency components and detail enhancement processing on the high-frequency components; In this embodiment, the low-frequency and high-frequency components obtained from the decomposition are processed separately, and the specific process is as follows: Low frequency components Low-frequency components of the inverted image After concatenation, the data is fed into the exposure correction network. This network uses a lightweight U-Net architecture, with both the encoder and decoder containing four convolutional layers. The output of this network is then compared to the previous output. Corrected low-frequency components of the same size .

[0032] High frequency components High-frequency components of the inverted image After stacking, the data is fed into a detail enhancement network. This network consists of three consecutive convolutional layers, and the final output is the high-frequency components with enhanced details. .

[0033] S03, reconstruct the processed low-frequency component and the high-frequency component to obtain at least one intermediate corrected image; In this embodiment, the Laplacian pyramid reconstruction algorithm is used for image reconstruction. The specific process is as follows: The processed low-frequency components Unprocessed high-frequency components and the enhanced top-level high-frequency components Upsampling and fusion operations are performed to reconstruct an intermediate corrected image. .

[0034] By adjusting the network parameters, this embodiment can generate two intermediate corrected images with different emphases. and For example, one intermediate corrected image focuses on optimizing global contrast, while the other focuses on restoring local details to meet the correction needs of different scenarios. The task of the fusion network is to intelligently select the best portion for each pixel from X, X1, and X2. The main methods are as follows: During the training phase, two subnetworks with different optimization objectives can be trained, resulting in different outputs.

[0035] Different loss function weights: Used to generate The network: When training its exposure correction and detail enhancement modules, its loss function may focus more on the smooth transition and naturalness of the overall brightness (e.g., L1 Loss has a higher weight).

[0036] Used to generate The network's training loss function may focus more on local contrast and detail sharpening (e.g., MS-SSIM Loss or GAN Loss have higher weights).

[0037] Result: During inference, the same input image, when passed through these two trained networks, yields results with different emphases. (More natural) and (Sharper).

[0038] Fine-tuning of network structure: The same network architecture can be used, but different model weight files can be used. One set of weights (Weight A) is optimized to produce conservative correction results. Another set of weights (Weight B) is optimized to produce aggressively enhanced results. ).

[0039] S04, adaptively fuse the input image with at least one of the intermediate correction images to output an exposure correction image.

[0040] Specifically, prior to the adaptive fusion step, the method further includes: S41, Obtain the original image containing alignment marks; S42, Identify the region of interest where the alignment markers are located in the original image; S43, Based on the identified region of interest, generate a region importance weight map with the same size as the input image, wherein pixels located within the region of interest are assigned a first weight value, pixels located outside the region of interest are assigned a second weight value, and the first weight value is greater than the second weight value. S44, input the input image, the at least one intermediate corrected image, and the region importance weight map into a fusion network, and output the corresponding weight vector; S45, the input image and the at least one intermediate correction image are weighted and fused using the weight vector to output the final exposure correction image.

[0041] Optionally, the step of identifying the region of interest includes: Based on the preset or historical alignment information of the mask and silicon wafer of the original image, the predicted regions in which the first alignment mark and the second alignment mark of the original image appear in the current original image are calculated using the alignment information calculation formula, wherein the original image contains the first alignment mark located on the mask and the second alignment mark located on the silicon wafer. Define the predicted region as the region of interest; The formula for calculating the alignment information is as follows:

[0042] in, and The coordinates of two feature points on the first alignment mark are obtained from the pre-stored reference image. and Let d be the coordinates of two feature points on the second alignment mark obtained from the pre-stored reference image, and let d be the distance between the two feature points on the second alignment mark. This represents the offset between the mask and the silicon wafer in the X direction. This represents the offset between the mask and the silicon wafer in the Y direction. The relative rotation angle between the mask and the silicon wafer.

[0043] Optionally, the step of identifying the region of interest includes: Preliminary edge detection or feature extraction is performed on the original image to directly locate the rough outline of the alignment mark; The rough outline is expanded outward by a preset number of pixels to form a bounding box, and the area within the bounding box is defined as the region of interest.

[0044] In this embodiment, the original image I containing mask alignment marks and silicon wafer alignment marks is first acquired.

[0045] Retrieves pre-stored reference image feature point coordinates from the system database, including mask marker points. and and silicon wafer markers and( .

[0046] Calculated based on the alignment information calculation formula , , An affine transformation is performed on the marked region in the reference image to predict its location in the current image I, and this region is defined as the region of interest.

[0047] Optionally, the original image $I$ is converted to grayscale and Canny edge detection is performed. The rough contours of all alignment marks are located using Hough transform or contour lookup functions. The smallest bounding rectangle of all contours is extended outward by 20 pixels to form a large bounding box, which is the region of interest.

[0048] Create a zero matrix $M$ (with dimensions $H \times W$) of the same size as the input image x. Set all pixel values ​​within the ROI region to 0.8 (first weight value) and pixel values ​​outside the region to 0.2 (second weight value). Then apply a Gaussian blur filter to $M$ to smooth the weight transition and avoid obvious boundaries after merging.

[0049] Specifically, the adaptive fusion step is implemented through the following formula:

[0050] in, For the final exposure-corrected image, The input image, and The intermediate corrected image, , , This represents the weight vector related to the image content generated by the fusion network.

[0051] In this embodiment, the input image X and two intermediate correction images are used. and The regional importance weight maps are stacked along the channel dimension to form an H×W×7 tensor (where 3 channels come from X, 3 channels come from X1, 3 channels come from X2, and 1 channel comes from M, for a total of 3+3+3+1=7 channels). This tensor is then input into a fusion network.

[0052] The network structure consists of four convolutional layers, with the last layer using the sigmoid activation function. The network outputs three weight maps of the same size as the input image. , , Each dimension (each weight map has a size of H×W×1) corresponds to... , The weights of X and X are given, and the weights at each pixel position are normalized: w1+w2+w3=1.

[0053] The final exposure-corrected image is generated by performing pixel-by-pixel weighted fusion according to the adaptive fusion formula. .

[0054] Alternatively, the method may further include: Determine whether the color balance of the exposure-corrected image meets the requirements; If the color balance of the exposure-corrected image does not meet the requirements, the exposure-corrected image is input into the prediction neural network for inference to predict the required red channel gain value and blue channel gain value for the current image. The exposure-corrected image is corrected using a correction formula based on the parameters of the red channel gain value and the blue channel gain value to obtain a re-corrected exposure-corrected image.

[0055] Furthermore, the prediction neural network is an improved network based on the ResNet34 architecture, and its improvements include: Replace the activation function from ReLU with Leaky ReLU; Replace the global average pooling operation at the end of the network with a local pooling operation; The final output layer of the network is a fully connected layer with an output dimension of 2, corresponding to the gain values ​​of the red channel and the blue channel, respectively. The predictive neural network is trained using a combined loss function; The combined loss function includes:

[0056] in, The loss function is Log-Cosh. For CosineEmbedding loss function, This is the combined loss function; The correction formula includes

[0057] in , , These represent the red, green, and blue channel components of the exposure-corrected image, respectively. The image is shown as the exposure-corrected image after further correction.

[0058] In this embodiment, the mean value of the exposure-corrected image Y across the RGB three channels is calculated. The color balance is assessed by judging the degree of deviation of mean(R) / mean(G) and mean(B) / mean(G) from the ideal value (usually 1). If the deviation exceeds a set threshold (e.g., 15%), the image is determined to be color unbalanced and color correction is required.

[0059] After scaling the exposure-corrected image Y to 224×224, it is input into the white balance prediction network for gain prediction.

[0060] This network is an improvement on the ResNet34 architecture, with specific improvements including: Replace all ReLU activation functions with Leaky ReLU (with a negative slope of 0.01); Remove the last two layers of the original network (global average pooling layer and fully connected layer). Replace it with a local average pooling layer with a kernel size of 4×4, followed by a fully connected layer with an output dimension of 2. The network ultimately outputs two scalar values: the red channel gain Rgain and the blue channel gain Bgain.

[0061] In summary, the image exposure correction method has the following effects: By employing a multi-scale decomposition and frequency division processing strategy, this method specifically addresses the critical challenges in exposure machine images: overexposure leading to saturated marker edges and underexposure resulting in low signal-to-noise ratios. After multi-scale decomposition of the input image and its inverse image, exposure correction is applied to the low-frequency components, while detail enhancement is performed on the high-frequency components. This effectively improves the overall exposure level while maximizing the preservation and enhancement of image detail information. This characteristic ensures the clear representation of minute features such as alignment marks in semiconductor images, providing high-quality, high-reliability input images for subsequent high-precision positioning algorithms. This fundamentally improves the performance basis of the alignment system and significantly enhances the accuracy of subsequent feature recognition.

[0062] By designing a region importance weight map, the computational resources of the fusion network and the attention of the alignment algorithm are prioritized for the marked regions (ROIs). The exposure correction effect and detail sharpness of the marked regions are optimized to the extreme, while the background regions maintain a natural transition. This "precise focus" mode minimizes the positioning error caused by image quality issues, directly reducing overlay error, which is of key significance for improving the yield of advanced process chips.

[0063] Adaptive fusion is achieved by employing dynamic weight vectors generated by a fusion network. This allows for dynamic adjustment of the fusion ratio between the input image and the intermediate corrected image based on the characteristics of the image content, resulting in stronger scene adaptability compared to fixed-weight fusion strategies. This enables the system to intelligently respond to image variations caused by different reflectivity materials and process layers, outputting stable and reliable correction results under various operating conditions. It reduces equipment alarms or production interruptions caused by image quality fluctuations, meeting the diverse image correction needs in semiconductor manufacturing. In terms of ROI positioning, the predictive method based on historical alignment information is extremely fast and adds almost no processing time, perfectly meeting the high-speed, continuous production requirements of the exposure machine; while the image-based ROI positioning method serves as a redundant backup and can still work reliably under initialization or abnormal conditions. The combination of the two ensures the continuity and reliability of the production process.

[0064] To address the color deviation issues arising from different thin films on silicon wafers under specific lighting conditions, the automatic white balance post-processing function uses a predictive neural network to perform color balance judgment and secondary correction on the corrected image. This effectively corrects color distortions related to process materials and avoids color casts that may occur during exposure correction. This ensures that the image reflects the true geometric shape rather than color interference, further improving the accuracy and consistency of alignment measurements. For processes such as semiconductor defect detection that rely on color information, this significantly enhances the reliability of image analysis.

[0065] Example 2 like Figure 2 As shown, a second embodiment of the present invention provides an image exposure correction system, the system comprising: The decomposition module 10 is used to acquire the input image to be corrected, and to perform multi-scale decomposition on the input image and its inverse image to obtain the corresponding low-frequency components and high-frequency components respectively. The processing module 20 is used to perform exposure correction processing on the low-frequency components using a neural network and to perform detail enhancement processing on the high-frequency components. Reconstruction module 30 is used to reconstruct the processed low-frequency component and the high-frequency component to obtain at least one intermediate corrected image; The fusion module 40 is used to adaptively fuse the input image with at least one of the intermediate correction images to output an exposure correction image.

[0066] The image exposure correction system provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0067] Example 3 like Figure 3 As shown, in the third embodiment of the present invention, the present invention provides the following technical solution: a computer, including a memory 202, a processor 201, and a computer program stored in the memory 202 and executable on the processor 201, wherein the processor 201 executes the computer program to implement the image exposure correction method as described above.

[0068] Specifically, the processor 201 may include a central processing unit, a specific integrated circuit, or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0069] Memory 202 may include a large-capacity memory for data or instructions. For example, and not limitingly, memory 202 may include a hard disk drive, floppy disk drive, solid-state drive, flash memory, optical disk drive, magneto-optical disk drive, magnetic tape drive, or Universal Serial Bus drive, or a combination of two or more of these. Where appropriate, memory 202 may include removable or non-removable media. Where appropriate, memory 202 may be internal or external to a data processing device. In a particular embodiment, memory 202 is non-volatile memory. In a particular embodiment, memory 202 includes read-only memory and random access memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM, an erasable PROM, an electrically erasable PROM, an electrically rewritable ROM, or flash memory, or a combination of two or more of these. Where appropriate, the RAM may be static random access memory (SRAM) or dynamic random access memory (DRAM), wherein DRAM may be fast page-mode DRAM, extended data output DRAM, synchronous DRAM, etc.

[0070] The memory 202 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 201.

[0071] The processor 201 implements the above-described image exposure correction method by reading and executing computer program instructions stored in the memory 202.

[0072] In some embodiments, the computer may further include a communication interface 203 and a bus 200. For example, Figure 3 As shown, the processor 201, memory 202, and communication interface 203 are connected through bus 200 and complete communication with each other.

[0073] The communication interface 203 is used to enable communication between the various modules, devices, units, and / or equipment in the embodiments of this application. The communication interface 203 can also enable data communication with other components such as external devices, image / data acquisition devices, databases, external storage, and image / data processing workstations.

[0074] Bus 200 includes hardware, software, or both, that couples computer components together. Bus 200 includes, but is not limited to, at least one of the following: data bus, address bus, control bus, expansion bus, local bus. For example, and not limitingly, bus 200 may include a graphics acceleration interface or other graphics bus, an enhanced industry standard architecture bus, a front-side bus, HyperTransport interconnect, an industry standard architecture bus, a wireless bandwidth interconnect, a low pin count bus, a memory bus, a WeChat architecture bus, a peripheral component interconnect bus, a PCI Express bus, a Serial Advanced Technology Attached Bus, a Video Electronics Standards Association local bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 200 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0075] Example 4 In the fourth embodiment of the present invention, in conjunction with the above-described image exposure correction method, the present invention provides the following technical solution: a storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-described image exposure correction method.

[0076] Those skilled in the art will understand that the data in the flowchart, or logic and / or steps otherwise described herein, for example, can be considered as a sequenced data table of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device. For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0077] More specific examples of readable media include: electrical connections with one or more wires, portable computer disk drives, random access memory, read-only memory, erasable and editable read-only memory, fiber optic devices, and portable optical disc read-only memory. Additionally, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0078] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0079] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An image exposure correction method, characterized in that, The method includes: The input image to be corrected is obtained, and the input image and its inverse image are decomposed into multiple scales to obtain the corresponding low-frequency components and high-frequency components respectively. The low-frequency components are subjected to exposure correction using a neural network, and the high-frequency components are subjected to detail enhancement processing. The processed low-frequency components and high-frequency components are reconstructed to obtain at least one intermediate corrected image; The input image is adaptively fused with at least one of the intermediate correction images to output an exposure-corrected image.

2. The image exposure correction method according to claim 1, characterized in that, Prior to the adaptive fusion step, the method further includes: Obtain the raw image containing alignment marks; Identify the region of interest where the alignment markers are located in the original image; Based on the identified region of interest, a region importance weight map with the same size as the input image is generated, wherein pixels located within the region of interest are assigned a first weight value, pixels located outside the region of interest are assigned a second weight value, and the first weight value is greater than the second weight value. The input image, the at least one intermediate corrected image, and the region importance weight map are input into a fusion network, and the corresponding weight vector is output. The input image and the at least one intermediate correction image are weighted and fused using the weight vector to output the final exposure-corrected image.

3. The image exposure correction method according to claim 2, characterized in that, The steps for identifying the region of interest include: Based on the preset or historical alignment information of the mask and silicon wafer of the original image, the predicted regions in which the first alignment mark and the second alignment mark of the original image appear in the current original image are calculated using the alignment information calculation formula, wherein the original image contains the first alignment mark located on the mask and the second alignment mark located on the silicon wafer. Define the predicted region as the region of interest; The formula for calculating the alignment information is as follows: in, and The coordinates of two feature points on the first alignment mark are obtained from the pre-stored reference image. and Let d be the coordinates of two feature points on the second alignment mark obtained from the pre-stored reference image, and let d be the distance between the two feature points on the second alignment mark. This represents the offset between the mask and the silicon wafer in the X direction. This represents the offset between the mask and the silicon wafer in the Y direction. The relative rotation angle between the mask and the silicon wafer.

4. The image exposure correction method according to claim 2, characterized in that, The steps for identifying the region of interest include: Preliminary edge detection or feature extraction is performed on the original image to directly locate the rough outline of the alignment mark; The rough outline is expanded outward by a preset number of pixels to form a bounding box, and the area within the bounding box is defined as the region of interest.

5. The image exposure correction method according to claim 2, characterized in that, The adaptive fusion step is achieved through the following formula: in, For the final exposure-corrected image, The input image, and The intermediate corrected image, , , This represents the weight vector related to the image content generated by the fusion network.

6. The image exposure correction method according to claim 1, characterized in that, The method further includes: Determine whether the color balance of the exposure-corrected image meets the requirements; If the color balance of the exposure-corrected image does not meet the requirements, the exposure-corrected image is input into the prediction neural network for inference to predict the required red channel gain value and blue channel gain value for the current image. The exposure-corrected image is corrected using a correction formula based on the parameters of the red channel gain value and the blue channel gain value to obtain a re-corrected exposure-corrected image.

7. The image exposure correction method according to claim 6, characterized in that, The prediction neural network is an improved network based on the ResNet34 architecture, and its improvements include: Replace the activation function from ReLU with Leaky ReLU; Replace the global average pooling operation at the end of the network with a local pooling operation; The final output layer of the network is a fully connected layer with an output dimension of 2, corresponding to the gain values ​​of the red channel and the blue channel, respectively. The predictive neural network is trained using a combined loss function; The combined loss function includes: in, The loss function is Log-Cosh. For CosineEmbedding loss function, This is the combined loss function; The correction formula includes in , , These represent the red, green, and blue channel components of the exposure-corrected image, respectively. The image is shown as the exposure-corrected image after further correction.

8. An image exposure correction system, characterized in that, The system includes: The decomposition module is used to acquire the input image to be corrected, and to decompose the input image and its inverse image into multiple scales to obtain the corresponding low-frequency components and high-frequency components respectively. The processing module is used to perform exposure correction processing on the low-frequency components using a neural network and to perform detail enhancement processing on the high-frequency components. The reconstruction module is used to reconstruct the processed low-frequency components and high-frequency components to obtain at least one intermediate corrected image; The fusion module is used to adaptively fuse the input image with at least one of the intermediate correction images to output an exposure correction image.

9. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the image exposure correction method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the image exposure correction method as described in any one of claims 1 to 7.

Images