Chip positioning method and system based on multi-focus image fusion

Abstract

The application discloses a chip positioning method and system based on multi-focus image fusion, and is applied to accurate positioning of integrated circuit chips. The method comprises the following steps: real-time monitoring of chip incoming signal, chip image acquisition; construction of chip image dataset and pre-processing, generation of training set and test set data; construction of a network model based on multi-focus image fusion, and training of the model through pre-processed data; fusion of multi-focus chip images by using the trained model, extraction of chip main bodies in the fused images and segmentation of backgrounds, and realization of accurate positioning. The application improves the definition and positioning accuracy of chip images, has high efficiency and robustness, and can be widely applied to the fields of semiconductor manufacturing and automatic detection. A corresponding system is also provided, realizing full-process automation from image acquisition to positioning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chip positioning, and in particular to a chip positioning method and system based on multi-focus image fusion. BACKGROUND

[0002] In recent years, with the wide application of chips in high-precision fields such as precision manufacturing, navigation systems and automatic control, chip positioning, as a basic problem in chip detection, has become increasingly important. The goal of chip positioning is to extract basic information from the overhead view of the chip, such as chip size, type (SOP, QFP, MBF) and angle with the horizontal plane. However, the pin damage problem often encountered in manual operation can directly affect the use of chips and weaken system performance, and an intelligent chip positioning and grasping solution is urgently needed.

[0003] Existing chip positioning technologies mainly include traditional methods and deep learning-based methods. Traditional methods use shape and grayscale features of images for screening, but it is difficult to meet the accuracy and efficiency requirements when facing complex and diverse chips. Deep learning-based methods use data sets and appropriate network models to achieve widespread application in image fusion, image segmentation and salient object detection. Currently, the advantages of such methods are gradually emerging, for example, the improved YOLOv8-seg model can realize real-time detection and segmentation of chips, reducing the computational load while maintaining high accuracy and speed, meeting the needs of industrial automated detection.

[0004] However, in practical applications, existing technologies still have deficiencies. Different types of chips with varying thicknesses are often mixed on the loading table, which makes it impossible for the depth of field of the imaging device to focus on all chips at the same time, and some chip images are out of focus, making it difficult for existing algorithms to accurately segment the main body of the chip. This out-of-focus problem has become a key bottleneck restricting the further development of deep learning algorithms in the field of chip positioning. In addition, with the increasing complexity of chip design, traditional layout methods are difficult to meet the performance, power and area (PPA) optimization requirements. For example, deep reinforcement learning (DRL) based layout optimization methods can dynamically adjust the netlist layout through experience learning, but they cannot effectively solve the chip positioning accuracy problem caused by out-of-focus.

[0005] In summary, existing chip positioning methods face the problem of image out-of-focus caused by different chip thicknesses, and there is an urgent need for a more robust and efficient technical solution to meet the high-precision chip positioning requirements in complex industrial scenarios. SUMMARY

[0006] To solve at least one of the above technical problems, the present application proposes a chip positioning method and system based on multi-focus image fusion.

[0007] The chip positioning method based on multi-focus image fusion comprises the following steps:

[0008] Real-time monitoring of the incoming signal of the chip, and collecting a chip image according to the incoming information of the chip;

[0009] Constructing a chip image dataset according to the chip image, pre-processing the chip image dataset, and constructing training set image data and test set image data to obtain pre-processed image data;

[0010] Constructing a network model based on multi-focus image fusion, and training the network model based on the pre-processed image data;

[0011] According to the trained network model, multi-focus chip image fusion is performed, the fused image is subjected to chip main body and background segmentation, and the chip is positioned according to the segmented main body and background.

[0012] In this scheme, the real-time monitoring of the incoming signal of the chip, and collecting a chip image according to the incoming information of the chip, specifically comprises:

[0013] Obtaining the depth of field adjustable range data of the image taking device, adjusting the depth of field of the image taking device to the maximum value according to the depth of field adjustable range data, and adjusting the focal length of the image taking device, and obtaining the focusing area of the image taking device;

[0014] Obtaining two groups of pre-set thickness chips for image acquisition, real-time monitoring of the incoming signal of the chip, and identifying the chip position information according to the incoming signal of the chip, and collecting a chip image according to the image taking device when the chip reaches the focusing area.

[0015] In this scheme, the chip image dataset is constructed according to the chip image, the chip image dataset is pre-processed, and training set image data and test set image data are constructed to obtain pre-processed image data, specifically comprising:

[0016] The collected chip image is constructed into a chip image dataset, and the chip image dataset is divided into training set image data and test set image data according to a pre-set proportion, and the pre-set proportion is 80% and 20% respectively;

[0017] Each image in the training set image data is cut into a plurality of non-overlapping image blocks of a pre-set pixel size to obtain pre-processed image data.

[0018] In this scheme, the network model based on multi-focus image fusion is constructed, and the network model is trained based on the pre-processed image data, specifically comprising:

[0019] The network model based on multi-focus image fusion is constructed, and the preprocessed image data is introduced into the reinforced dense connection module of the network model to extract enhanced high-frequency and low-frequency features;

[0020] The enhanced high-frequency and low-frequency features are mixed and input into the reinforced network respectively to extract reinforced high-frequency features and reinforced low-frequency features;

[0021] The reinforced high-frequency features and the reinforced low-frequency features are fused based on the cross attention module of DeepLabv3 prior;

[0022] The loss function of the network model in the training process is calculated in real time, and the weight parameters of the network model are optimized.

[0023] In the scheme, the preprocessed image data is introduced into the reinforced dense connection module to extract enhanced high-frequency and low-frequency features, specifically:

[0024] The preprocessed image data is introduced into the reinforced dense connection module, and the reinforced dense connection module includes a dense convolution layer, a transformation layer and a reinforced module layer, and the dense convolution layer and the transformation layer are connected in parallel;

[0025] The dense convolution layer and the transformation layer are connected in parallel;

[0026] The dense convolution layer and the transformation layer are connected in parallel;

[0027] In the scheme, the preprocessed image data is introduced into the reinforced dense connection module to extract enhanced high-frequency and low-frequency features, specifically:

[0028] The enhanced high-frequency and low-frequency feature maps are mixed and input into the reinforced network to extract reinforced high-frequency features and reinforced low-frequency features, and the reinforced network calculation formula is as follows:

[0029] F EH =E(F H +F L )×F H ,

[0030] F EL =E(F H +F L )×F L ,

[0031] Wherein, F EH represents the enhanced high-frequency feature, F ELdenotes the enhanced low-frequency feature, F H denotes the high-frequency feature before enhancement, F L denotes the low-frequency feature before enhancement, E denotes a spatial attention enhancement operation.

[0032] In this scheme, the cross-attention module based on DeepLabv3 prior fuses the enhanced high-frequency feature and the enhanced low-frequency feature, specifically:

[0033] The enhanced high-frequency feature and the enhanced low-frequency feature are introduced into the cross-attention module based on DeepLabv3 prior, and the attention calculation formula of the cross-attention module is as follows:

[0034]

[0035] Wherein, Q, K, V represent query, key and value respectively, the query, key and value are calculated by multiplying the input enhanced high-frequency feature and the enhanced low-frequency feature with trainable matrix, d k is the dimension of K, B is a relatively learnable position encoding, and softmax is a normalization function;

[0036] By executing the attention function h times and connecting the results to the multi-head attention, the feature marks generated by the multi-head attention layer are refined by two multi-layer perceptron layers with GELU activation layers;

[0037] After executing the multi-head attention and before the multi-layer perceptron, the enhanced high-frequency feature and the enhanced low-frequency feature are introduced into the normalization layer for normalization operation, and the normalization operation is represented as:

[0038]

[0039] Given two local window features X1 and X2 from different domains, the weight matrix W Q , W V According to the normalization operation projected into Q, K and V, the normalized enhanced high-frequency feature and the normalized enhanced low-frequency feature are fused based on the cross-attention module, and the fusion process is:

[0040]

[0041] For Q1 from focal length 1, it merges the cross-domain information by performing attention weighting on K2 and V2 from focal length 2, while retaining the information in focal length 1 through residual connection, where MHA is a multi-head attention layer.

[0042] When the complementary information fusion of different focal lengths is completed, the CNN-based image reconstruction unit maps the fused enhanced high-frequency feature and the fused enhanced low-frequency feature back to the image space to generate a fused image If .

[0043] In this scheme, the loss function of the network model in the training process is calculated in real time, and the weight parameters of the network model are optimized, specifically:

[0044] In the network model training process, by continuously reducing the error between the prediction quality score and the target quality score output by the network model, the ReLU activation function and the Adam optimizer are used as the optimization algorithm for automatically updating the learning rate of the network model, wherein the calculation formula of the loss function is:

[0045] L = γL aux + μL texture ,

[0046] Where L aux represents auxiliary intensity loss, L texture represents gradient loss, γ and μ are weight factors for controlling the contribution of each term, and the gradient loss is designed to suppress the important details from the input image. The formula is as follows:

[0047]

[0048] Where H and W represent the height and width of the source image respectively, ‖·‖ represents the l1 norm, max(·) represents the maximum value selection element by element, I f is the fusion result image, I1 and I2 are the source images in two focus states respectively.

[0049] In this scheme, the network model is trained based on the network model after training, and the chip image fusion is performed. The chip main body and background are segmented, and the chip is positioned according to the segmented main body and background, specifically:

[0050] Obtain the chip top view under different focal lengths, import the chip top view into the trained multi-focus image fusion network model, and segment the chip main body and chip pin from the background according to the different gray values by using the YOLOv8-seg segmentation algorithm;

[0051] Based on the feature extraction algorithm, the chip pixel coordinates, area, angle and chip type information of the segmented chip main body and chip pin and background are extracted, and the chip is positioned according to the chip pixel coordinates, area, angle and chip type information.

[0052] The second aspect of the present application also provides a chip positioning system based on multi-focus image fusion, comprising a memory and a processor, the memory comprising a chip positioning method based on multi-focus image fusion program, and the chip positioning method based on multi-focus image fusion is implemented when the processor executes the chip positioning method based on multi-focus image fusion program.

[0053] The present application discloses a chip positioning method and system based on multi-focus image fusion, which is applied to the accurate positioning of integrated circuit chips. The method comprises the following steps: real-time monitoring of chip incoming signal and acquisition of chip image; construction of chip image dataset and pre-processing to generate training set and test set data; construction of a network model based on multi-focus image fusion and training of the model through pre-processed data; fusion of multi-focus chip images using the trained model, extraction of chip main body in the fused image and segmentation of the background to achieve accurate positioning. The present application improves the clarity and positioning accuracy of chip images, has high efficiency and robustness, and can be widely applied to the fields of semiconductor manufacturing and automated detection. The corresponding system is also provided to realize the full-process automation from image acquisition to positioning. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A flowchart of a chip positioning method based on multi-focus image fusion of the present application is shown;

[0055] Figure 2 A schematic diagram of an image capturing system of the present application is shown;

[0056] Among them, Figure 2 Among them,

[0057] 1. Optical camera, 2. Medium hole light source, 3. Lens, 4. IC chip, 5. Feeding platform.

[0058] Figure 3 A network structure schematic diagram of a multi-focus image fusion model of the present application is shown. DETAILED DESCRIPTION

[0059] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0060] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the protection scope of the present application is not limited by the specific embodiments disclosed below.

[0061] Figure 1A flow chart of a chip positioning method based on multi-focus image fusion is shown.

[0062] As shown in the figure, Figure 1 the first aspect of the present application provides a chip positioning method based on multi-focus image fusion, comprising:

[0063] S102, real-time monitoring of the incoming signal of the chip, and collecting chip images according to the incoming information of the chip;

[0064] S104, constructing a chip image dataset according to the chip image, pre-processing the chip image dataset, and constructing training set image data and test set image data to obtain pre-processed image data;

[0065] S106, constructing a network model based on multi-focus image fusion, and training the network model based on the pre-processed image data;

[0066] S108, multi-focus chip image fusion based on the trained network model, chip main body and background segmentation of the fused image, and chip positioning based on the segmented main body and background.

[0067] According to the embodiment of the present application, the S102 is specifically:

[0068] Obtaining the depth of field adjustable range data of the image taking device, adjusting the depth of field of the image taking device to the maximum value according to the depth of field adjustable range data, and adjusting the focal length of the image taking device to obtain the focusing area of the image taking device;

[0069] Obtaining two groups of chips with preset thickness for image acquisition, real-time monitoring of the incoming signal of the chip, identifying the chip position information according to the incoming signal of the chip, and collecting chip images according to the image taking device when the chip reaches the focusing area.

[0070] According to the embodiment of the present application, the S104 is specifically:

[0071] Obtaining training dataset and pre-processing it; wherein the training dataset contains 2 groups, adjusting the image taking device to make the depth of field maximum, adjusting the focal length to make the thickest chip and the thinnest chip respectively at the best focusing position, collecting two images as multi-focus input of the dataset, and the image taking system is as shown in the figure Figure 2 ;

[0072] For the images in the above dataset, they are randomly divided into two parts, of which 80% of the image samples are used for training and the remaining 20% are used for testing. In this way, the model can be effectively verified on both training and test data.

[0073] For the images used for training, the images are cropped into multiple non-overlapping 64x64 pixel image blocks. This helps the network learn more detailed features and improve the accuracy of image fusion.

[0074] According to an embodiment of the present application, the S106, in particular:

[0075] The cropped image blocks are input into an enhance dense block (EDB) in the multi-focus image fusion model. Each EDB has a dense convolutional neural network (DC), a transformer (TR), and an enhance module (EM). In order to prevent interference between the extraction of high-frequency features and low-frequency features, the DC layer and the TR layer use a parallel connection method. The inputs of the DC layer and the TR are input into the EM layer through weighted splicing. The spatial attention mechanism of the EM layer enhances the low-frequency and high-frequency features respectively. After four EDB layers, a 48x16-dimensional feature map is obtained.

[0076] The enhanced low-frequency and high-frequency feature maps are mixed and input into the enhancement network, and the following formula is calculated:

[0077] F EH =E(F H +F L )×F H ,

[0078] F EL =E(F H +F L )×F L ,

[0079] where F EH represents the enhanced high-frequency feature, F EL represents the enhanced low-frequency feature, F H represents the high-frequency feature before enhancement, F L represents the low-frequency feature before enhancement, and E represents the spatial attention enhancement operation.

[0080] The enhanced high-frequency feature F EH and the enhanced low-frequency feature F EL are input into the cross attention module (CA) based on DeepLabv3 prior. The attention calculation formula is as follows:

[0081]

[0082] where Q, K, V represent Query, Key and Value respectively, are computed by input multiplied by trainable matrices respectively, d k is the dimension of K, B is a relatively learnable position encoding, and softmax is a normalization function.

[0083] By performing h times attention function and concatenating the results to multi-head self-attention, where h is set to 4 in the work. Next, the feature tokens generated by the Multi Head Attention (MHA) layer are refined by two Multilayer Perceptron (MLP) layers with GELU activation layer. The Layler Normalization (LN) is performed after MHA and before MLP. The complete process can be represented as:

[0084]

[0085] Given two local window features X1 and X2 from different domains, the weight matrix W Q , W V is used to project them into Q, K and V. The whole process of cross-attention based fusion is formalized as:

[0086]

[0087] For Q1 from focal length 1, it merges the cross-domain information by performing attention weighting on K2 and V2 from focal length 2, while preserving the information in focal length 1 through a residual connection, and vice versa.

[0088] After fully fusing the complementary information of different focal lengths, we design a CNN-based image reconstruction unit to map the fused depth features back to the image space. Deploying a CNN-based image reconstruction unit H CR to reduce the number of channels and generate the fused image I f , which is represented as:

[0089] I f = H CR (F D ),

[0090] The loss function is calculated, and the weight parameters are optimized. When training the network model, the error between the prediction quality score and the target quality score output by the network model is continuously reduced, and the weight parameters of the network model are continuously optimized. The activation function uses the ReLU activation function, the optimization algorithm uses the Adam (Adaptive moment estimation) automatic learning rate update algorithm, the batch-size size is set to 64, the learning rate is fixed to 0.001, and the loss function calculation formula is as follows:

[0091] L = γL aux + μL texture ,

[0092] Wherein, L aux represents the auxiliary intensity loss, L texture represents the gradient loss. γ and μ are weight factors for controlling the contribution of each term. The gradient loss is designed to suppress the fusion result to retain important details from the input image, and its formula is as follows:

[0093]

[0094] Where H and W represent the height and width of the source image respectively. ‖·‖ represents the l1 norm, max(·) means the maximum value selection element by element, and I f is the fusion result image, I1 and I2 are the source images in two focus states respectively.

[0095] For the auxiliary intensity loss, an excellent image fusion algorithm expects to generate a fusion image with appropriate intensity according to the global apparent intensity information of the source image. To this end, we designed the following auxiliary intensity loss to guide our fusion model to capture appropriate intensity information:

[0096]

[0097] According to the embodiment of the present application, the S108, specifically:

[0098] The trained model is used to fuse the chip top view under different focal lengths, the chip main body and chip pins are segmented from the background according to the different gray values through the YOLOv8-seg segmentation algorithm, and the basic information such as pixel coordinates, area, angle and chip type of the chip is extracted through the feature extraction algorithm.

[0099] The second aspect of the present application also provides a chip positioning system based on multi-focus image fusion, which comprises a memory and a processor, the memory comprises a chip positioning method based on multi-focus image fusion program, and the chip positioning method based on multi-focus image fusion is executed by the processor to realize the steps of the chip positioning method based on multi-focus image fusion according to any one of the above.

[0100] The application discloses a chip positioning method and system based on multi-focus image fusion, and is applied to accurate positioning of integrated circuit chips.

[0101] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. The described device embodiments are only schematic. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling, direct coupling or communication connection between the components can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0102] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units; they can be located in one place or distributed on multiple network units; part or all of the units can be selected according to actual needs to achieve the purpose of the present embodiment.

[0103] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be realized in the form of hardware or in the form of hardware plus software functional unit.

[0104] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, the foregoing program can be stored in a computer readable storage medium, and the program executes the steps of the method embodiments when executed; and the foregoing storage medium includes a mobile storage device, a read-only memory (ROM), a random access memory (RAM), a magnetic disc or an optical disc, and various storage medium capable of storing program codes.

[0105] Alternatively, the integrated unit of the present application can be stored in a computer readable storage medium if it is realized in the form of a software function module and sold or used as an independent product. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes a mobile storage device, a ROM, a RAM, a magnetic disc or an optical disc, and various storage medium capable of storing program codes.

[0106] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A chip positioning method based on multi-focus image fusion, characterized in that, The method comprises the following steps: Real-time monitoring of the incoming signal of the chip, collecting the chip image according to the incoming information of the chip, specifically: Obtain the depth of field adjustable range data of the image taking device, adjust the depth of field of the image taking device to the maximum according to the depth of field adjustable range data, and adjust the focal length of the image taking device to obtain the focusing area of the image taking device; Obtain two groups of preset thickness chips for image acquisition, real-time monitoring of the incoming signal of the chip, identifying the chip position information according to the incoming signal of the chip, when the chip reaches the focusing area, collecting the chip image according to the image taking device; According to the chip image, a chip image dataset is constructed, the chip image dataset is preprocessed, and a training set image data and a test set image data are constructed to obtain preprocessed image data; A network model based on multi-focus image fusion is constructed, and the network model is trained based on the preprocessed image data, specifically: Construct a network model based on multi-focus image fusion, and import the preprocessed image data into the reinforced dense connection module of the network model to extract enhanced high-frequency and low-frequency features; The enhanced high-frequency and low-frequency features are mixed and input into the reinforced network to extract reinforced high-frequency and low-frequency features; The reinforced high-frequency and low-frequency features are fused based on the cross attention module of DeepLabv3 prior; Calculate the loss function of the network model in the training process in real time, and optimize the weight parameters of the network model; According to the trained network model, multi-focus chip image fusion is performed, the chip main body and background are segmented after image fusion, and the chip is positioned according to the segmented main body and background, specifically: Obtain the chip top view under different focal lengths, import the chip top view into the trained multi-focus image fusion network model, and segment the chip main body and chip pin from the background according to the different gray values through the YOLOv8-seg segmentation algorithm; Based on the feature extraction algorithm, the chip pixel coordinates, area, angle and chip type information of the segmented chip main body and chip pin are extracted, and the chip is positioned according to the chip pixel coordinates, area, angle and chip type information.

2. The chip positioning method based on multi-focus image fusion according to claim 1, characterized in that, The chip image dataset is constructed, the chip image dataset is divided into training set image data and test set image data according to a preset proportion, and the preset proportion is 80% and 20% respectively; Each image in the training set image data is cropped into a plurality of non-overlapping image blocks of a preset pixel size to obtain preprocessed image data. The preprocessed image data is imported into the reinforced dense connection module of the network model to extract enhanced high-frequency and low-frequency features, specifically:

3. The chip positioning method based on multi-focus image fusion according to claim 1, characterized in that, ​ The preprocessed image data is imported into a reinforced dense connection module, the reinforced dense connection module comprising a dense convolutional layer, a transformation layer and a reinforced module layer, the dense convolutional layer being connected in parallel with the transformation layer; The dense convolutional layer and the transformation layer are weighted and spliced by receiving the imported preprocessed image data, and the weighted and spliced preprocessed image data is input into the reinforced module layer; The reinforced module layer performs reinforcement operation on the low-frequency and high-frequency features of the preprocessed image data according to the spatial attention mechanism, to obtain high-frequency and low-frequency features of a preset dimension.

4. The chip positioning method based on multi-focus image fusion according to claim 1, characterized in that, The enhanced high-frequency and low-frequency features are mixed and input into a reinforced network to extract reinforced high-frequency features and reinforced low-frequency features, specifically: The enhanced high-frequency and low-frequency features are mixed and input into a reinforced network to extract reinforced high-frequency features and reinforced low-frequency features, and the reinforced network calculation formula is as follows: ， ， wherein, represents an enhanced high-frequency feature, represents an enhanced low-frequency feature, represents a pre-enhanced high-frequency feature, represents a pre-enhanced low-frequency feature, represents a spatial attention enhancement operation.

5. The chip positioning method based on multi-focus image fusion according to claim 1, characterized in that, The cross-attention module based on DeepLabv3 prior fuses the reinforced high-frequency features and the reinforced low-frequency features, specifically: The cross-attention module based on DeepLabv3 prior fuses the reinforced high-frequency features and the reinforced low-frequency features, and the attention calculation formula of the cross-attention module is as follows: ， wherein Q, K, V represent query, key and value respectively, which are calculated by inputted reinforced high-frequency features and reinforced low-frequency features multiplied by trainable matrix respectively, is the dimension of K, B is a relatively learnable position encoding, is a normalization function; By executing the attention function h times and connecting the results to the multi-head attention, h is set to 4, and the feature labels generated by the multi-head attention layer are refined by two multi-layer perceptron layers with GELU activation layers; After executing the multi-head attention and before the multi-layer perceptron, the reinforced high-frequency features and the reinforced low-frequency features are imported into a normalization layer for normalization operation, and the normalization operation is represented as: ， ， Given two local window features from different domains and a weight matrix shared across different windows , According to the normalization operation, the normalized reinforcement high-frequency features and reinforcement low-frequency features are projected into Q, K and V, and then fused based on the cross attention module. The fusion process is as follows: ， For Q1 from focus 1, it merges cross-domain information by performing attention weighting on K2 and V2 from focus 2, while preserving the information in focus 1 through a residual connection, where is a multi-head attention layer; When the complementary information fusion of different focal lengths is completed, the CNN-based image reconstruction unit maps the fused enhanced high-frequency features and enhanced low-frequency features back to the image space to generate a fused image .

6. The chip positioning method based on multi-focus image fusion according to claim 1, characterized in that, The system includes a memory and a processor, the memory including a chip positioning method based on multi-focus image fusion program, the chip positioning method based on multi-focus image fusion being executed by the processor to realize the steps of the chip positioning method based on multi-focus image fusion of any one of claims 1-6. The system includes a memory and a processor, the memory including a chip positioning method based on multi-focus image fusion program, the chip positioning method based on multi-focus image fusion being executed by the processor to realize the steps of the chip positioning method based on multi-focus image fusion of any one of claims 1-6. ， where, denotes the auxiliary strength loss, denotes the gradient loss, and is a weight factor for controlling the contribution of each term, by designing the gradient loss to suppress the fusion result to keep important details from the input images, which is formulated as follows: ， ， where H and W represent the height and width of the source image, respectively, denotes norm, refers to the element-wise maximum selection, is the fusion result image, and are the source images in the two focus states, respectively.

7. A multi-focus image fusion based chip positioning system, characterized in that, ​

Images