Polarization three-dimensional imaging method and device for monocular endoscope

Abstract

The application discloses a monocular endoscope-oriented polarization three-dimensional imaging method and device. The method comprises the following steps: firstly, a polarization three-dimensional imaging data set suitable for a medical endoscope scene is constructed; then, based on the polarization characteristics of the tissue structure in the endoscope environment, a neural network model fusing a double-domain cross attention mechanism is designed and constructed. The network is effectively trained by introducing a fusion loss function, and the trained model is saved. Subsequently, the tissue image collected by the endoscope in real time is input into the trained model, and the normal vector information of the tissue surface is output. The normal vector solving method proposed by the application has the advantages of high precision and strong robustness in combination with the physical principle of deep learning and polarization imaging, and is suitable for complex endoscopic medical scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of endoscopy, in particular to a polarization three-dimensional imaging method and device for monocular endoscopy. BACKGROUND

[0002] With the development of minimally invasive surgery technology, endoscopes are increasingly widely used in clinical diagnosis and treatment. However, the traditional monocular endoscopic imaging system can only provide two-dimensional images, lacking accurate perception of tissue depth and three-dimensional topography, which limits the accurate judgment of doctors on the structure of lesions and the fine degree of operation. In order to improve the spatial perception ability of the endoscopic system, researchers have proposed various three-dimensional reconstruction techniques, including binocular vision, structured light, and light field imaging methods. However, these techniques usually face problems such as complex equipment, large system volume, or high requirements for imaging environment, which are not easy to promote in clinical practice.

[0003] Polarization imaging, as a technology that uses light wave vibration direction information to obtain object surface structure and material characteristics, has gradually attracted attention in biomedical imaging in recent years. In particular, when dealing with highly scattering biological tissue surfaces, polarization information can effectively improve the robustness and accuracy of topography reconstruction. Although polarization three-dimensional imaging has unique advantages in obtaining object surface structure information, it still faces many challenges in practical applications, especially in dealing with complex reflection types (such as mixed specular and diffuse reflection) and solving the ambiguity of surface normal vectors.

[0004] Polarization three-dimensional imaging technology based on deep learning trains neural networks in passive lighting conditions to learn the surface normal distribution of the target object through data-driven methods. This approach improves the flexibility and adaptability of normal estimation to some extent. However, existing methods still have obvious shortcomings in restoring the fine texture details of the target surface and accurately estimating the normal vector. During the reconstruction process, the fine texture area of the target may lose information, which may lead to inaccurate normal estimation and reduce the overall reconstruction accuracy. At the same time, the medical training dataset for polarization three-dimensional imaging is very scarce, which to a large extent also limits the improvement of model performance. SUMMARY

[0005] In view of the shortcomings of the prior art, the present application provides a polarization three-dimensional imaging method and device for monocular endoscopy, which constructs a more stable and expressive polarization representation input by approximately decoupling specular reflection and diffuse reflection components and further encoding the polarization angle. This improvement can more comprehensively extract polarization and texture information from the image, effectively reduce the loss of details during reconstruction, and improve the accuracy of surface normal estimation.

[0006] To achieve the above-mentioned purpose, the present application provides a polarization three-dimensional imaging method for monocular endoscopy, comprising the following steps:

[0007] S1: Construct a polarized three-dimensional imaging dataset for medical endoscopes;

[0008] S2: Construct a network with dual-domain cross-attention fusion based on the polarization characteristics of tissues in endoscopic scenes;

[0009] S3: Divide the polarized three-dimensional imaging dataset in step S1 into a training set and a validation set, construct a fusion loss function, train the network in step S2 using the training set, use the validation set to guide model convergence, obtain the trained model and save it;

[0010] S4: Input the tissue images collected in real time by the endoscope into the model saved in step S3, output the tissue surface normal vector, and finally reconstruct the three-dimensional topography of the tissue surface.

[0011] Further, step S1 includes the following sub-steps:

[0012] S11: Build a polarization metrology monocular endoscope imaging system; modify the endoscope so that it is equipped with a polarization camera at the rear end of the imaging system, collect polarization images with polarization angles of 0°, 45°, 90°, and 135° as the intensity branch domain Input1 for network input;

[0013] S12: Build a polarization metrology monocular endoscope and three-dimensional scanner image registration system, fix the endoscope and the scanner to ensure that the imaging poses of the two are the same at each shooting; through stereo calibration, obtain the intrinsic and extrinsic parameters between the endoscope and the scanner, and unify the fields of view of the scanner and the endoscope;

[0014] S13: Use the four polarization images in sub-step S11 to propose a way to approximately calculate the specular reflection component I S and the diffuse reflection component I D :

[0015] ;

[0016] wherein, represents the specular reflection amplitude, represents the constant reflection component; calculate the Stokes vector-based parameters , :

[0017] ;

[0018] wherein, represents the degree of polarization, represents the analyzer angle; , , , These represent polarization images with polarization angles of 0°, 45°, 90°, and 135°, respectively; for the analyzer angle... Further encoding yields:

[0019] ;

[0020] in, Indicates the encoding polarization angle;

[0021] Based on the above , , , The physical branch domain Input2 consists of four physical priors that serve as the network input.

[0022] S14: Capture the three-dimensional point cloud of the tissue using a scanner, and convert the three-dimensional point cloud into a point cloud under the endoscopic field of view using the external parameters calibrated in sub-step S12; then calculate the surface normal vector field based on the difference method, which is used as the true value Output of the network output.

[0023] S15: Randomly crop the input image into N images to augment the dataset.

[0024] Furthermore, in step S2, the network with dual-domain cross-attention fusion uses UNet as the backbone network, which consists of an encoder and a decoder.

[0025] Furthermore, the network with dual-domain cross-attention fusion includes a feature extraction module, a cross-attention hybrid domain module, a fusion module, and a normal vector output module;

[0026] The feature extraction module adopts the encoder of UNet. In the encoder stage, the intensity branch domain Input1 and the physical branch domain Input2 are respectively input into the residual network to extract their respective local features. The feature map after downsampling by the residual network enters the cross-attention hybrid domain module for learning, so as to explore the potential correlation between branch domains and perceive global information. The decoder cascades multiple upsampling blocks to obtain feature maps. The feature map of the encoder and the feature map of the decoder are fused by the fusion module to obtain fused features. The normal vector output module converts the fused features into a normal vector for each pixel.

[0027] Furthermore, the Bi-Fusion module is used as a fusion module, mainly composed of channel attention and spatial attention; the spatial attention uses a CBAM block, which acts as a spatial filter to enhance local details and suppress irrelevant regions; the channel attention uses an SE block to enhance the ability to extract local details; both channel attention and spatial attention include pooling, fully connected, and activation operations.

[0028] Further, in step S3, the network is trained by constructing a fusion loss function defined as:

[0029] ;

[0030] wherein is a cosine consistency loss function, is a gradient continuity loss function; respectively represent , weight parameters of, used to regulate the weight of different loss terms.

[0031] Further, in step S3, the polarized three-dimensional imaging data set of step S1 is divided into a training set and a validation set; the network with double-domain cross-attention fusion in step S2 is preliminarily trained using the training set, and the validation set is used to guide the model to converge; during the training process, the learning rate is dynamically adjusted, and a front-large-later-small gradient strategy is adopted to accelerate the early convergence speed and stabilize the later optimization process; at the same time, an early stopping mechanism is introduced, when the model fusion loss function is less than the set threshold within the set training number, the training is terminated in advance to prevent overfitting.

[0032] To achieve the above purpose, the application further provides a polarized three-dimensional imaging device for monocular endoscope, comprising one or more processors for realizing the polarized three-dimensional imaging method for monocular endoscope.

[0033] To achieve the above purpose, the application further provides an electronic device, comprising a memory and a processor, the memory is coupled with the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to realize the polarized three-dimensional imaging method for monocular endoscope.

[0034] To achieve the above purpose, the application further provides a computer readable storage medium, which stores a computer program, the program is executed by a processor to realize the polarized three-dimensional imaging method for monocular endoscope.

[0035] The beneficial effects of the application are as follows:

[0036] 1、The application decouples the specular reflection and diffuse reflection components by approximation, and further encodes the polarization angle, to construct a more stable and expressive polarized representation input.

[0037] 2、The application constructs a network with double-domain cross-attention fusion, which can more comprehensively extract the polarization and texture information in the image through the extraction and fusion of the double-domain branch, effectively reduces the detail loss in the reconstruction process, and improves the accuracy of surface normal estimation.

[0038] 3、The fusion loss function constructed in the application comprehensively considers the overall accuracy and spatial continuity of the output normal vector, effectively reduces the artifact phenomenon, and improves the continuity of the structure, thereby helping to more accurately restore the three-dimensional topography of the surface. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flow chart of a monocular endoscope-oriented polarization three-dimensional imaging method of an embodiment of the application;

[0040] Figure 2 is a network framework structure diagram of a network framework with double-domain cross-attention fusion proposed in an embodiment of the application;

[0041] Figure 3 is a network structure schematic diagram of a cross-attention module proposed in an embodiment of the application;

[0042] Figure 4 is a network structure schematic diagram of a Bi-Fusion fusion module proposed in an embodiment of the application;

[0043] Figure 5 is a normal vector prediction result map for tissue of an embodiment of the application;

[0044] Figure 6 is a structure schematic diagram of a monocular endoscope-oriented polarization three-dimensional imaging device in an embodiment of the application;

[0045] Figure 7 is a schematic diagram of an electronic device in an embodiment of the application. DETAILED DESCRIPTION

[0046] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers are used to indicate the same or similar components. The embodiments described in the following exemplary embodiments do not represent all the embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0047] The present application will be described in detail below with reference to the accompanying drawings and preferred embodiments, and the objects and effects of the present application will become more apparent. It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0048] As shown in Figure 1 , the present application provides a monocular endoscope-oriented polarization three-dimensional imaging method, comprising the following steps:

[0049] S1: Construct a polarization three-dimensional imaging data set for a medical endoscope.

[0050] Step S1 specifically comprises the following sub-steps:

[0051] S11: Build a polarization metrology monocular endoscope imaging system; retrofit the endoscope so that its imaging rear end is equipped with a polarization camera that can collect polarization images at four angles as the intensity branch domain Input1 of the network input, as shown in Figure 2 . The relationship between the light intensity observed by polarization imaging and the polarization angle of the incident light can be expressed as:

[0052]

[0053] wherein, and represent the upper and lower limits of , respectively, represents the analyzer (such as a polaroid) angle, represents the diffuse reflection component, represents the non-polarized mirror reflection component, represents the mirror reflection amplitude, represents the constant reflection component.

[0054] S12: Build a polarization metrology monocular endoscope and three-dimensional scanner image registration system, use a custom bracket to fix the endoscope and scanner to ensure that the imaging poses of the two are the same at each shooting. Through stereo calibration, the intrinsic and extrinsic parameters between the endoscope and the scanner are obtained, so that the field of view of the scanner and the field of view of the endoscope can be unified.

[0055] S13: Use the four polarization images in sub-step S11 to propose a way to approximately calculate the mirror reflection component I S and the diffuse reflection component I D :

[0056]

[0057] The degree of polarization , the analyzer angle are parameters based on Stokes vectors.

[0058]

[0059] wherein, , , , represent the polarization images with polarization angles of 0°, 45°, 90°, and 135°, respectively; the analyzer angle is further encoded to address the disadvantage of polarization angle π wrapping.

[0060]

[0061] wherein, denotes the encoding polarization angle.

[0062] Based on the above , , , 4 physical priors as the physical branch domain Input2 of the network input, as shown in Figure 2 .

[0063] S14: Capture the three-dimensional point cloud of the tissue by a high-precision scanner, and convert it into a point cloud under the field of view of the endoscope by using the external parameters calibrated in sub-step S12. Then, calculate the surface normal vector field based on the difference method as the true value Output of the network output.

[0064] S15: In order to preserve the original physical information of the polarization data, the input image is randomly cut into N images, further realizing the enhancement of the data set.

[0065] S2: Based on the polarization characteristics of the tissue under the endoscopic scene, a network with double-domain cross-attention fusion is constructed.

[0066] The network architecture includes a feature extraction module, a cross-attention hybrid domain module, a fusion module, and a normal vector output module, as shown in Figure 2 .

[0067] Unet is used as the backbone network, which is a U-shaped network architecture composed of an encoder and a decoder. The feature extraction module follows the encoder of UNet. In the encoder stage, the intensity branch domain Input1 and the physical branch domain Input2 are input into ResNet (Residual Network) respectively to extract their respective local features. The ResNet downsampling is divided into four stages, and the resolution of the output feature map E is H / 2×W / 2×64, H / 4×W / 4×128, H / 8×W / 8×256, and H / 16×W / 16×512 in turn.

[0068] The down-sampled feature map enters the cross-attention hybrid domain module to learn to excavate the potential correlation between the branch domains and perceive the global information, as shown in Figure 3 . This is a bidirectional learning process to improve the feature quality of the double domains. Cross-attention is constructed using Query, Key, and Value, which are obtained by 1×1 convolution and reshaping operation: . The discriminative feature transformation from the intensity branch domain to the physical branch domain is denoted as :

[0069]

[0070] wherein denotes a pair of feature maps; , , denote the height, width, and channel number of the image respectively, d k is a scale factor, denotes a transpose operation. In Q 1 and K 2 , the correlation between the two domains is established, and then the attention matrix is generated by Softmax(·). The mixed domain feature is generated by V 2 weighting. In order to learn different attention distributions, the cross-attention is extended to a multi-head structure.

[0071]

[0072] wherein , is the corresponding weight, and n is the number of attention heads, and the present application takes n = 8; denotes cross-attention. Similarly, is a discriminative feature transformation from the physical branch domain to the intensity branch domain, which can be easily calculated by the same method. Finally, D and are connected to output D.

[0073] The decoder cascades multiple up-sampling blocks to obtain feature maps, and fuses the features of the encoder part and the features of the decoder through a fusion module.

[0074] The Bi-Fusion module is used as a fusion module to fuse the feature maps of the encoder and the decoder, as shown in Figure 4 , mainly composed of channel attention and spatial attention. The spatial attention adopts CBAM (Convolutional Block Attention Module), which is used as a spatial filter to enhance local details and suppress irrelevant areas. Since the convolutional features have weak global context processing ability and are easily disturbed by noise, the spatial attention is processed on the feature maps of the encoder. The channel attention adopts SE (Squeeze-and-Excitation), which can enhance the ability to extract local details, and the decoder feature maps processed by the cross-attention mixed domain module are relatively easy to lack local details, so the channel attention is used for processing. Both channels contain pooling, fully connected, and activation operations. Specifically, the fusion feature F is obtained by the following operation:

[0075]

[0076] where, and are the weight parameters of and , is the Hadamard product, Conv is a 3x3 size convolutional layer, is the interaction feature, denotes the channel attention output, denotes the channel attention input, denotes the spatial attention output, denotes the spatial attention input.

[0077] The resulting feature F can effectively capture the global and local context of the current spatial resolution.

[0078] Normal vector output module: This module is composed of a fully connected layer, which is used to convert the feature F obtained by the fusion module into the normal vector of each pixel. It uses the fused global and local features to generate the final normal vector prediction, and is closely related to the previous modules - the feature extraction provides detailed information, the cross attention enhances the discriminative ability, and the fusion module provides spatial context, so as to ensure the accuracy and physical rationality of the normal vector.

[0079] S3: Construct a fusion loss function to train the network, obtain a trained model and save it.

[0080] The fusion loss function is defined as:

[0081]

[0082] where, is the cosine consistency loss function, is the gradient continuity loss function. are the weight parameters of , , which are used to regulate the weight of different loss terms.

[0083]

[0084] where, , denote the coordinate parameters, is the ground truth of the normal vector, is the normal vector output in the testing stage of the deep network, is the gradient calculation function.

[0085] The polarized three-dimensional imaging data set obtained in step S1 is divided into a training set and a verification set in a ratio of 9:1. First, the network with the dual-domain cross-attention fusion in step S2 is preliminarily trained using the training set, and the verification set is used to guide the model to converge. During the training process, the learning rate is dynamically adjusted, and a "large front and small back" gradient strategy is adopted to speed up the early convergence speed and stabilize the later optimization process. At the same time, an early stopping mechanism is introduced, when the model converges within a set number of training rounds, the fusion loss function Loss is less than the set threshold, the training is terminated in advance to prevent overfitting. The trained network model is saved.

[0086] S4: input the tissue image collected in real time by the endoscope into the model saved in step S3, and output the tissue surface normal vector, as shown in formula (1). Figure 5 The method provided by the present application can accurately estimate the target normal vector to restore the three-dimensional topography, and provides important theoretical support and practical value for high-quality three-dimensional endoscopic imaging in clinical practice.

[0087] Corresponding to the foregoing embodiments of the polarized three-dimensional imaging method for monocular endoscopes, the present application also provides embodiments of a polarized three-dimensional imaging device for monocular endoscopes.

[0088] Referring to Figure 6 , the polarized three-dimensional imaging device for monocular endoscopes provided by the embodiments of the present application comprises one or more processors for implementing the polarized three-dimensional imaging method for monocular endoscopes in the foregoing embodiments.

[0089] The embodiments of the polarized three-dimensional imaging device for monocular endoscopes of the present application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiments can be realized by software, or by hardware or a combination of software and hardware. Taking software realization as an example, as a logically meaningful device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the device with data processing capability where the device is located. From the hardware level, as shown in Figure 6 , it is a hardware structure diagram of the device with data processing capability where the polarized three-dimensional imaging device for monocular endoscopes of the present application is located. In addition to the processor, the memory, the network interface, and the non-volatile memory shown in Figure 6 , the device with data processing capability where the device in the embodiments is located usually includes other hardware according to the actual functions of the device with data processing capability, and details are not described here.

[0090] The implementation process of the functions and roles of each unit in the above device is specifically described in the implementation process of the corresponding steps in the above method, and is not described here.

[0091] For the apparatus embodiment, since it basically corresponds to the method embodiment, the relevant part can be seen from the part of the method embodiment. The apparatus embodiment described above is only illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purpose of the present application scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0092] Corresponding to the foregoing embodiment of the polarization three-dimensional imaging method for monocular endoscope, the embodiment of the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the polarization three-dimensional imaging method for monocular endoscope as described above. As shown in the Figure 7 The polarization three-dimensional imaging method for monocular endoscope provided by the embodiment of the present application is a hardware structure diagram of any data processing capable device. In addition to the processor, memory, DMA controller, disk, and non-volatile memory shown in Figure 7 In addition to the processor, memory, DMA controller, disk, and non-volatile memory shown in the foregoing embodiment, any data processing capable device in which the apparatus is located usually includes other hardware according to the actual function of the data processing capable device, which will not be described again.

[0093] Corresponding to the foregoing embodiment of the polarization three-dimensional imaging method for monocular endoscope, the embodiment of the present application also provides a computer readable storage medium having a program stored thereon, which is executed by a processor to implement the polarization three-dimensional imaging method for monocular endoscope in the foregoing embodiment.

[0094] The computer readable storage medium can be an internal storage unit of any data processing capable device, such as a hard disk or a memory. The computer readable storage medium can also be any data processing capable device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both an internal storage unit of any data processing capable device and an external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the data processing capable device, and can also be used to temporarily store data that has been output or will be output.

[0095] Those skilled in the art can understand that the above description is only the preferred examples of the present application and is not used to limit the present application, and although the present application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacements for part of the technical features. Any modification, equivalent replacement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A polarization three-dimensional imaging method for monocular endoscope, characterized by, The method comprises the following steps: S1: constructing a polarization three-dimensional imaging data set for a medical endoscope; S2: constructing a network with double-domain cross-attention fusion based on the polarization characteristics of tissues in an endoscopic scene; The network with double-domain cross-attention fusion adopts UNet as a backbone network and is composed of an encoder and a decoder; The network with double-domain cross-attention fusion comprises a feature extraction module, a cross-attention mixed domain module, a fusion module and a normal vector output module; Polarization images with polarization angles of 0°, 45°, 90° and 135° are collected as an intensity branch domain Input1 input to the network, and four physical priors including a specular reflection component, a diffuse reflection component, parameters of a Stokes vector and an encoded polarization angle are collected as a physical branch domain Input2 input to the network; the feature extraction module uses the encoder of UNet, and in the encoder stage, the intensity branch domain Input1 and the physical branch domain Input2 are input to a residual network respectively to extract local features thereof; the feature map after down-sampling of the residual network enters the cross-attention mixed domain module to learn, so as to mine the potential correlation between the branch domains and perceive global information; a plurality of up-sampling blocks are cascaded in the decoder to obtain a feature map, the feature map of the encoder is fused with the feature map of the decoder through the fusion module to obtain a fusion feature, and the fusion feature is converted into a normal vector of each pixel by the normal vector output module; The Bi-Fusion module is used as the fusion module and is composed of channel attention and spatial attention; the spatial attention adopts a CBAM block, which is used as a spatial filter to enhance local details and suppress irrelevant areas; the channel attention adopts an SE block, which is used to enhance the extraction ability of local details; the channel attention and the spatial attention both comprise pooling, full connection and activation operations; S3: dividing the polarization three-dimensional imaging data set in step S1 into a training set and a validation set, constructing a fusion loss function, training the network in step S2 by using the training set, guiding the model to converge by using the validation set, obtaining a trained model and saving the trained model; S4: inputting a tissue image collected in real time by an endoscope into the model saved in step S3 to output a tissue surface normal vector and finally reconstruct a three-dimensional topography of a tissue surface.

2. The monocular endoscope-oriented polarimetric 3D imaging method according to claim 1, characterized in that, Step S1 comprises the following sub-steps: S11: building a polarization metrology monocular endoscope imaging system; modifying the endoscope so that the imaging rear end is equipped with a polarization camera, and collecting polarization images with polarization angles of 0°, 45°, 90° and 135° as an intensity branch domain Input1 input to the network; S12: building a polarization metrology monocular endoscope and three-dimensional scanner image registration system, fixing the endoscope and the scanner to ensure that the imaging poses of the two are the same at each shooting; through stereo calibration, the intrinsic parameters and extrinsic parameters between the endoscope and the scanner are obtained, so as to unify the fields of view of the endoscope and the scanner; S13: With the 4 polarized images in sub-step S11, a way to approximately calculate the specular reflection component I S and the diffuse reflection component I D is proposed: ； wherein represents the specular reflection amplitude quantity, represents the constant reflection component; the calculation of the parameters based on the Stokes vector , : ； wherein represents the degree of polarization, represents the analyzer angle; , , , represent the polarized images with polarization angles of 0°, 45°, 90°, 135°, respectively; the analyzer angle is further encoded to obtain: ； wherein denotes the encoded polarization angle; Based on the above , , , 4 physical priors as physical branch domain input2 of the network S14: Capture the three-dimensional point cloud of the tissue by the scanner, and convert the three-dimensional point cloud into the point cloud under the field of view of the endoscope by using the external parameters calibrated in sub-step S12; then calculate the surface normal vector field based on the difference method as the true value Output of the network output; S15: Randomly cut the input image into N images to enhance the data set.

3. The monocular endoscope-oriented polarimetric 3D imaging method according to claim 1, characterized in that, In step S3, the network is trained by constructing a fusion loss function defined as: ； wherein is a cosine consistency loss function, is a gradient continuity loss function; respectively represent , a weight parameter of for regulating the weight of different loss terms.

4. The monocular endoscope-oriented polarimetric 3D imaging method of claim 1, wherein, In step S3, the polarized three-dimensional imaging data set of step S1 is divided into a training set and a validation set; the network with the dual-domain cross-attention fusion in step S2 is preliminarily trained by using the training set, and the model is guided to converge by using the validation set; in the training process, the learning rate is dynamically adjusted, and a gradient strategy of large in front and small in back is adopted to accelerate the convergence speed in the early stage and stabilize the optimization process in the later stage; at the same time, an early stopping mechanism is introduced, when the fusion loss function is less than the set threshold within the set training number of times, the training is terminated in advance to prevent overfitting.

5. A polarization three-dimensional imaging device for monocular endoscope, characterized by, The processor is configured to implement the polarized three-dimensional imaging method for monocular endoscope according to any one of claims 1-4.

6. An electronic device comprising a memory and a processor, characterized in that The memory is coupled to the processor, and is configured to store program data, and the processor is configured to execute the program data to implement the polarized three-dimensional imaging method for monocular endoscope according to any one of claims 1-4.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the polarized three-dimensional imaging method for monocular endoscope according to any one of claims 1-4.

Images