A method and device for denoising a laser speckle image

Abstract

The application relates to a laser speckle image denoising method and device, which combines the complementary advantages of a laser intensity image and a laser phase image, extracts and fuses a contour feature map of a laser speckle image and a curvature feature map and a normal vector feature map of a laser phase image, provides a neural network with abundant geometric context, and further adds geometric feature constraints in the neural network to form a hierarchical system. The synergistic effect of the geometric constraints enables the denoising model to accurately maintain the edge contour, local surface morphology, surface orientation and overall topological structure of the image when removing speckle noise, significantly improves the geometric structure integrity of the denoised image, and effectively solves the problem that denoising and structure preservation are difficult to balance and diagnostic information is lost in the background technology. The application provides a more reliable and accurate solution for medical imaging application scenarios, and has significant technical progress and application value.

Description

Technical Field

[0001] This application relates to the field of laser speckle image denoising technology, and more particularly to a method and apparatus for denoising laser speckle images. Background Technology

[0002] In the field of medical imaging, laser imaging technology, due to its advantages such as high resolution, fast imaging speed, and non-invasiveness, is widely used in scenarios such as pathological slide observation, skin lesion diagnosis, and three-dimensional imaging of internal organs, becoming one of the core technologies for clinical diagnosis and medical research. However, the high coherence of lasers can induce random interference phenomena, forming laser speckle noise. This noise manifests as randomly distributed bright and dark spots in the image, which can severely interfere with the grayscale uniformity and detail clarity of medical images. In medical settings, speckle noise may obscure the edge contours of diseased tissues, tiny lesions, and surface structural features, making it difficult for doctors to accurately identify the extent and morphology of lesions, and even leading to misdiagnosis or missed diagnosis. Therefore, denoising of laser speckle images is a crucial prerequisite for the practical application of medical laser imaging technology, and its denoising effect directly determines the diagnostic value of medical images.

[0003] Currently, laser speckle denoising technology mainly utilizes digital image processing methods, including traditional filtering methods and neural network denoising methods. Traditional filtering methods, such as mean filtering, wavelet filtering, and Lee filtering, do not require additional hardware and have low costs, but they are prone to damaging high-frequency details and geometric structures of the image. Neural network denoising methods are based on models such as U-Net, which achieve noise differentiation through data training. However, existing models mostly rely on pixel grayscale similarity and do not fully combine the geometric structural characteristics of medical images, making it difficult to balance denoising and structure preservation.

[0004] Current digital image processing methods still have significant shortcomings, particularly: existing neural network denoising methods lack effective geometric constraint mechanisms, fail to fully utilize the structural advantages of laser phase images, struggle to balance denoising effects with geometric integrity, and often employ single-modal inputs, making it impossible to consider both the details of intensity images and the structural advantages of phase images. This can easily lead to incomplete or excessive denoising, resulting in the loss of diagnostic information. Summary of the Invention

[0005] A first aspect of this application provides a method for denoising laser speckle images, the method comprising: Obtain laser speckle samples and their corresponding laser phase samples; The contour feature map of the laser speckle sample, the curvature feature map and the normal vector feature map of the laser phase sample are extracted, and a dual-modal geometric fusion feature map is determined based on the contour feature map, the curvature feature map and the normal vector feature map. A denoising model is trained based on the dual-modal geometric fusion feature map, the laser speckle samples, and the laser phase samples; wherein, the denoising model is used to output the denoised image corresponding to the laser speckle samples, and the loss function of the denoising model includes: ; ; ; ; ; in, Let be the loss function of the denoising model. For mean square error loss, For contour consistency constraints, For hypercurvature consistency constraints, For the normal vector direction constraint, For geometric topological consistency constraints, These are the weight coefficients for the corresponding constraints; For dynamic thresholds, The correlation coefficient, It is a natural exponential function. Based on the threshold, The angle between the normal vectors, Mean square error, The Gaussian curvature fitted to the denoised image, The Gaussian curvature extracted from the laser phase sample. The average curvature fitted to the denoised image. The average curvature extracted from the laser phase sample. The Gaussian curvature gradient of the denoised image. To pass Corrected Gaussian curvature gradient, The Gaussian curvature gradient of the laser phase sample; In response to a denoising command, the system acquires a target laser speckle image and a corresponding target laser phase image, and denoises the target laser speckle image based on the target laser phase image and the trained denoising model.

[0006] The laser speckle image denoising method provided in this application has at least the following beneficial effects: This method extracts the contour feature map of the laser speckle image and the curvature feature map and normal vector feature map of the laser phase image, and fuses them into a dual-modal geometric fusion feature map, which provides a rich geometric context for the denoising model. This feature fusion strategy enables the model to understand the geometric structure of the image more comprehensively during the training process, rather than relying solely on pixel-level grayscale information. This method introduces a loss function containing multiple geometric constraints during the training process of the denoising model. This loss function includes not only the traditional mean squared error loss. It also integrates contour consistency constraints. Hypercurvature Consistency Constraint Normal vector direction constraints and geometric topological consistency constraints At the same time, a hierarchical system is formed, that is, geometric topological consistency constraints are first used. Locking in the overall structure of the target region prevents the network model's output from deviating completely from the target shape during the initial training phase, providing a basic framework for subsequent constraints; then, the underlying curvature... Normal vector constraint pass The system uses a linkage mechanism, employing the accuracy of the normal vector orientation to correct curvature calculation deviations, ensuring that local surface details are consistent with the original laser phase image and compensating for local distortions; then, contour consistency constraints are applied. By dynamic threshold With topological constraints Related, The dynamic thresholding mechanism makes contour preservation more flexible and robust. The synergistic effect of these geometric constraints enables the denoising model to accurately preserve the edge contours, local surface morphology, surface orientation, and overall topology of the image while removing speckle noise, significantly improving the geometric integrity of the denoised image. This effectively solves the problems of balancing denoising and structure preservation, as well as the loss of diagnostic information, in background techniques.

[0007] A second aspect of this application provides a noise reduction apparatus for laser speckle images, the apparatus comprising: The training sample acquisition module is used to acquire laser speckle samples and their corresponding laser phase samples; The geometric feature extraction module is used to extract the contour feature map of the laser speckle sample, as well as the curvature feature map and normal vector feature map of the laser phase sample, and to determine the dual-modal geometric fusion feature map based on the contour feature map, the curvature feature map and the normal vector feature map; The model training module is used to train a denoising model based on the dual-modal geometric fusion feature map, the laser speckle samples, and the laser phase samples; wherein, the denoising model is used to output the denoised image corresponding to the laser speckle samples, and the loss function of the denoising model includes: ; ; ; ; ; in, Let be the loss function of the denoising model. For mean square error loss, For contour consistency constraints, For hypercurvature consistency constraints, For the normal vector direction constraint, For geometric topological consistency constraints, These are the weight coefficients for the corresponding constraints; For dynamic thresholds, The correlation coefficient, It is a natural exponential function. Based on the threshold, The angle between the normal vectors, Mean square error, The Gaussian curvature fitted to the denoised image, The Gaussian curvature extracted from the laser phase sample. The average curvature fitted to the denoised image. The average curvature extracted from the laser phase sample. The Gaussian curvature gradient of the denoised image. To pass Corrected Gaussian curvature gradient, The Gaussian curvature gradient of the laser phase sample; The target image denoising module is used to respond to denoising commands, acquire a target laser speckle image and a corresponding target laser phase image, and denoise the target laser speckle image based on the target laser phase image and the trained denoising model.

[0008] A third aspect of this application provides an electronic device including at least one controller and a memory for communicatively connecting to the controller; the memory stores instructions executable by the at least one controller, the instructions being executed by the at least one controller to cause the at least one controller to perform a laser speckle image denoising method as described in the first aspect of this application.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform a laser speckle image denoising method as described in the first aspect of this application.

[0010] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of a laser speckle image denoising method provided in an embodiment of this application; Figure 2 This is a schematic diagram of a laser speckle image denoising method provided in another embodiment of this application; Figure 3 This is a schematic diagram of a laser speckle image denoising device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0014] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0015] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or function in a specific orientation, and therefore should not be construed as a limitation of this application.

[0016] In medical laser imaging, the random interference phenomenon of laser speckle noise disrupts the uniformity of image grayscale and spatial continuity, thus interfering with the edge contours of lesions and the features of small lesions.

[0017] For example, in the context of laser confocal microscopy diagnosis of skin lesions, when acquiring skin tissue images, speckle noise makes the boundaries of melanoma appear discontinuous and blurred, making it impossible for doctors to accurately identify the extent and depth of the lesion, thus affecting the judgment of the nature of the lesion.

[0018] If the above problems are not addressed, distortion of the geometric structure of medical images will reduce the reliability of clinical diagnosis, potentially leading to incorrect treatment decisions, delaying disease intervention, and adversely affecting patient safety.

[0019] like Figure 1 One embodiment of this application provides a method for denoising laser speckle images, the method comprising: Step S110: Obtain the laser speckle sample and its corresponding laser phase sample; Step S120: Extract the contour feature map of the laser speckle sample, and extract the curvature feature map and normal vector feature map of the laser phase sample. Based on the contour feature map, curvature feature map and normal vector feature map, determine the dual-modal geometric fusion feature map. Step S130: Based on the dual-modal geometric fusion feature map, laser speckle samples, and laser phase samples, a denoising model is trained; wherein, the denoising model is used to output the denoised image corresponding to the laser speckle samples, and the loss function of the denoising model includes: ; ; ; ; ; in, The loss function of the denoising model. For mean square error loss, For contour consistency constraints, For hypercurvature consistency constraints, For the normal vector direction constraint, For geometric topological consistency constraints, These are the weight coefficients for the corresponding constraints; For dynamic thresholds, The correlation coefficient, It is a natural exponential function. Based on the threshold, The angle between the normal vectors, Mean square error, The Gaussian curvature fitted to the denoised image, Gaussian curvature extracted from laser phase samples. The average curvature fitted to the denoised image. The average curvature extracted from the laser phase sample. The Gaussian curvature gradient of the denoised image, To pass Corrected Gaussian curvature gradient, The Gaussian curvature gradient of the laser phase sample; Step S140: In response to the denoising command, acquire the target laser speckle image and the corresponding target laser phase image, and denoise the target laser speckle image based on the target laser phase image and the trained denoising model.

[0020] Because laser speckle images are essentially the imaging result of light intensity signals after laser light is reflected from a target, they primarily characterize the grayscale distribution and reflectivity of the target surface, reflecting the two-dimensional contrast of light and dark (such as the grayscale difference between diseased and normal tissues), and can intuitively present the two-dimensional outline of the target. Laser phase images, on the other hand, are essentially the imaging result of phase changes during laser propagation, primarily characterizing the three-dimensional geometric structure of the target. Their pixel values ​​are directly related to the height and orientation of the target surface, accurately reflecting the three-dimensional morphology of the target. The two types of images are highly complementary. Therefore, the starting point of this embodiment is to combine the complementary advantages of laser intensity images and laser phase images, learn the geometric features in both types of images, and add geometric feature constraints to the neural network to achieve efficient removal of laser speckle while preserving image details and geometric structural integrity.

[0021] In S110 of this embodiment, both types of samples are formed based on the area of ​​the object (such as an individual lesion) acquired by the detector, and both are used in the training process of the model (neural network). It should be noted that since the laser speckle image is formed based on the acquisition of the object area by a conventional laser imaging detector, the phase image of the object area can be obtained at the same time using phase imaging technology.

[0022] In S120 of this embodiment, the contour feature map of the laser speckle sample can be obtained by the Canny edge detection algorithm. Then, the curvature feature map of the laser phase image reflecting the curvature of the target surface is calculated by the second-order partial derivative. The normal vector feature map of the laser phase image reflecting the orientation of the target surface is calculated by the gradient.

[0023] In this embodiment, before step S120, the laser speckle sample and its corresponding laser phase sample can be preprocessed, for example: (1) Normalization: Map the pixel values ​​of the two types of images to the [0,1] range to eliminate the difference in grayscale; (2) Preliminary noise reduction: Gaussian filtering is used to remove non-speckle random noise; (3) Size alignment: Image translation and rotation deviations are corrected by image registration algorithms such as SIFT feature matching.

[0024] In S130 of this embodiment, taking the denoising model U-net as an example, it includes an encoder and a decoder as well as a jump connection located between the encoder and the decoder.

[0025] First, the dual-modal geometric fusion feature map, laser speckle samples, and laser phase samples are input into the U-net model for tensor fusion to obtain the encoder's input features.

[0026] In this embodiment, the U-net network loss function includes: ; ; in, These are the weight coefficients for the corresponding constraints.

[0027] The following describes several constraints set in this embodiment: The first constraint is This is the mean squared error loss, which is used to ensure basic denoising performance. It's worth noting that mean squared error loss is common knowledge in the field and will not be elaborated upon here.

[0028] The second constraint is The constrained denoised image (i.e., the final output of the network model) is consistent with the topological structure (such as connected regions, holes, etc.) of the original laser speckle sample. In some embodiments, it can be: ; in, For the denoised image, the first Class topological feature parameters, The first laser speckle sample Class topological feature parameters, This represents the total number of categories of topological features. This is the function for finding the maximum value.

[0029] Under normal circumstances, The value can be 4, for example, : Represents the number of connected regions; : Represents the number of holes; : Represents the number of inflection points in the contour; : Represents the average distance between adjacent inflection points. The denominator is set to... This can prevent the denominator from being 0.

[0030] The third constraint is The purpose is to ensure that the contour obtained in the denoised image output by the model is consistent with the original laser speckle sample.

[0031] In some embodiments, it can be: ; in, For contour consistency constraints, This is the set of Canny edge detection contour pixels for the denoised image output by the model. It is the Canny edge detection contour pixel set of the original laser speckle sample.

[0032] Importantly, this embodiment sets a dynamic threshold within this constraint. Subject to dynamic threshold The impact, namely .

[0033] in, The purpose of this is: due to the dynamic threshold Loss value from topology constraints Regulation, when The smaller, The lower the level of topological distortion, the stricter the contour constraints; conversely, the higher the level of topological distortion, the more relaxed the constraints. This allows for strict contour constraints when topological distortion is low, and for prioritizing overall preservation when topological distortion is high. Under normal circumstances... It can be set to 0.15. The value can be set from 0.8 to 1.2.

[0034] The fourth constraint is The normal vector of the denoised image is constrained to have the same orientation as the normal vector of the original image.

[0035] In some embodiments, it can be: ; in, For the corresponding pixels of the denoised image Surface normal vector, This is the surface normal vector of the corresponding pixel in the original laser speckle image.

[0036] It is worth noting that, during the process of solving using the normal vector, it can be determined that the ratio of the dot product to the magnitude is the angle between them. cosine value Subsequent embodiments of this application utilize the cosine value corresponding to this included angle. This enables the control of the hyperbola consistency constraint and the control of the calculation of the bimodal geometric fusion feature map, as detailed in subsequent embodiments.

[0037] The fifth constraint is Constrain the curvature value and gradient change trend.

[0038] In some embodiments, it can be: ; ; in, The angle between the normal vectors, The Gaussian curvature fitted to the denoised image, Gaussian curvature extracted from laser phase samples. The average curvature fitted to the denoised image. The average curvature is extracted from the laser phase sample; these parameters can be calculated using second-order partial derivatives. The Gaussian curvature gradient of the denoised image, To pass Corrected Gaussian curvature gradient, The Gaussian curvature gradient of the laser phase sample.

[0039] In this embodiment, hypercurvature consistency constraint Normal vector direction constraint Between To regulate, that is, through Correcting the curvature gradient ensures that local surface details remain consistent with the original laser phase image, compensating for local distortions.

[0040] Unlike conventional loss function construction, this embodiment constructs a hierarchical system under several consistency constraints. This is because a single constraint can lead to inconsistencies, with different constraints interfering with each other and reducing the accuracy and reliability of the constraint effect. Therefore, this embodiment proposes a clear hierarchical logic, where the topological structure (top layer) determines the macroscopic shape of the target region, the contour structure (middle layer) defines the two-dimensional boundary of the target, and curvature and normal vectors (bottom layer) characterize the local surface details of the target. The relationship between the four types of constraints strictly follows this inherent logic, ensuring that the constraint process conforms to the geometric characteristics of the target and avoiding constraint disconnect or conflict. At the same time, leveraging the advantages of laser dual-modal imaging, topological and contour constraints rely on the boundary accuracy of the intensity image, while curvature and normal vector constraints rely on the structural stability of the phase image. The relationship further strengthens the fusion of dual-modal advantages, making the constraints more reliable.

[0041] 1) Top-level topology constraints ( First, lock the overall structure of the target region (such as connected regions, holes, etc.) to prevent the network model output from completely deviating from the target shape in the early stage of training, and to provide a basic framework for subsequent constraints. 2) Mid-level contour constraints ( ), through dynamic threshold With topological constraints With linkage, when topological distortion is small, contour constraints are more stringent (i.e.) Smaller (Reduce), accurately lock the boundary, when the topological distortion is large, relax the constraints to prioritize the preservation of the whole (i.e. Larger (Raise), to avoid further distortion of the overall structure due to excessive constraint of the contour; 3) Curvature Normal vector constraint pass Linking the two is crucial because the normal vector reflects the orientation of the target area's surface, while curvature reflects the degree of bending of the target area's surface; the two are inherently related: for targets with the same orientation, the curvature calculation must be based on a unified directional reference. If the orientation of the normal vector deviates significantly, the curvature calculation will suffer severe deviations, leading to distortion of local details. Therefore, the cosine value of the angle between the normal vectors is used. Correcting curvature gradient ,when When the curvature is small, the deviation in the orientation of the normal vector is large, which weakens the curvature gradient, reduces the curvature constraint weight, and avoids misleading the network with incorrect curvature information; when the orientation of the normal vector is consistent, the original curvature gradient is retained, the curvature constraint is strengthened, and the accuracy of the local curvature is ensured.

[0042] This hierarchical association in the embodiment allows constraints to cover the entire structure from the whole to the parts without losing the overall structure or ignoring details. It avoids structural distortion caused by a single constraint, improves the model training effect, and ultimately improves the denoising effect when the model is used.

[0043] In step S140, after model training, the target laser speckle image can be denoised based on the target laser phase image and the trained denoising model. The specific process is as follows: Step S1411: Extract the target contour feature map from the target laser speckle image, and extract the target curvature feature map and target normal vector feature map from the target laser phase image. Based on the target contour feature map, target curvature feature map and target normal vector feature map, determine the target dual-modal geometric fusion feature map. Step S1412: Input the target dual-modal geometric fusion feature map, the target laser phase image and the target laser speckle image into the trained denoising model to obtain the denoising result of the output target laser speckle image.

[0044] In this embodiment, as described in the background section, traditional digital image processing methods (such as mean filtering and wavelet filtering), while low-cost, inherently suffer from the tendency to damage high-frequency details and geometric structures of images, which is unacceptable in medical diagnosis. Existing neural network denoising methods, although achieving noise differentiation to some extent, primarily rely on pixel grayscale similarity and lack effective geometric constraint mechanisms. They fail to fully utilize the structural advantages of laser phase images, resulting in a difficulty in balancing denoising and structure preservation. This often manifests as incomplete or excessive denoising, leading to the loss of diagnostic information.

[0045] This embodiment leverages the complementary advantages of laser intensity images and laser phase images, effectively overcoming the limitations of existing technologies by introducing multimodal input and a comprehensive geometric constraint mechanism. Specifically, by acquiring laser speckle samples and corresponding laser phase samples for model training, dual-modal data utilization is achieved. This contrasts with existing technologies that often employ single-modal input. Furthermore, the laser phase image provides stable geometric structure information, compensating for the shortcomings of laser speckle images in structure preservation.

[0046] Furthermore, this method extracts the contour feature map of the laser speckle image and the curvature feature map and normal vector feature map of the laser phase image, and fuses them into a dual-modal geometric fusion feature map, providing a rich geometric context for the denoising model. This feature fusion strategy enables the model to understand the geometric structure of the image more comprehensively during training, rather than relying solely on pixel-level grayscale information.

[0047] Of particular importance, the denoising model training process in this embodiment introduces a loss function that includes multiple geometric constraints. This loss function includes not only the traditional mean squared error loss. It also integrates contour consistency constraints. Hypercurvature Consistency Constraint Normal vector direction constraints and geometric topological consistency constraints At the same time, a hierarchical system is formed, that is, geometric topological consistency constraints are first used. Locking in the overall structure of the target region prevents the network model's output from deviating completely from the target shape during the initial training phase, providing a basic framework for subsequent constraints; then, the underlying curvature... Normal vector constraint pass( The system uses a linkage mechanism, employing the accuracy of the normal vector orientation to correct curvature calculation deviations, ensuring that local surface details are consistent with the original laser phase image and compensating for local distortions; then, contour consistency constraints are applied. By dynamic threshold ( ) and topological constraints Related, The dynamic threshold mechanism makes the contour more flexible and robust.

[0048] The synergistic effect of these geometric constraints enables the denoising model to accurately preserve the edge contours, local surface morphology, surface orientation, and overall topology of the image while removing speckle noise. This contrasts sharply with existing neural network denoising methods that lack effective geometric constraints, significantly improving the geometric structural integrity of the denoised image and effectively solving the problems of balancing denoising and structure preservation, as well as the loss of diagnostic information, in background techniques.

[0049] In some embodiments of this application, the denoising model includes an encoder, a feature fusionist, and a decoder; The encoder consists of four cascaded convolutional blocks, each containing two 3×3 convolutional layers, one batch normalization layer, and one max pooling layer. The feature fusion unit is connected to the last convolutional block in the cascade. The feature fusion unit is used to fuse the output features of the encoder with the bimodal geometric fusion feature map to obtain intermediate fused features. The decoder consists of four cascaded deconvolutional blocks, each containing one 3×3 deconvolutional layer, two 3×3 convolutional layers, and one batch normalization layer; the decoder is connected to the feature fusion unit. The process of calculating intermediate fusion features includes: ; ; in, This is a dual-modal geometric fusion feature map. For the output characteristics of the encoder, For attention mechanism functions, It is the sigmoid activation function. This is the weight matrix. For bias terms, Characterizing the goodness of fit to geometric constraints, It is the weighting coefficient.

[0050] In this embodiment, the denoising model achieves effective denoising of laser speckle images through the collaborative work of the encoder, feature fusion unit, and decoder. First, the encoder performs multi-level feature extraction and downsampling on the input laser speckle sample, gradually capturing the semantic information of the image. During this process, the convolutional layer, batch normalization layer, and max pooling layer in each convolutional block work together to ensure effective feature extraction and representation. Subsequently, the feature fusion unit intelligently fuses the deep features extracted by the encoder with a pre-determined bimodal geometric fusion feature map. This fusion is not a simple superposition, but rather an attention mechanism that dynamically adjusts the weights of the encoder features and geometric features based on the geometric constraint fit, thereby generating intermediate fused features containing rich semantic and geometric information. This dynamic weighted fusion mechanism allows the model to more flexibly utilize the geometric structure information provided by the laser phase image, compensating for the shortcomings of single image features. Finally, the decoder receives the intermediate fused features and gradually upsamples and refines the features through cascaded deconvolutional blocks, ultimately reconstructing a high-quality denoised image. The deconvolutional layer, convolutional layer, and batch normalization layer in the decoder work together to ensure effective restoration of image details and suppression of artifacts.

[0051] Meanwhile, this embodiment uses attention to extract deep features from the encoder's output. and Secondary fusion can further enhance the fusion effect. And crucially, Geometric constraint fit was added. When geometric constraint fit The higher the value, the more the attention is biased towards regions with complete geometric structures. This can avoid the attention being misled by speckle noise, prioritize the focus on effective regions with complete geometric structures, enhance target feature extraction, suppress noise features, improve the accuracy of network denoising, and reduce artifact generation.

[0052] In some embodiments of this application, the calculation process of the dual-modal geometric fusion feature map includes: ; in, For contour feature map, This is a curvature feature map. For the normal vector feature map, These are the weight coefficients for the corresponding feature maps.

[0053] Modal geometry fusion feature maps are key representations used to fuse geometric information from laser speckle samples and laser phase samples. They aim to integrate geometric features from different modes into a unified representation to provide comprehensive geometric context information for denoising models. Their role is to guide the denoising model to remove speckle noise while preserving image structure and details, thereby improving the quality of the denoised image. A modal geometry fusion feature map can be a multi-channel tensor, where each channel encodes a specific geometric property or a combination thereof.

[0054] Contour feature maps represent the boundary or edge information of objects in an image. In laser speckle image denoising, contours are important structural features that need to be accurately preserved. The extraction of these feature maps can be achieved through various edge detection algorithms, such as the Canny operator, the Sobel operator, and the Laplacian operator. Their purpose is to capture regions in the image where there are abrupt changes in brightness or texture, which typically correspond to the shape boundaries of objects.

[0055] Curvature feature maps describe the degree of curvature of local shapes on an image surface. Their function is to provide a quantitative description of whether a surface is concave, convex, flat, or saddle-shaped, helping denoising models distinguish regions with different geometric shapes. Feature maps can be obtained by calculating the local curvature (such as Gaussian curvature or mean curvature) at image pixels. For example, curvature can be calculated using the second derivative information of the image (such as the Hessian matrix).

[0056] The normal vector feature map represents the local orientation of an image surface at each pixel. In laser phase images, the normal vector provides crucial information about the surface orientation, which is essential for reconstructing 3D shapes and understanding surface geometry. This feature map can be obtained by calculating the gradient information of the image or using 3D reconstruction techniques; for example, the surface normal vector can be estimated based on the gradient of the phase image. Its purpose is to provide vector information about the local orientation of the surface, helping denoising models maintain directional consistency when dealing with regions with complex geometries.

[0057] By adjusting these weight details, the importance of different geometric information can be optimized according to specific application scenarios or data characteristics. For example, fixed values ​​can be set based on experience, or adaptive learning can be performed through the training process, enabling the model to better balance the influence of different features. These coefficients ensure the flexibility and adaptability of the fusion process.

[0058] This embodiment introduces the cosine value of the angle between the normal vectors, which can adjust the curvature features according to the relative directional relationship between the normal vectors, so that they are better consistent with the overall geometric structure. This helps to reduce potential errors caused by inconsistencies in direction when fusing different geometric features, thereby improving the accuracy of the fused feature map.

[0059] In some embodiments of this application, The possible values ​​include: ; in, The weighted sum coefficient, Integers ranging from 1 to 4 It is a natural exponential function.

[0060] This embodiment incorporates dynamic weight adjustment, meaning that the constraint weights are adaptively adjusted during the training process using a dynamic weight formula. In the early stages of training: the network has a weak ability to distinguish between noise and target features. If topological disorder occurs, the weight of topological constraint will be increased automatically to correct the overall structure first; if contour offset occurs, the weight of contour constraint will be increased to accurately lock the boundary. Later in the training process: network output distortion decreases, the loss values ​​of each constraint tend to be balanced, the weights automatically tend to be balanced, and image artifacts caused by excessive constraints are avoided; Specifically, the loss value reflects the current model in the [missing information - likely a specific timeframe or period]. The degree of fit or error on a geometric feature, when the loss value of a certain constraint term... A larger value indicates that the model performs poorly in that aspect. The value will be smaller; conversely, when When the value is smaller (indicating that the model performs well in this aspect). The value of is relatively large. This design ensures that when the model performs poorly on a certain constraint, its corresponding weight is relatively reduced, thus preventing that constraint from over-dominating training; conversely, when a constraint performs well, its weight is relatively increased, encouraging the model to continue optimizing that aspect. (The denominator is not directly related to the design.) It applies to all geometric constraint terms. The values ​​are summed to normalize the weights, ensuring that all weight coefficients are equal. The sum of these factors remains relatively stable to a certain extent, allowing the proportion of each weighting coefficient in the total weight to be dynamically adjusted. Used to control the total weight of all geometric constraint terms.

[0061] In this embodiment, during model training, the weight coefficient for each geometric constraint is no longer a fixed value, but is dynamically calculated based on the current loss value of that constraint. When the loss value of a geometric constraint is large, it indicates that the model's fit in that aspect is poor, and the weight of that constraint will be relatively reduced through the calculation of the exponential function. Conversely, when the loss value of a geometric constraint is small, it indicates that the model's fit in that aspect is good, and its corresponding weight will be relatively increased. This design allows the model to adaptively adjust its focus on different geometric features during training, avoiding the over-dominance of poorly performing constraints and preventing well-performing constraints from being ignored. By normalizing all exponential terms and multiplying them by the total weight coefficient, the proportion of the weights of all geometric constraints in the total loss function is ensured to be dynamic and reasonable. This dynamic weight adjustment mechanism allows the model to more effectively balance the optimization objectives of each geometric constraint, thereby better learning and preserving the geometric structural information of the laser speckle image during training, improving the training efficiency of the denoising model and the final denoising quality.

[0062] In some embodiments of this application, a skip connection is provided between each convolutional block of the encoder and the corresponding deconvolutional block of the decoder; the skip connection includes: Calculate the fusion confidence weights based on the contour consistency constraint and the hypercurvature consistency constraint: ; ; ; in, To incorporate confidence weights; The weighted fusion feature map is obtained by weighting the fusion confidence weights and the bimodal geometric fusion feature map: ; in, The output features of the skip connections in the convolutional block, For weighted fusion feature maps; The weighted fused feature map is passed to the corresponding deconvolution block.

[0063] Skip connections can effectively alleviate the vanishing gradient problem, allowing the network to be trained to greater depths. Skip connections can be implemented in various ways; for example, the output feature map of a certain layer of the encoder can be concatenated with the input feature map of the corresponding layer of the decoder along the channel dimension, or they can be added element-wise.

[0064] In this embodiment, weight It is calculated based on contour consistency constraints and hyperbola consistency constraints. By using the inverses of these constraint terms or their related functions as weights, the model can adaptively adjust its dependence on different feature information during training according to the degree to which the current denoising effect satisfies the geometric constraints. For example, when the geometric constraints are highly satisfied, the weights can tend to retain more original image features; when the satisfaction is low, the weights of geometrically guided features can be increased.

[0065] The overall working principle of this embodiment is as follows: In the encoder-decoder architecture for laser speckle image denoising, in order to overcome the problem of spatial detail loss that may occur during encoder downsampling and to more effectively utilize geometric information to guide the denoising process, a skip connection with a dynamic weighted fusion mechanism is introduced. Specifically, the output features of each convolutional block of the encoder... Instead of simply passing the data directly to the corresponding deconvolution block in the decoder, it first participates in an intelligent weighted fusion process. In this fusion process, it first considers the contour consistency constraints of the current denoising model output. and hypercurvature consistency constraints Dynamically calculate the fusion confidence weights Contour consistency constraints and hypercurvature consistency constraints As part of the loss function, it reflects the degree to which the denoised image matches the real image geometrically. When the values ​​of these constraint terms are small (i.e., the denoising effect is geometrically better), the calculated... The higher the value, the better the model's understanding of the output features extracted by the current encoder. (Containing rich spatial details) has a high confidence level. Conversely, when and When the value is large (i.e., the geometric matching degree is poor), The lower the level, the more the model will rely on the dual-modal geometric fusion feature map. .

[0066] Subsequently, this dynamically calculated Used for and Perform weighted fusion to generate a weighted fusion feature map. This means that if High, (Raw spatial features from the encoder) in The dominant position is achieved in the middle, thus ensuring that the decoder can receive more fine spatial information from the original image, if Lower, then The increased contribution of explicit geometrically guided features from the laser phase image provides a stronger geometric correction signal for the decoder even with poor geometric matching; ultimately, this intelligently weighted fusion... It is passed to the corresponding deconvolution block of the decoder.

[0067] In this way, this embodiment achieves adaptive and dynamic integration of encoder multi-scale spatial features and dual-modal geometric fusion feature maps. It not only retains the advantage of skip connections in transmitting fine spatial information, but more importantly, it introduces an intelligent decision-making mechanism based on geometric constraint feedback. This allows the model to flexibly adjust its dependence on different types of features according to the real-time effect of denoising. This enables the decoder to accurately restore image details when reconstructing denoised images, while ensuring the integrity and consistency of the geometric structure. As a result, it significantly improves the denoising quality of laser speckle images, especially in complex geometric structures and texture regions.

[0068] like Figure 2 For ease of explanation, this embodiment provides the following steps: Step S910, construct the denoising model; the model includes: The encoder consists of four cascaded convolutional blocks, each containing two 3×3 convolutional layers, one batch normalization layer, and one max pooling layer. The feature fusion unit is connected to the last convolutional block in the cascade; The decoder consists of four cascaded deconvolutional blocks, each containing one 3×3 deconvolutional layer, two 3×3 convolutional layers, and one batch normalization layer; the decoder is connected to the feature fusion unit. Step S920, data processing; Determine laser intensity image and laser phase image ; Laser intensity image extracted using the Canny operator Contour feature map ; Laser phase image extracted using second-order partial derivatives curvature feature map Through gradient calculation and extraction of laser phase image Normal vector feature map .

[0069] Fusion laser intensity image Contour feature map Laser phase image and normal vector feature map ,get: ; in It is a dual-modal geometric fusion feature map; Obtained through training. It is obtained by calculating the normal vector.

[0070] Step S930, Model Training; First, the laser intensity image Laser phase image and dual-modal geometric fusion feature maps Channel splicing was performed to obtain the tensor. ; For the image height and width.

[0071] Then, the tensor Input to encoder ,get ,in: ; ; in, for and The fusion characteristics. It is the sigmoid activation function. This is the weight matrix. Bias term; geometric constraint fit The higher the value, the more attention is drawn to areas with intact geometric structures.

[0072] Then, The input is fed into the decoder to obtain the final denoised image output by the decoder. .

[0073] In this configuration, each convolutional block of the encoder is connected to a corresponding deconvolutional block of the decoder via skip connections; these skip connections include: Calculate the fusion confidence weights: ; ; ; The weighted fusion feature map is obtained: ; The weighted fused feature map is passed to the corresponding deconvolution block. The output features are the skip connections of the convolutional block.

[0074] The following are four constraints: 1) Contour consistency constraint: ; ; ; 2) Hypercurvature consistency constraint: ; ; 3) Normal vector constraint: ; During the calculation of the normal vector, the ratio of the dot product to the magnitude is obtained as the cosine of the angle between the normal vectors. .

[0075] 4. Geometric topological consistency constraints: ; The loss function is described below: ; ; The actual label pixel value, Weight Calculation method: ; in, This is the weighted sum coefficient.

[0076] Step S940, Model Usage: Acquire and preprocess laser speckle images and laser phase images; The contour features of the laser speckle image and the curvature feature map and normal vector feature map of the laser phase image are extracted and fused to obtain a dual-modal geometric fusion feature map; The preprocessed bimodal image is fused with bimodal geometric features. Figure 1 The same input is used to train the network, and the denoised image of the original image is obtained.

[0077] The following is a set of experimental examples: Experimental environment: Four NVIDIA RTX3060 graphics cards with 12GB of video memory, an Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz, 128GB of memory, Ubuntu 20.04.1 operating system, CUDA version 11.3, Python 3.8 programming language, and PyTorch version 1.12.1 deep learning framework. All results were obtained from this experimental environment.

[0078] This experiment used two datasets, both suitable for medical laser imaging scenarios, as follows: (1) Medical laser speckle-phase dual-modal dataset (collected from Guangzhou Respiratory Center): The training set contains 8,000 laser speckle images, corresponding phase images and real label images, and the test set contains 2,000 images of the same type. The images are all from clinical pathological slides and laser imaging samples of skin lesions, covering different speckle intensities (weak, medium and strong) and different lesion types, which are in line with actual application scenarios.

[0079] (2) LSID publishes laser speckle dataset; Experimental Objective: To verify the denoising effect and geometric structure preservation capability of the proposed (bimodal + geometrically constrained neural network), to verify the effectiveness of each core module through ablation experiments, and to compare with existing mainstream denoising algorithms. The key evaluation indicators are: denoising effect (PSNR peak signal-to-noise ratio, SSIM structural similarity), geometric structure preservation capability (contour consistency, curvature deviation), and model efficiency (number of parameters, inference speed).

[0080] Ablation experiments (verifying the effectiveness of each module): Based on the improved U-Net model, dual-modal input and geometric constraint modules were added respectively to verify the impact of each module on the denoising effect and geometric structure preservation. The experimental results are shown in Table 1.

[0081] Table 1

[0082] Comparative experiments (comparison with mainstream algorithms): Traditional denoising algorithms (Lee filtering, wavelet filtering) and existing neural network denoising algorithms (basic U-Net, U-Net++) were selected and tested under the same experimental environment and the same dataset to verify the superiority of the proposed solution. The experimental results are shown in Table 2.

[0083] Table 2

[0084] Analysis of experimental results: (1) Ablation experiments show that after adding geometric constraint modules to the dual-modal input, all indicators are improved, proving that geometric constraints and the correlation between modules can effectively avoid geometric structure distortion; (2) Comparative experiments show that the PSNR and SSIM of this application are superior to traditional filtering algorithms and existing neural network algorithms. According to the experimental data, it can be proved that this application can take into account the denoising effect, geometric structure preservation and model efficiency, and adapt to the needs of real-time medical imaging. (3) Experiments have verified the effectiveness of the core design of this application. The synergistic effect of dual-modal fusion and geometric constraints solves the pain point of existing algorithms in balancing denoising and structure preservation. It can accurately preserve the contour, shape and surface details of lesions in medical images and meet the needs of clinical diagnosis.

[0085] like Figure 3 One embodiment of this application provides a noise reduction device for laser speckle images, the device comprising: The training sample acquisition module 1100 is used to acquire laser speckle samples and their corresponding laser phase samples; The geometric feature extraction module 1200 is used to extract the contour feature map of the laser speckle sample, as well as the curvature feature map and normal vector feature map of the laser phase sample, and to determine the dual-modal geometric fusion feature map based on the contour feature map, curvature feature map and normal vector feature map. The model training module 1300 is used to train a denoising model based on dual-modal geometric fusion feature maps, laser speckle samples, and laser phase samples. The denoising model outputs the denoised image corresponding to the laser speckle sample, and its loss function includes: ; ; ; ; ; in, The loss function of the denoising model. For mean square error loss, For contour consistency constraints, For hypercurvature consistency constraints, For the normal vector direction constraint, For geometric topological consistency constraints, These are the weight coefficients for the corresponding constraints; For dynamic thresholds, The correlation coefficient, It is a natural exponential function. Based on the threshold, The angle between the normal vectors, Mean square error, The Gaussian curvature fitted to the denoised image, Gaussian curvature extracted from laser phase samples. The average curvature fitted to the denoised image. The average curvature extracted from the laser phase sample. The Gaussian curvature gradient of the denoised image, To pass Corrected Gaussian curvature gradient, The Gaussian curvature gradient of the laser phase sample; The target image denoising module 1400 is used to respond to denoising commands, acquire the target laser speckle image and the corresponding target laser phase image, and denoise the target laser speckle image based on the target laser phase image and the trained denoising model.

[0086] It should be noted that the laser speckle image denoising device provided in this embodiment and the laser speckle image denoising method described above are based on the same inventive concept. Therefore, the content of the laser speckle image denoising method in the above embodiment is also applicable to the content of the laser speckle image denoising device in this embodiment, and will not be repeated here.

[0087] like Figure 4 One embodiment of this application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for denoising laser speckle images. The electronic device includes: At least one battery; At least one memory; At least one processor; At least one program; The program is stored in memory, and the processor executes at least one program to implement the laser speckle image denoising method described above in this disclosure.

[0088] Electronic devices can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0089] The electronic devices according to embodiments of this application will now be described in detail.

[0090] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure. The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to perform a laser speckle image denoising method according to an embodiment of this disclosure.

[0091] The input / output interface 1800 is used to implement information input and output. The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900); The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0092] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the detection method of the pressurized water reactor containment pressure control system described above.

[0093] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0094] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0095] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0096] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0097] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0098] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0099] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0101] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0102] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0103] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

Claims

1. A method for denoising laser speckle images, characterized in that, The method includes: Obtain laser speckle samples and their corresponding laser phase samples; The contour feature map of the laser speckle sample, the curvature feature map and the normal vector feature map of the laser phase sample are extracted, and a dual-modal geometric fusion feature map is determined based on the contour feature map, the curvature feature map and the normal vector feature map. A denoising model is trained based on the dual-modal geometric fusion feature map, the laser speckle samples, and the laser phase samples; wherein, the denoising model is used to output the denoised image corresponding to the laser speckle samples, and the loss function of the denoising model includes: ； ； ； ； ； in, Let be the loss function of the denoising model. For mean square error loss, For contour consistency constraints, For hypercurvature consistency constraints, For the normal vector direction constraint, For geometric topological consistency constraints, These are the weight coefficients for the corresponding constraints; For dynamic thresholds, The correlation coefficient, It is a natural exponential function. Based on the threshold, The angle between the normal vectors, Mean square error, The Gaussian curvature fitted to the denoised image, The Gaussian curvature extracted from the laser phase sample. The average curvature fitted to the denoised image. The average curvature extracted from the laser phase sample. The Gaussian curvature gradient of the denoised image. To pass Corrected Gaussian curvature gradient, The Gaussian curvature gradient of the laser phase sample; In response to a denoising command, the system acquires a target laser speckle image and a corresponding target laser phase image, and denoises the target laser speckle image based on the target laser phase image and the trained denoising model.

2. The method for denoising laser speckle images according to claim 1, characterized in that, The denoising model includes an encoder, a feature fusionist, and a decoder; The encoder includes four cascaded convolutional blocks, each of which contains two 3×3 convolutional layers, one batch normalization layer, and one max pooling layer. The feature fusion unit is connected to the last convolutional block in the cascade, and the feature fusion unit is used to fuse the output features of the encoder with the bimodal geometric fusion feature map to obtain intermediate fused features; The decoder includes four cascaded deconvolutional blocks, each of which contains one 3×3 deconvolutional layer, two 3×3 convolutional layers, and one batch normalization layer; the decoder is connected to the feature fusion unit. The process of fusing the output features of the encoder with the dual-modal geometric fusion feature map to obtain intermediate fused features includes: ； ； in, This is the dual-modal geometric fusion feature map. The output characteristics of the encoder, For attention mechanism functions, It is the sigmoid activation function. This is the weight matrix. For bias terms, It is the weighting coefficient.

3. The method for denoising laser speckle images according to claim 2, characterized in that, The process of determining the dual-modal geometric fusion feature map based on the contour feature map, the curvature feature map, and the normal vector feature map includes: ； in, The contour feature map, The curvature feature map, The normal vector feature map, These are the weight coefficients for the corresponding feature maps.

4. The method for denoising laser speckle images according to claim 2, characterized in that, The calculation process for the value includes: ； in, The weighted sum coefficient, Integers ranging from 1 to 4 It is a natural exponential function.

5. The method for denoising laser speckle images according to claim 2, characterized in that, The contour consistency constraints include: ； in, The first denoised image Class topological feature parameters, The first laser speckle sample Class topological feature parameters, The total number of classes of topological features. This is the function for finding the maximum value.

6. The method for denoising laser speckle images according to claim 2, characterized in that, A skip connection is provided between each convolutional block of the encoder and each corresponding deconvolutional block of the decoder; the skip connection includes: Calculate the fusion confidence weights based on the contour consistency constraint and the hypercurvature consistency constraint: ； ； ； in, The fusion confidence weights; The weighted fusion feature map is obtained by weighting the fusion confidence weights with the bimodal geometric fusion feature map: ； in, The output features of the skip connections of the convolutional block, The weighted fusion feature map; The weighted fusion feature map is passed to the corresponding deconvolution block.

7. The method for denoising laser speckle images according to claim 2, characterized in that, The denoising of the target laser speckle image based on the target laser phase image and according to the trained denoising model includes: Extract the target contour feature map from the target laser speckle image, and extract the target curvature feature map and target normal vector feature map from the target laser phase image. Based on the target contour feature map, the target curvature feature map and the target normal vector feature map, determine the target dual-modal geometric fusion feature map. The target dual-modal geometric fusion feature map, the target laser phase image, and the target laser speckle image are input into the trained denoising model to obtain the denoising result of the output target laser speckle image.

8. A noise reduction device for laser speckle images, characterized in that, The device includes: The training sample acquisition module is used to acquire laser speckle samples and their corresponding laser phase samples; The geometric feature extraction module is used to extract the contour feature map of the laser speckle sample, as well as the curvature feature map and normal vector feature map of the laser phase sample, and to determine the dual-modal geometric fusion feature map based on the contour feature map, the curvature feature map and the normal vector feature map; The model training module is used to train a denoising model based on the dual-modal geometric fusion feature map, the laser speckle samples, and the laser phase samples; wherein, the denoising model is used to output the denoised image corresponding to the laser speckle samples, and the loss function of the denoising model includes: ； ； ； ； ； in, Let be the loss function of the denoising model. For mean square error loss, For contour consistency constraints, For hypercurvature consistency constraints, For the normal vector direction constraint, For geometric topological consistency constraints, These are the weight coefficients for the corresponding constraints; For dynamic thresholds, The correlation coefficient, It is a natural exponential function. Based on the threshold, The angle between the normal vectors, Mean square error, The Gaussian curvature fitted to the denoised image, The Gaussian curvature extracted from the laser phase sample. The average curvature fitted to the denoised image. The average curvature extracted from the laser phase sample. The Gaussian curvature gradient of the denoised image. To pass Corrected Gaussian curvature gradient, The Gaussian curvature gradient of the laser phase sample; The target image denoising module is used to respond to denoising commands, acquire a target laser speckle image and a corresponding target laser phase image, and denoise the target laser speckle image based on the target laser phase image and the trained denoising model.

9. An electronic device, characterized in that, It includes at least one controller and a memory for communicatively connecting with the controller; the memory stores instructions executable by the at least one controller, which, when executed by the at least one controller, causes the at least one controller to perform a laser speckle image denoising method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions for causing a computer to perform a laser speckle image denoising method as described in any one of claims 1 to 7.

Images