Image Data Reconstruction Method and System Based on Spatial Region Awareness and Mask-Guided Design

Abstract

This invention provides an image data reconstruction method and system based on spatial region awareness and mask guidance, belonging to the field of image reconstruction technology. It compares the images in the training set with images in the training noise data to obtain a consistency comparison result. Based on the consistency comparison result, the anomaly segmentation network is fine-tuned to ensure that the segmentation results maintain a high degree of consistency in semantics and spatial distribution. The generated region-aware mask ensures that the decoding attention is always focused on the anomaly region, thereby improving the accuracy of anomaly noise removal while maintaining the physical structure and spatial continuity of the original data.

Description

Technical Field

[0001] This invention relates to the field of image reconstruction technology, and in particular to an image data reconstruction method and system based on spatial region perception and mask guidance. Background Technology

[0002] From a broad perspective of image reconstruction, 3D image data can be considered a special type of image with strong physical meaning and complex structure. The acquisition of this type of image data is generally affected by various degradation factors. These anomalous noises in 3D space often exhibit irregular distribution, abrupt changes in intensity, and diverse morphologies (such as stripes, clumps, or isolated points). The core challenge lies in recovering a clear, complete, and physically reliable original signal or structure from observation data contaminated by complex noise. This reconstruction problem is evident in multiple fields, including medical imaging, remote sensing, and industrial non-destructive testing. Among these, the removal of irregular anomalies is a common key technical challenge for improving the quality and usability of 3D image data.

[0003] Traditional 3D image denoising and reconstruction methods are mainly based on signal processing, statistical learning, and interpolation theory. In the spatial domain, methods such as median filtering and mean filtering smooth noise through local neighborhood operations, but they are extremely sensitive to outliers with abrupt changes in intensity (such as impulse noise) and easily lead to blurring or loss of details such as edges and textures. In the transform domain, methods such as frequency domain filtering or wavelet thresholding attempt to separate noise at specific frequencies or scales; however, their separation effect decreases significantly when there is severe aliasing between the noise spectrum and the effective signal spectrum. Methods based on statistics, such as principal component analysis or dictionary learning, assume that the data can be represented by a set of low-dimensional or sparse dictionaries, but these methods have limited modeling capabilities for anomalous noise with highly irregular spatial distributions and lacking global sparsity or stationary statistical properties. In addition, reconstruction methods based on geostatistics (such as Kriging interpolation) or variational models can introduce spatial continuity priors, but their reconstruction quality heavily depends on the precise setting of empirical parameters (such as variogram functions and regularization weights), resulting in insufficient adaptability and robustness for real-world scenarios with complex structures and varied noise patterns. The fundamental limitation of these traditional methods is that they adopt a globally uniform or simple prior-based processing paradigm, lacking the ability to specifically perceive and differentiate specific degraded areas in an image. Therefore, it is difficult to achieve the optimal balance between "completely removing noise" and "perfectly preserving structure".

[0004] With the development of deep learning technology, especially the maturity of convolutional neural networks and autoencoder architectures, data-driven image reconstruction methods have made groundbreaking progress. These models can learn complex nonlinear mappings from degraded observations to clear images from massive amounts of data, surpassing traditional methods in many image restoration tasks. Furthermore, researchers have introduced advanced architectures such as attention mechanisms, generative adversarial networks, and diffusion models to improve the quality and realism of generated images. However, when these advanced models are directly applied to 3D images containing strong anomalous noise with spatial locality, irregular shape, and varying intensity, their inherent limitations become apparent. Mainstream models typically use uniform convolutional kernels to process the entire image. This indiscriminate processing makes it difficult for the network to efficiently concentrate its learning capacity on anomalous noise that occupies only a small portion of the image. The consequences are twofold: firstly, noise suppression may be incomplete, especially in areas where noise and normal structural features are similar; secondly, over-processing to remove strong noise often damages or distorts the original structure and details of most normal areas, a problem known as over-smoothing or structural distortion. Summary of the Invention

[0005] This invention provides an image data reconstruction method and system based on spatial region perception and mask guidance, which aims to improve the accuracy of abnormal noise removal while maintaining the physical structure and spatial continuity of the original data.

[0006] To achieve the above objectives, the present invention provides an image data reconstruction method based on spatial region awareness and mask guidance, comprising:

[0007] Step 1: Acquire 3D geological data for training;

[0008] Step 2: Multi-dimensional slicing of the training 3D geological data to obtain slice data, and cyclic cross-sampling of the slice data to obtain multiple sets of image patches as multiple training sets;

[0009] Step 3: For each training set, input the training set into the anomaly segmentation network for identification to obtain the training segmentation result, which includes anomaly images and clean images. Then, process the training segmentation result to obtain training noise data and training region-aware mask.

[0010] Step 4: For each training set, perform consistency comparison between the images in the training set and the images in the training noise data to obtain the consistency comparison results. Fine-tune the anomaly segmentation network based on the consistency comparison results to obtain the trained anomaly segmentation network. Train the diffusion model based on the training segmentation results and the training region-aware mask to obtain the image reconstruction model.

[0011] Step 5: Input the three-dimensional geological data to be processed into the trained anomaly segmentation network for segmentation, obtain the segmentation result, and generate a region-aware mask on the segmentation result to obtain the masked image data;

[0012] Step 6: Input the masked image data into the image reconstruction model for 3D reconstruction to obtain the denoised 3D image.

[0013] Furthermore, before performing multi-dimensional slicing on the training 3D geological data, the following steps are also included:

[0014] Two mutually perpendicular virtual coordinate axes are established with the geometric center of the training 3D geological data as the origin;

[0015] By rotating the virtual coordinate axes around the center point, training 3D geological data from multiple perspectives can be obtained.

[0016] Furthermore, the training 3D geological data is sliced ​​in multiple dimensions to obtain slice data, including:

[0017] The training 3D geological data is horizontally cut along the Z-axis direction to obtain a horizontal cross-section.

[0018] Vertical sections are obtained by cutting the training 3D geological data perpendicularly along the X-axis or Y-axis direction.

[0019] The equation of the slice plane is constructed using a three-dimensional rotation matrix, and then the slope slices at arbitrary angles are generated based on the equation of the slice plane.

[0020] Furthermore, cyclic cross-sampling is performed on the sliced ​​data to obtain multiple sets of image patches as multiple training sets, including:

[0021] The sampling parameters of the three-dimensional geological data used for training are determined based on the training rounds of the model;

[0022] Random sampling is performed on horizontal sections, vertical sections, and inclined slices based on sampling parameters to obtain multiple sets of initial image patches;

[0023] Perform geometric transformations on each initial image patch to obtain multiple sets of image patches, which serve as multiple training sets.

[0024] Furthermore, the segmentation results used for training are processed to obtain training noise data and training region-aware masks, including:

[0025] The clean images from the segmentation results used in training are input into the anomaly generation network for processing to obtain noisy training data.

[0026] Abnormal images in the training segmentation results are processed using an adaptive thresholding method or morphological operations to obtain a training region-aware mask.

[0027] Furthermore, before training the diffusion model based on the segmentation results and the region-aware mask used for training, the following steps are also included:

[0028] Identify independent connected regions in the region-aware mask used for training;

[0029] Filtering rules are designed based on the attributes of independent connected regions;

[0030] Based on filtering rules, independent connected regions are filtered or modified to obtain optimized training region-aware masks.

[0031] Furthermore, the diffusion model is trained based on the segmentation results and the region-aware mask used for training, including:

[0032] Using the standard forward diffusion process, Gaussian noise is added to the clean image in the training segmentation results to generate a noisy image corresponding to the clean image.

[0033] The noise image corresponding to the clean image is labeled using the training region-aware mask to obtain the labeled noise image.

[0034] The diffusion model is trained using clean images and labeled noisy images.

[0035] Furthermore, during the training of the diffusion model based on clean images and labeled noisy images, a warm start and cosine annealing strategy is used to schedule the learning rate of the diffusion model.

[0036] Furthermore, a warm-start and cosine annealing strategy is used to schedule the learning rate of the diffusion model, including:

[0037] In the initial stage of model training, a warm start period is applied, during which the learning rate of the diffusion model is linearly increased from zero or a minimum value to a preset maximum threshold.

[0038] The learning rate of the diffusion model decays from a preset maximum threshold to a preset minimum threshold based on the curve of the cosine function.

[0039] The present invention also provides an image data reconstruction system based on spatial region awareness and mask guidance, comprising:

[0040] The acquisition module is used to acquire 3D geological data for training.

[0041] The sampling module is used to slice the training 3D geological data in multiple dimensions to obtain slice data, and to perform cyclic cross-sampling on the slice data to obtain multiple sets of image patches as multiple training sets.

[0042] The processing module is used to input the training set into the anomaly segmentation network for identification for each training set, and obtain the segmentation result for training. The segmentation result for training includes anomaly images and clean images. The processing module is used to process the segmentation result for training to obtain training noise data and training region-aware mask.

[0043] The training module is used to perform consistency comparison between the images in the training set and the images in the training noise data for each training set, obtain the consistency comparison results, fine-tune the anomaly segmentation network based on the consistency comparison results, obtain the trained anomaly segmentation network, and train the diffusion model based on the training segmentation results and the training region-aware mask to obtain the image reconstruction model.

[0044] The generation module is used to input the three-dimensional geological data to be processed into the trained anomaly segmentation network for segmentation, obtain the segmentation result, and generate a region-aware mask on the segmentation result to obtain the masked image data.

[0045] The reconstruction module is used to input the masked image data into the image reconstruction model for 3D reconstruction, and obtain a denoised 3D image.

[0046] The above-described solution of the present invention has the following beneficial effects:

[0047] This invention involves multi-dimensional slicing of training 3D geological data to obtain slice data, followed by cyclic cross-sampling within the slice data to obtain multiple sets of image patches as training sets. For each training set, the training set is input into an anomaly segmentation network for identification, yielding training segmentation results. These results are then processed to obtain training noise data and a training region-aware mask. For each training set, the images in the training set are compared with the images in the training noise data to obtain consistency comparison results. Based on these results, the anomaly segmentation network is fine-tuned to obtain a trained anomaly segmentation network. Finally, a diffusion model is trained using the training segmentation results and the training region-aware mask to obtain an image reconstruction model. The 3D geological data to be processed is input into a trained anomaly segmentation network for segmentation, resulting in segmentation results. A region-aware mask is then generated on the segmentation results to obtain masked image data. The masked image data is input into an image reconstruction model for 3D reconstruction to obtain a denoised 3D image. Compared with existing technologies, this invention performs consistency comparison between the images in the training set and the images in the training noise data to obtain consistency comparison results. Based on the consistency comparison results, the anomaly segmentation network is fine-tuned to ensure that the segmentation results maintain a high degree of consistency in semantics and spatial distribution. The generated region-aware mask ensures that the decoding attention is always focused on the anomaly region, thereby improving the accuracy of anomaly noise removal while maintaining the physical structure and spatial continuity of the original data.

[0048] Other beneficial effects of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of cyclic cross-sampling in an embodiment of the present invention;

[0051] Figure 3 This is a schematic diagram of the abnormal segmentation network in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the diffusion model in an embodiment of the present invention. Detailed Implementation

[0053] To make the technical problems, solutions, and advantages of this invention clearer, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0054] In the description of this invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0055] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a locking connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0056] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0057] This invention addresses existing problems by providing an image data reconstruction method and system based on spatial region perception and mask guidance.

[0058] like Figure 1 As shown, embodiments of the present invention provide an image data reconstruction method based on spatial region awareness and mask guidance, comprising:

[0059] Step 1: Acquire 3D geological data for training;

[0060] Step 2: Multi-dimensional slicing of the training 3D geological data to obtain slice data, and cyclic cross-sampling of the slice data to obtain multiple sets of image patches as multiple training sets;

[0061] Step 3: For each training set, input the training set into the anomaly segmentation network for identification to obtain the training segmentation result, which includes anomaly images and clean images. Then, process the training segmentation result to obtain training noise data and training region-aware mask.

[0062] Step 4: For each training set, perform consistency comparison between the images in the training set and the images in the training noise data to obtain the consistency comparison results. Fine-tune the anomaly segmentation network based on the consistency comparison results to obtain the trained anomaly segmentation network. Train the diffusion model based on the training segmentation results and the training region-aware mask to obtain the image reconstruction model.

[0063] Step 5: Input the three-dimensional geological data to be processed into the trained anomaly segmentation network for segmentation, obtain the segmentation result, and generate a region-aware mask on the segmentation result to obtain the masked image data;

[0064] Step 6: Input the masked image data into the image reconstruction model for 3D reconstruction to obtain the denoised 3D image.

[0065] Specifically, before performing multi-dimensional slicing on the training 3D geological data, the following steps are also included:

[0066] Two mutually perpendicular virtual coordinate axes are established with the geometric center of the training 3D geological data as the origin;

[0067] To ensure that the slices cover the entire space, the virtual coordinate axes are rotated around the center point to obtain training 3D geological data from multiple perspectives.

[0068] In embodiments of the present invention, such as Figure 2 As shown, the virtual coordinate axes are the X-axis and Y-axis. The direction of the virtual coordinate axes can be initialized based on the principal component analysis method of the data or the Manhattan world hypothesis (such as the alignment direction of the walls and the ground).

[0069] Specifically, the orientation of the virtual coordinate axes can be initialized based on principal component analysis of the data or the Manhattan world hypothesis (such as wall and floor alignment orientation), including:

[0070] Extract the spatial coordinates of all points in the training 3D geological data and calculate their mean;

[0071] This mean value is determined as the geometric center of the three-dimensional geological data and used as the origin of the virtual coordinate axis;

[0072] If based on the principal component analysis method, the covariance matrix is ​​solved after decentering the coordinates of all points. The two orthogonal eigenvectors with the largest spatial distribution variance are selected by eigenvalue decomposition and used as the directions of the virtual X-axis and Y-axis, respectively.

[0073] If based on the Manhattan world hypothesis, the three-dimensional geological data is first smoothed by Gaussian and the three-dimensional Sobel gradient is calculated. The unit gradient direction of all values ​​in the three-dimensional volume is extracted and clustered by K-means. Then, the orthogonal Manhattan main direction is obtained by Schmidt orthogonalization. Combining the orientation requirements of wall and ground alignment, two mutually perpendicular directions are selected as the directions of the virtual X-axis and Y-axis.

[0074] In this embodiment of the invention, the rotation scheme can adopt equal-interval angle sampling, for example, rotating once every 10°, rotating a total of 18 times for 180°, generating 18 sets of three-dimensional geological data for training from different perspectives. In order to avoid redundant slices, the perspective distribution can also be optimized based on information entropy.

[0075] Specifically, the training 3D geological data is sliced ​​in multiple dimensions to obtain slice data, including:

[0076] The training 3D geological data is horizontally cut along the Z-axis to obtain a horizontal cross-section, in order to preserve lateral continuity.

[0077] Vertical sections are obtained by cutting the training 3D geological data vertically along the X-axis or Y-axis direction to capture longitudinal variations.

[0078] The equation of the slice plane is constructed using a three-dimensional rotation matrix, and then the slope slices at arbitrary angles are generated based on the equation of the slice plane.

[0079] In this embodiment of the invention, the equation of the slice plane is:

[0080] ;

[0081] in, It is the normal vector of the plane. These are the coordinates of the inclined plane slice.

[0082] Specifically, when generating arbitrarily angled slices, the virtual coordinate axes with the geometric center as the origin are first rotated around the center point by the target angle based on a three-dimensional rotation matrix, thus obtaining the normal vector of the slice plane at the corresponding angle. And determine the equation of the slice plane. Then, iterate through all voxel coordinates of the 3D geological data used for training. Substituting this value into the plane equation, and calculating the voxel values ​​that satisfy the equation through linear interpolation, we can fully capture the spatial distribution characteristics of the three-dimensional geological data at this tilt angle.

[0083] Specifically, cyclic cross-sampling is performed on the sliced ​​data to obtain multiple sets of image patches as multiple training sets, including:

[0084] The sampling parameters of the three-dimensional geological data used for training are determined based on the training rounds of the model;

[0085] Random sampling is performed on horizontal sections, vertical sections, and inclined slices based on sampling parameters to obtain multiple sets of initial image patches;

[0086] Perform geometric transformations on each initial image patch to obtain multiple sets of image patches, which serve as multiple training sets.

[0087] Specifically, the sampling parameters include spatial location. In this embodiment of the invention, the Markov chain Monte Carlo method is used to guide the image patch to shift towards the abnormal edge region based on the model loss gradient of the previous training round. It also includes size and orientation. In this embodiment of the invention, based on an incremental learning strategy, a large size is used to capture global features in the initial stage of model training, and the size is gradually reduced in the later stage to focus on local details. It also includes sampling weights. In this embodiment of the invention, the sampling weights are updated according to the information entropy of the slice. The sampling weights of high-entropy regions, i.e., regions with complex textures, are larger, which means that a higher sampling rate is required.

[0088] In this embodiment of the invention, the size and aspect ratio of multiple initial image blocks obtained by random sampling in horizontal sections, vertical sections and inclined slices according to sampling parameters are variable, ranging from 0.5 to 2.0; at the same time, it is necessary to avoid oversampling high-density areas.

[0089] In this embodiment of the invention, geometric transformation is performed on each group of initial image blocks, including:

[0090] Each initial image block is rotated 30° and then scaled by 0.8 to 1.2 times, and finally flipped horizontally or vertically.

[0091] In this embodiment of the invention, before inputting the training set into the anomaly segmentation network for identification, 5% of the image blocks in the training set are randomly selected and pre-trained large model is used for preliminary segmentation. Then, interactive correction combined with manual verification tools is used to correct obvious errors in the preliminary segmentation results.

[0092] It should be noted that the pre-trained large model can be a deep learning-based image segmentation model (SAM, SegmentAnything Model) or a semantic segmentation network DeepLabV3+.

[0093] Specifically, the segmentation results used for training are processed to obtain training noise data and training region-aware masks, including:

[0094] The clean images from the segmentation results used in training are input into the anomaly generation network for processing to obtain noisy training data.

[0095] Abnormal images in the training segmentation results are processed using an adaptive thresholding method or morphological operations to obtain a training region-aware mask.

[0096] like Figure 3As shown, in this embodiment of the invention, the backbone network of the anomaly segmentation network adopts a U-Net or Transformer encoder-decoder structure, which supports multi-scale feature extraction and embeds a spatial region attention mechanism in the decoder to calculate the spatial weights of the feature maps. Regions with high weights correspond to suspected anomaly regions to enhance spatial continuity. The loss function of the anomaly segmentation network consists of Dice loss, cross-entropy loss, and geometric constraint function. The geometric constraint function is a smoothness penalty function for anomaly regions.

[0097] It should be noted that the adaptive thresholding method is an algorithm that automatically determines the global threshold. Its principle is to traverse the pixels of the abnormal image in the training segmentation results, calculate the inter-class variance of the pixels in the foreground and background classes, and select the threshold that maximizes the inter-class variance as the optimal segmentation threshold. Then, the optimal segmentation threshold is used to divide the image block into multiple small blocks, and a mask is generated independently for each small block to better adapt to local contrast changes.

[0098] It should be noted that the process of generating a mask based on morphological operations, such as opening operations, involves first eroding the abnormal images in the training segmentation results, and then dilating the eroded images. This can effectively eliminate isolated noise and small spikes in the mask, and can separate regions that are unintentionally connected without significantly changing the area of ​​the remaining regions.

[0099] Specifically, before training the diffusion model based on the segmentation results and the region-aware mask used for training, the following steps are also included:

[0100] Identify independent connected regions in the region-aware mask used for training;

[0101] Filtering rules are designed based on the attributes of independent connected regions;

[0102] Based on filtering rules, independent connected regions are filtered or modified to obtain optimized training region-aware masks.

[0103] Specifically, the process of designing filtering rules based on the properties of independent connected regions is as follows:

[0104] First, extract the core attributes of each independent connected region in the training region-aware mask, including region area, shape features (such as aspect ratio and roundness), spatial location (such as relative distance from the core region), and gray-scale statistical features (such as mean and variance).

[0105] The filtering rules are set to retain connected regions with an area greater than a preset threshold, a shape that conforms to the natural distribution of geological bodies (such as an aspect ratio within a reasonable range) and located within the effective geological exploration range, while removing isolated noise regions with too small an area, abnormal regions with distorted shapes, and irrelevant regions that are outside the effective exploration range. At the same time, morphological smoothing correction is performed on effective regions with irregular edges but whose core attributes meet the requirements.

[0106] During the design process, the initial threshold range of each attribute is first determined by combining prior geological knowledge. Then, the threshold parameters are calibrated by statistically analyzing the attribute distribution characteristics of normal geological areas in the segmentation results. Finally, a filtering rule is formed that mainly uses attribute threshold screening and supplemented by morphological correction.

[0107] Specifically, the diffusion model is trained based on the segmentation results and the region-aware mask used for training, including:

[0108] Using the standard forward diffusion process, Gaussian noise is added to the clean image in the training segmentation results to generate a noisy image corresponding to the clean image.

[0109] The noise image corresponding to the clean image is labeled using the training region-aware mask to obtain the labeled noise image.

[0110] The diffusion model is trained using clean images and labeled noisy images.

[0111] This invention generates a noisy image corresponding to a clean image based on the principle of linear noise modulation. The noise intensity is controlled by a variance scheme (such as linear or cosine scheduling). The core of the principle of linear noise modulation is to use noise as a modulation signal and shift its spectrum to the vicinity of the carrier frequency through linear modulation, thereby generating an interference signal. In this process, Gaussian white noise controls a certain parameter of the high-frequency carrier and makes it change linearly with the noise signal. The result is a linear shift of the original noise spectrum structure in the spectrum, rather than the generation of new frequency components.

[0112] Specifically, the noisy image corresponding to the clean image is labeled using the region-aware mask trained on the training, resulting in the labeled noisy image, including:

[0113] Use the noisy image and its corresponding clean image as a set of training data;

[0114] By encoding the training region-aware mask into a pixel-level weight map through spatial multiplication and attaching it to each set of training data, abnormal regions are explicitly identified and given higher reconstruction priority during training, thereby guiding the diffusion model to focus on the abnormal noise regions specified by the mask during the recovery process.

[0115] like Figure 4As shown, in this embodiment of the invention, the diffusion model is based on the U-Net diffusion model backbone. A cross-attention mechanism is added to the encoder, and the feature map is modulated under mask conditions. Time step information is injected into the residual block through sinusoidal modulation embedding. Noise data at any time, the current diffusion time step, and the corresponding region-aware mask are used as joint inputs. The goal is to predict the noise added to the data. Therefore, the network structure needs to be able to fuse the spatial conditional information of the mask. In the forward diffusion recovery stage of model training, conditional modulation is performed under the guidance of the region-aware mask. For abnormal regions identified by the mask, stronger or more directional noise perturbations are applied to make them degenerate to a pure noise state more quickly. For normal regions outside the mask, the original signal structure is preserved. The network parameters are optimized by maximizing the lower bound of evidence. This process jointly optimizes all latent variables related to the diffusion process. Gradient clipping is used to limit the gradient norm and avoid gradient explosion. Multi-stage training is adopted. The unconditional model is pre-trained first, and then the conditional branches are fine-tuned to improve convergence efficiency. This optimization strategy helps to stabilize training and accelerate convergence.

[0116] The preferred approach is to use a warm start and cosine annealing strategy to schedule the learning rate of the diffusion model during the training process based on clean images and labeled noisy images.

[0117] Specifically, a warm-start and cosine annealing strategy is used to schedule the learning rate of the diffusion model, including:

[0118] In the initial stage of model training, a warm start period is applied, during which the learning rate of the diffusion model is linearly increased from zero or a minimum value to a preset maximum threshold.

[0119] The learning rate of the diffusion model decays from a preset maximum threshold to a preset minimum threshold based on the curve of the cosine function. Its mathematical formula is typically expressed as:

[0120] ;

[0121] in, Indicates the learning rate. This indicates the preset minimum threshold. This indicates the preset maximum threshold. Indicates the current iteration step or cycle count. This represents half the length of the annealing cycle, which is half the period of the cosine curve. This smooth descent helps the model to fine-tune parameters later in the training process in order to approach a better local minimum.

[0122] This invention embodiment involves multi-dimensional slicing of training 3D geological data to obtain slice data, followed by cyclic cross-sampling within the slice data to obtain multiple sets of image patches as training sets. For each training set, the training set is input into an anomaly segmentation network for identification, yielding training segmentation results. These results are then processed to obtain training noise data and a training region-aware mask. For each training set, the images in the training set are compared with the images in the training noise data to obtain consistency comparison results. Based on these results, the anomaly segmentation network is fine-tuned to obtain a trained anomaly segmentation network. Finally, a diffusion model is trained using the training segmentation results and the training region-aware mask to obtain an image reconstruction model. The 3D geological data to be processed is input into a trained anomaly segmentation network for segmentation, resulting in segmentation results. A region-aware mask is then generated on the segmentation results to obtain masked image data. The masked image data is input into an image reconstruction model for 3D reconstruction to obtain a denoised 3D image. Compared with existing technologies, this embodiment of the invention performs consistency comparison between the images in the training set and the images in the training noise data to obtain consistency comparison results. Based on the consistency comparison results, the anomaly segmentation network is fine-tuned to ensure that the segmentation results maintain a high degree of consistency in semantics and spatial distribution. The generated region-aware mask ensures that the decoding attention is always focused on the anomaly region, thereby improving the accuracy of anomaly noise removal while maintaining the physical structure and spatial continuity of the original data.

[0123] Corresponding to the image data reconstruction method based on spatial region awareness and mask guidance described in the above embodiments, this invention also provides an image data reconstruction system based on spatial region awareness and mask guidance, comprising:

[0124] The acquisition module is used to acquire 3D geological data for training.

[0125] The sampling module is used to slice the training 3D geological data in multiple dimensions to obtain slice data, and to perform cyclic cross-sampling on the slice data to obtain multiple sets of image patches as multiple training sets.

[0126] The processing module is used to input the training set into the anomaly segmentation network for identification for each training set, and obtain the segmentation result for training. The segmentation result for training includes anomaly images and clean images. The processing module is used to process the segmentation result for training to obtain training noise data and training region-aware mask.

[0127] The training module is used to perform consistency comparison between the images in the training set and the images in the training noise data for each training set, obtain the consistency comparison results, fine-tune the anomaly segmentation network based on the consistency comparison results, obtain the trained anomaly segmentation network, and train the diffusion model based on the training segmentation results and the training region-aware mask to obtain the image reconstruction model.

[0128] The generation module is used to input the three-dimensional geological data to be processed into the trained anomaly segmentation network for segmentation, obtain the segmentation result, and generate a region-aware mask on the segmentation result to obtain the masked image data.

[0129] The reconstruction module is used to input the masked image data into the image reconstruction model for 3D reconstruction, and obtain a denoised 3D image.

[0130] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for reconstructing image data based on spatial region perception and mask guidance, characterized in that, include: Step 1: Acquire 3D geological data for training; Step 2: Multi-dimensional slicing is performed on the training 3D geological data to obtain slice data, and cyclic cross-sampling is performed on the slice data to obtain multiple sets of image patches as multiple training sets; Step 3: For each training set, input the training set into the anomaly segmentation network for identification to obtain the training segmentation result. The training segmentation result includes anomaly images and clean images. Then, process the training segmentation result to obtain training noise data and training region-aware mask. Step 4: For each training set, perform consistency comparison between the images in the training set and the images in the training noise data to obtain consistency comparison results. Fine-tune the anomaly segmentation network based on the consistency comparison results to obtain the trained anomaly segmentation network. Train the diffusion model based on the training segmentation results and the training region-aware mask to obtain the image reconstruction model. Step 5: Input the three-dimensional geological data to be processed into the trained anomaly segmentation network for segmentation to obtain the segmentation result, and generate a region-aware mask on the segmentation result to obtain the masked image data. Step 6: Input the masked image data into the image reconstruction model for three-dimensional reconstruction to obtain a denoised three-dimensional image.

2. The image data reconstruction method based on spatial region perception and mask guidance according to claim 1, characterized in that, Before performing multi-dimensional slicing on the training 3D geological data, the following steps are also included: Two mutually perpendicular virtual coordinate axes are established with the geometric center of the training 3D geological data as the origin; Rotating the virtual coordinate axes around the center point yields three-dimensional geological data for training from multiple perspectives.

3. The image data reconstruction method based on spatial region perception and mask guidance according to claim 1, characterized in that, The training 3D geological data is sliced ​​in multiple dimensions to obtain slice data, including: The training three-dimensional geological data is horizontally cut along the Z-axis direction to obtain a horizontal cross-section. The training three-dimensional geological data is vertically cut along the X-axis or Y-axis direction to obtain a vertical profile. A three-dimensional rotation matrix is ​​used to construct the equation of the slicing plane, and then an inclined slice with an arbitrary angle is generated based on the equation of the slicing plane.

4. The image data reconstruction method based on spatial region perception and mask guidance according to claim 3, characterized in that, Cyclic cross-sampling is performed on the slice data to obtain multiple sets of image patches as multiple training sets, including: The sampling parameters of the three-dimensional geological data used for training are determined according to the training rounds of the model; Random sampling is performed on the horizontal section, the vertical section, and the inclined slice according to the sampling parameters to obtain multiple sets of initial image blocks; Perform geometric transformations on each initial image patch to obtain multiple sets of image patches, which serve as multiple training sets.

5. The image data reconstruction method based on spatial region perception and mask guidance according to claim 4, characterized in that, The segmentation results used for training are processed to obtain training noise data and training region-aware masks, including: The clean image from the training segmentation result is input into the anomaly generation network for processing to obtain training noise data; Abnormal images in the training segmentation results are processed using an adaptive thresholding method or morphological operations to obtain a training region-aware mask.

6. The image data reconstruction method based on spatial region perception and mask guidance according to claim 1, characterized in that, Before training the diffusion model based on the training segmentation results and the training region-aware mask, the method further includes: Identify independent connected regions in the region-aware mask used for training; Filtering rules are designed based on the attributes of the independent connected regions; Based on the filtering rules, the independent connected regions are filtered or modified to obtain an optimized training region-aware mask.

7. The image data reconstruction method based on spatial region perception and mask guidance according to claim 1, characterized in that, The diffusion model is trained based on the segmentation results used for training and the region-aware mask used for training, including: Using a standard forward diffusion process, Gaussian noise is added to the clean image in the training segmentation result to generate a noisy image corresponding to the clean image; The noise image corresponding to the clean image is labeled using the training region-aware mask to obtain the labeled noise image. The diffusion model is trained based on the clean image and the labeled noisy image.

8. The image data reconstruction method based on spatial region perception and mask guidance according to claim 7, characterized in that, During the training of the diffusion model based on the clean image and the labeled noisy image, a warm start and cosine annealing strategy are used to schedule the learning rate of the diffusion model.

9. The image data reconstruction method based on spatial region perception and mask guidance according to claim 8, characterized in that, The learning rate of the diffusion model is scheduled using a warm-start and cosine annealing strategy, including: In the initial stage of model training, a warm start period is applied, during which the learning rate of the diffusion model is linearly increased from zero or a minimum value to a preset maximum threshold. The learning rate of the diffusion model decreases from the preset maximum threshold to the preset minimum threshold based on the curve of the cosine function.

10. An image data reconstruction system based on spatial region perception and mask guidance, characterized in that, include: The acquisition module is used to acquire 3D geological data for training. The sampling module is used to slice the training 3D geological data in multiple dimensions to obtain slice data, and to perform cyclic cross-sampling in the slice data to obtain multiple sets of image patches as multiple training sets. The processing module is used to input the training set into the anomaly segmentation network for identification for each training set, and obtain the segmentation result for training. The segmentation result for training includes anomaly images and clean images. The processing module is used to process the segmentation result for training to obtain training noise data and training region-aware mask. The training module is used to perform consistency comparison between the images in the training set and the images in the training noise data for each training set, obtain the consistency comparison result, fine-tune the anomaly segmentation network according to the consistency comparison result, obtain the trained anomaly segmentation network, and train the diffusion model based on the training segmentation result and the training region-aware mask to obtain the image reconstruction model. The generation module is used to input the three-dimensional geological data to be processed into the trained anomaly segmentation network for segmentation, obtain the segmentation result, and generate a region-aware mask on the segmentation result to obtain the masked image data. The reconstruction module is used to input the masked image data into the image reconstruction model for three-dimensional reconstruction to obtain a denoised three-dimensional image.

Images