Deep learning-based weak intercalated layer semantic segmentation method and device, and medium

Abstract

The invention discloses a weak intercalated layer semantic segmentation method and device based on deep learning and a medium, and belongs to the technical field of semantic segmentation. The method comprises the following steps: acquiring and preprocessing weak intercalated layer image data, generating normalized image data and enhanced image data, processing the normalized image data and the enhanced image data, and generating a multi-scale fusion feature map; processing the multi-scale fusion feature map through an attention network guided by geological prior, generating an attention weighted feature map, processing the attention weighted feature map, and generating an initial segmentation prediction map; calculating the difference between the initial segmentation prediction map and a real label, and generating optimized gradient data; iteratively updating model parameters based on the optimized gradient data to generate a final semantic segmentation model; and based on the final semantic segmentation model, performing segmentation operation on the target weak intercalated layer image to generate an intercalated layer segmentation result graph. According to the method, the technical effect of improving the low-contrast weak intercalated layer segmentation precision and generalization ability is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semantic segmentation, and particularly relates to a soft interlayer semantic segmentation method based on deep learning, a device and a medium. BACKGROUND

[0002] In geological exploration and engineering safety monitoring, semantic segmentation of soft interlayers (such as rock fissures and fault zones) is a key link for evaluating geological stability and risk. However, traditional image processing methods rely on manually designed features (such as edge detection and texture analysis), which are difficult to cope with the characteristics of low contrast, thin structure and complex texture of soft interlayers. Although the model based on deep learning can extract multi-scale features, the ASPP module with fixed hole rate parameters is not adaptive to the curvature change of the interlayer, and the small target segmentation accuracy is low. In addition, the lack of guidance of geological prior knowledge leads to limited generalization ability of the model in the data scarce scene.

[0003] The current technology also faces the contradiction between the scarcity of labeled data and the deployment efficiency: the labeling of soft interlayers requires the participation of professional personnel, which is costly and difficult to obtain on a large scale; at the same time, the existing model mostly uses large-scale networks (such as ResNet-101), which is difficult to balance the real-time performance and computing resource demand on embedded devices. Although existing research has tried to introduce dynamic feature perception, attention mechanism and adaptive loss function, the technical solutions still have problems such as parameter fixation, insufficient integration of geological priori knowledge and insufficient lightweight optimization.

[0004] Therefore, how to improve the segmentation accuracy and generalization ability of low-contrast soft interlayers becomes a technical problem to be solved. SUMMARY

[0005] The embodiments of the present application provide a soft interlayer semantic segmentation method based on deep learning, a device and a medium, to solve the technical problem of improving the segmentation accuracy and generalization ability of low-contrast soft interlayers.

[0006] In a first aspect, the embodiments of the present application provide a soft interlayer semantic segmentation method based on deep learning, characterized in that the method comprises: acquiring soft interlayer image data, wherein the soft interlayer image data comprises geological exploration images, engineering detection images and synthetic data; preprocessing the soft interlayer image data to generate normalized image data and enhanced image data; constructing a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate a multi-scale fusion feature map; the dynamic feature perception architecture comprises an adaptive dilated convolution mechanism and a cross-scale feature interaction module; processing the multi-scale fusion feature map through a geological prior guided attention network to generate an attention weighted feature map; the geological prior guided attention network fuses channel attention weights, spatial attention weights and geological structure parameter constraint factors; processing the attention weighted feature map based on a preset encoder-decoder structure to generate an initial segmentation prediction map; calculating the difference between the initial segmentation prediction map and a real label through an adaptive loss function module to generate optimization gradient data; iteratively updating model parameters based on the optimization gradient data to generate a final semantic segmentation model; and performing a segmentation operation on a target soft interlayer image based on the final semantic segmentation model to generate an interlayer segmentation result map.

[0007] In an implementation manner of the present application, the soft interlayer image data is preprocessed to generate normalized image data and enhanced image data, specifically comprising: extracting multi-level feature maps of an input image through an encoder to generate a first level feature map, a second level feature map, a third level feature map and a fourth level feature map; processing the fourth level feature map through an adaptive dilated convolution mechanism to generate a dynamic dilated rate parameter and multi-branch output features; fusing the first level feature map, the second level feature map, the third level feature map and the multi-branch output features through a cross-scale feature interaction module to generate a multi-scale fusion feature map; introducing residual connection in a decoder to process the fourth level feature map and the multi-scale fusion feature map to generate a high resolution feature map.

[0008] In an implementation manner of the present application, the fourth level feature map is processed through the adaptive dilated convolution mechanism to generate the dynamic dilated rate parameter and the multi-branch output features, specifically comprising: performing a global average pooling operation on an input feature map to generate a local feature vector; inputting the local feature vector into a learnable parameter matrix to generate a dynamic dilated rate parameter through Softmax normalization; adjusting a convolution kernel receptive field according to the dynamic dilated rate parameter to perform a multi-branch dilated convolution operation on the input feature map to generate multi-branch output features.

[0009] In an implementation form of the present application, the dynamic feature-aware architecture is constructed to process the normalized image data and the enhanced image data to generate a multi-scale fusion feature map, specifically comprising: extracting channel features of the multi-scale fusion feature map to generate channel attention weights through a fully connected layer; extracting spatial features of the channel-weighted feature map to generate spatial attention weights through a convolution layer; generating an interlayer geometric parameter map through a pre-trained geological feature extractor and converting it into a geological constraint factor; and fusing the spatial attention weights and the geological constraint factor to generate an attention-weighted feature map.

[0010] In an implementation form of the present application, the interlayer geometric parameter map is generated through a pre-trained geological feature extractor and converted into a geological constraint factor, specifically comprising: extracting an interlayer curvature parameter and an extension direction parameter from the input image to generate a geological structure parameter tensor; performing a convolution operation on the geological structure parameter tensor to generate an initial constraint factor; and normalizing the initial constraint factor through a Sigmoid function to generate a geological constraint factor.

[0011] In an implementation form of the present application, the difference between the initial segmentation prediction map and the real label is calculated through an adaptive loss function module to generate optimization gradient data, specifically comprising: calculating a Dice loss value and a Focal loss value of the initial segmentation prediction map and the real label; dynamically assigning a Dice loss weight and a Focal loss weight according to the interlayer density parameter to generate a weighted joint loss value; extracting an edge mask of the real label through an edge detection algorithm to calculate an edge-aware loss value; and fusing the weighted joint loss value and the edge-aware loss value to generate optimization gradient data.

[0012] In an implementation form of the present application, the Dice loss weight and the Focal loss weight are dynamically assigned according to the interlayer density parameter to generate a weighted joint loss value, specifically comprising: counting the number of interlayer pixels in a unit area to generate an interlayer density parameter; inputting the interlayer density parameter into a Sigmoid function to generate a Dice loss dynamic weight; and calculating a Focal loss dynamic weight according to the Dice loss dynamic weight.

[0013] In an implementation form of the present application, the edge mask of the real label is extracted through an edge detection algorithm to calculate an edge-aware loss value, specifically comprising: performing a Canny edge detection operation on the real label to generate a binary edge mask; extracting pixel prediction values of edge regions in the initial segmentation prediction map to generate an edge prediction subset; and calculating a cross-entropy loss of the edge prediction subset and the real edge mask to generate an edge-aware loss value.

[0014] In a second aspect, the embodiments of the present application also provide a soft interlayer semantic segmentation device based on deep learning, which is characterized by comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: acquire soft interlayer image data, wherein the soft interlayer image data comprises geological exploration images, engineering detection images and synthetic data; pre-process the soft interlayer image data to generate normalized image data and enhanced image data; construct a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate multi-scale fusion feature maps; the dynamic feature perception architecture comprises an adaptive atrous convolution mechanism and a cross-scale feature interaction module; process the multi-scale fusion feature maps through a geological prior guided attention network to generate attention weighted feature maps; the geological prior guided attention network fuses channel attention weights, spatial attention weights and geological structure parameter constraint factors; process the attention weighted feature maps based on a preset encoder-decoder structure to generate an initial segmentation prediction map; calculate the difference between the initial segmentation prediction map and a real label through an adaptive loss function module to generate optimization gradient data; iteratively update model parameters based on the optimization gradient data to generate a final semantic segmentation model; and perform segmentation operation on a target soft interlayer image based on the final semantic segmentation model to generate an interlayer segmentation result map.

[0015] In a third aspect, the embodiments of the present application also provide a non-volatile computer storage medium based on deep learning of soft interlayer semantic segmentation, which stores computer executable instructions, and the computer executable instructions are configured to: acquire soft interlayer image data, wherein the soft interlayer image data comprises geological exploration images, engineering detection images and synthetic data; pre-process the soft interlayer image data to generate normalized image data and enhanced image data; construct a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate multi-scale fusion feature maps; the dynamic feature perception architecture comprises an adaptive atrous convolution mechanism and a cross-scale feature interaction module; process the multi-scale fusion feature maps through a geological prior guided attention network to generate attention weighted feature maps; the geological prior guided attention network fuses channel attention weights, spatial attention weights and geological structure parameter constraint factors; process the attention weighted feature maps based on a preset encoder-decoder structure to generate an initial segmentation prediction map; calculate the difference between the initial segmentation prediction map and a real label through an adaptive loss function module to generate optimization gradient data; iteratively update model parameters based on the optimization gradient data to generate a final semantic segmentation model; and perform segmentation operation on a target soft interlayer image based on the final semantic segmentation model to generate an interlayer segmentation result map.

[0016] The embodiments of the present application provide a method, device, and medium for semantic segmentation of weak interlayers based on deep learning, which have at least the following technical effects: Through the adaptive dilated convolution mechanism of the dynamic feature perception architecture, the receptive field of the convolution kernel is dynamically adjusted according to the local features of the input image, effectively enhancing the edge capture capability of slender and curved structures; combined with the shallow details and deep semantic fusion strategy of the cross-scale feature interaction module, the accuracy of interlayer recognition in low-contrast areas is significantly improved, reducing missed segmentation.

[0017] The geological prior-guided attention network integrates channel attention weights, spatial attention weights, and geological structure parameter constraint factors. It guides the feature weighting process through interlayer curvature and extension direction parameters, enhances the model's semantic understanding of complex geological structures, and maintains stable segmentation robustness in data-scarce scenarios.

[0018] The adaptive loss function module jointly combines weighted Dice-FocalLoss and edge-aware loss, dynamically allocates loss weights based on the interlayer density parameter, and focuses on optimizing the segmentation effect of small targets; at the same time, it introduces an edge detection constraint mechanism to enhance the accuracy of interlayer boundary segmentation and significantly reduce edge blur errors.

[0019] The lightweight deployment solution compresses the model size through knowledge distillation technology, combines parameter quantization with hardware acceleration strategies, and achieves high frame rate inference on embedded devices. While ensuring segmentation accuracy, it effectively balances computing resource consumption and real-time requirements to meet the engineering application requirements of geological exploration field monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A flowchart of a method for semantic segmentation of weak interlayers based on deep learning provided in an embodiment of the present application; Figure 2 A schematic diagram of a dynamic feature perception architecture in a weak interlayer semantic segmentation method based on deep learning provided in an embodiment of the present application; Figure 3 A schematic diagram of the internal structure of a weak interlayer semantic segmentation device based on deep learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to make the purposes, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0022] The embodiments of the present application provide a soft interlayer semantic segmentation method and device based on deep learning, to solve the technical problem of improving the segmentation accuracy and generalization ability of low-contrast soft interlayer.

[0023] The technical solutions provided by the embodiments of the present application will be described in detail below with reference to the drawings.

[0024] Figure 1 A soft interlayer semantic segmentation flowchart based on deep learning is provided for the embodiments of the present application. As shown in Figure 1 The soft interlayer semantic segmentation method based on deep learning provided by the embodiments of the present application specifically includes the following steps: Step 1, obtaining soft interlayer image data, the soft interlayer image data including geological exploration images, engineering detection images and synthetic data.

[0025] Soft interlayer image data refers to original images containing geological weak zones (such as rock layer cracks and faults), including geological exploration images (satellite or aerially photographed geological structure images), engineering detection images (on-site collected engineering structure images) and synthetic data (simulated images generated by simulation software). These data are used to train and verify the semantic segmentation model to improve the recognition ability of low-contrast interlayer. Geological exploration image acquisition: download the geological images of a certain mining area from a satellite platform (such as Sentinel-2), the images covering the fault and rock layer distribution area. The resolution of these images is not less than 10 meters / pixel, ensuring that the fine interlayer structure can be captured.

[0026] Engineering detection image acquisition: using a drone device to collect high-resolution images (resolution ≥ 1920x1080) at a certain tunnel engineering site B, focusing on shooting the slope or rock surface to cover the crack and mud interlayer target.

[0027] Synthetic data generation: using a geological simulation software (such as Petrel) to create simulated images, simulating soft interlayers with different curvatures and extension directions. The simulation parameters include interlayer density (0.1~0.5 pixels / unit area) and texture complexity to supplement the lack of real data. All image data are stored in RGB format and ensure that there is no personal sensitive information.

[0028] Step 2. Preprocess the weak interlayer image data to generate normalized and augmented image data.

[0029] Preprocessing includes normalization (standardizing pixel value range) and augmentation (enhancing data diversity through transformation). This generates normalized image data (pixel values unified to the range [0, 1]) and augmented image data (images after applying geometric or color transformations). These operations aim to improve the model's generalization ability for low-contrast scenes.

[0030] Normalization: Perform a linear mapping on the original image to scale the pixel values from the original range to the interval [0, 1]. The specific formula is: pixel value = (original value - minimum value) / (maximum value - minimum value). This eliminates the effects of lighting differences and ensures input consistency.

[0031] Data Augmentation: Apply random transformations to generate augmented image data, including: Geometric Augmentation: Random rotation (angle range ±30°), scaling (scale 0.8~1.2 times) to simulate different shooting angles.

[0032] Color Augmentation: Adjust brightness (shift ±0.2), contrast (scale ±0.3) to enhance the visibility of low-contrast interlayers.

[0033] Augmented data expands the size of the data set, reducing the risk of overfitting.

[0034] In a specific example, image preprocessing includes: 1) Image Normalization Normalize the pixel values of the input image to the range [0, 1], with the formula where I is the initial image, is the normalized image.

[0035] 2) Data Augmentation Geometric Transformation: Random rotation (±30°), translation (±5%), scaling (0.8~1.2 times); Color Transformation: Adjust brightness (±0.2), contrast (±0.3), saturation (±0.3); Synthetic Data Generation: Generate interlayer distribution maps through geological structure simulation software and fuse them with real images to enhance data diversity.

[0036] 3) Data Division Divide the data set into training set, validation set and test set according to the ratio of 7:2:1 to ensure the generalization ability of the model.

[0037] Step 2.1, extracting multi-level feature maps of the input image by the encoder to generate a first-level feature map, a second-level feature map, a third-level feature map and a fourth-level feature map.

[0038] The encoder is a convolutional neural network (such as ResNet-101) for extracting multi-level feature maps from the input image, and the higher the level, the richer the semantic information. The first-level feature map (F1) retains the detailed edges, the second-level feature map (F2) integrates the local texture, the third-level feature map (F3) extracts the structural features, and the fourth-level feature map (F4) contains high-level semantics.

[0039] Input the normalized image into the pre-trained ResNet-101 encoder.

[0040] For example, in the four-level output feature maps: : resolution 56x56, channel number 256, capturing high-frequency details such as interlayer edges.

[0041] : resolution 28x28, channel number 512, integrating local texture information.

[0042] : resolution 14x14, channel number 1024, extracting interlayer structural features.

[0043] : resolution 7x7, channel number 2048, encoding high-level semantics such as interlayer overall distribution.

[0044] Through level progression, the model takes into account both the details of the slender structure and the global semantics.

[0045] Step 2.2, processing the fourth-level feature map using an adaptive dilated convolution mechanism to generate a dynamic dilated rate parameter and a multi-branch output feature.

[0046] The adaptive dilated convolution mechanism (AAC) dynamically adjusts the dilated rate (convolution kernel spacing) according to the local features of the image to optimize the receptive field size. The dynamic dilated rate parameter is generated by a learnable matrix, and the multi-branch output feature is a set of feature maps after applying different dilated rate convolutions, enhancing the adaptability to changes in interlayer curvature.

[0047] Step 2.2.1, performing a global average pooling operation on the input feature map to generate a local feature vector.

[0048] Performing a global average pooling operation on the input feature map: on ​​​​Global average pooling (GAP) is performed on the feature map (size 7×7×2048) to compress the spatial dimensions and generate a local feature vector (size 1×1×2048). This captures the global context of the feature map and avoids neglecting thin and long layers.

[0049] Step 2.2.2: Input the local feature vector into the learnable parameter matrix and generate the dynamic void ratio parameter through Softmax normalization.

[0050] Input the local feature vector into the learnable parameter matrix: The local feature vector is input into the learnable parameter matrix (size 2048×K, where K is the number of branches), normalized by the Softmax function, and outputs the dynamic dilation rate parameter ri (for example, ri∈[1, 6]). Softmax ensures that the parameter sums to 1, making the dilation rate adaptive to the input features.

[0051] Step 2.2.3: Adjust the receptive field of the convolution kernel according to the dynamic dilation rate parameter, perform a multi-branch dilation convolution operation on the input feature map, and generate a multi-branch output feature.

[0052] Adjust the convolution kernel receptive field according to the dynamic void rate parameter: A multi-branch atrous convolution operation is applied to the feature map, with each branch using an independent atrous ratio ri (e.g., branch 1: r = 2, branch 2: r = 4), generating a multi-branch output feature (size 7 × 7 × 512). For example, in areas with large interlayer curvature, ri is automatically increased to expand the receptive field and improve the ability to capture slender structures.

[0053] Step 2.3: The first-level feature map, the second-level feature map, the third-level feature map, and the multi-branch output features are fused through a cross-scale feature interaction module to generate a multi-scale fused feature map.

[0054] The Cross-Scale Feature Interaction (CSFI) module fuses feature maps from different levels (including multi-branch output features) to generate a multi-scale fused feature map. This module integrates shallow details and deep semantics through up / down sampling and splicing operations, addressing the problem of low-level feature loss.

[0055] Upsampling high-level features: The feature map is upsampled to 28×28 resolution, The multi-branch output feature maps are upsampled to 28×28 resolution and bilinear interpolation is used to maintain smoothness.

[0056] Downsampling low-level features: The feature map is downsampled to 28×28 resolution and average pooling is used to reduce noise.

[0057] Concatenate all feature maps: concatenate along the channel dimension (after downsampling), 、 (after upsampling), and multi-branch output features (after upsampling).

[0058] Perform 1×1 convolution fusion: Apply convolution kernels to fuse the concatenated features and generate a multi-scale fused feature map (size 28×28×1024). This preserves interlayer edge details while enhancing semantic consistency.

[0059] Step 2.4: Introduce a residual connection in the decoder to process the fourth-level feature map and the multi-scale fusion feature map to generate a high-resolution feature map.

[0060] Residual connections directly transfer high-level features in the decoder to avoid information loss. The high-resolution feature map is the output generated by upsampling the fused features, with a resolution similar to the input image (e.g., 56×56).

[0061] Will The feature map is upsampled to 28×28 resolution.

[0062] Perform residual addition: add the upsampled The features are added to the multi-scale fusion feature map, and the formula is expressed as: Output feature = ↑+Fusion feature map.

[0063] Upsampling through the decoder: Use transposed convolution to upsample the feature map to the original resolution (e.g., 224×224) to generate a high-resolution feature map, ensuring clear mezzanine segmentation boundaries.

[0064] Step 3: Construct a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate a multi-scale fusion feature map. The dynamic feature perception architecture includes an adaptive dilated convolution mechanism and a cross-scale feature interaction module.

[0065] The dynamic feature perception architecture is a neural network structure for enhanced image feature extraction. It includes an adaptive atrous convolution mechanism (AAC) and a cross-scale feature interaction module (CSFI). It processes normalized image data (pixel values ​​in the range [0, 1]) and enhanced image data (images after geometric and color transformations), generating a multi-scale fused feature map (a fusion of feature maps at different levels of detail and semantics). By dynamically adjusting the receptive field and cross-scale fusion, this architecture improves the recognition of weak interlayers (such as low-contrast cracks).

[0066] Adaptive Atrous Convolution (AAC) application: Input normalized image data or enhanced image data (resolution 224×224).

[0067] The AAC module is used to dynamically adjust the dilation rate (convolution kernel spacing) of the convolution kernel. For example, on the input feature map, the dilation rate is automatically increased or decreased according to the local sandwich curvature: the dilation rate is increased (e.g., r = 6) when the curvature is large to capture elongated structures, and the dilation rate is decreased (e.g., r = 2) when the curvature is small to preserve details.

[0068] The output multi-scale feature branch, each branch corresponding to a feature map of different receptive field (size 7x7x512), ensures that the model adapts to the sandwich shape change.

[0069] The cross-scale feature interaction module (CSFI) fuses: The multi-scale feature branch output by AAC and the low-level features of the encoder (e.g. ) are fused.

[0070] The resolution is unified to 28x28 by upsampling (high-level features) and downsampling (low-level features), and then all the features are spliced.

[0071] 1x1 convolution is applied to fuse the spliced features, generating a multi-scale fused feature map (size 28x28x1024). This feature map integrates shallow edge details (e.g., crack texture) and deep semantic information (e.g., sandwich distribution), solving the omission problem of traditional methods in low-contrast areas.

[0072] In one specific example, as shown in Figure 2 : The dynamic feature perception architecture adopts a structure combining adaptive atrous convolution (AAC) and cross-scale feature interaction (CSFI). The AAC module calculates the dilation rate parameter based on local features , with the formula , where is a learnable parameter, GAP is a global average pooling operation, and F is the input feature map. The CSFI module fuses shallow detail features and deep semantic features through upsampling, downsampling, and splicing operations.

[0073] AAC: Local feature extraction: on the input feature map , the local feature is extracted by global average pooling (GAP).

[0074] Dilation rate generation: input the learnable parameter (K is the number of output channels) to generate the dilation rate parameter The model can dynamically adjust the receptive field according to the local features of the input image, enhancing the sensitivity to the edges of weak interlayer.

[0075] Dynamic convolution operation: apply a convolution kernel with a dilation rate of to each branch of the input feature map F, extract multi-scale features, and improve the ability to capture slender structures (such as fault lines), output feature , the formula is .

[0076] Parameter optimization: automatically adjust through backpropagation, so that the model can adaptively adjust the dilation rate during training to adapt to the curvature changes of the interlayer.

[0077] The AAC module calculates the dilation rate parameter based on local features, avoiding the lack of adaptability of fixed dilation rate to changes in interlayer curvature. For example, in areas with large interlayer curvature, the dilation rate is automatically increased to enhance the ability to perceive slender structures.

[0078] CSFI: Feature pyramid construction: extract output feature maps at four levels of the encoder ResNet-101 .

[0079] Multi-scale fusion strategy: upsample high-level features and to resolution, downsample low-level features to resolution, and concatenate them with , then fuse them through 1x1 convolution with the formula . This multi-scale feature fusion preserves the edge information of slender structures while enhancing the semantic understanding of complex geological structures.

[0080] Residual connection: introduce a residual block in the decoder, upsample the high-level features of the encoder and directly pass them to the decoder to avoid information loss, improve the model's ability to model long-distance dependencies of weak interlayers, and the formula is , then generate the final output through pointwise convolution and upsampling.

[0081] Step 3.1, extract the channel features of the multi-scale fusion feature map, and generate channel attention weights through a fully connected layer.

[0082] Channel features are global statistical information of feature maps (such as the average activation value of each channel), and channel attention weights are weight vectors (weight values of each channel) generated through a fully connected layer, used to emphasize key channels (such as feature channels related to interlayers).

[0083] Channel feature extraction: Input multi-scale fused feature map (size 28x28x1024).

[0084] Perform global average pooling (GAP) operation: compute the average value of each channel on spatial dimensions (28x28), generate channel feature vector (size 1x1x1024). This compresses redundant information and preserves global context related to weak interlayer.

[0085] Channel attention weight generation: Input channel feature vector into fully connected layer (FC), which contains a learnable parameter matrix (size 1024x1024).

[0086] Normalize output by Sigmoid function, generate channel attention weight vector (size 1x1x1024). Weight values are in the range [0, 1], high weight indicates that the channel is more critical to interlayer identification (such as texture or edge channel). For example, in the area with dense cracks, the weight automatically enhances the relevant channel.

[0087] This process enhances the model's sensitivity to key features and improves its robustness to low-contrast interlayers.

[0088] Step 3.2, extract spatial features of channel weighted feature map, generate spatial attention weight through convolution layer.

[0089] Channel weighted feature map is the feature map after applying channel attention weight (size 28x28x1024), spatial feature is the spatial distribution information of feature map (such as position-related activation value), spatial attention weight is the weight matrix generated by convolution layer (weight value of each pixel), used to emphasize key spatial regions (such as interlayer edges).

[0090] Channel weighted feature map generation: Multiply channel attention weight vector with multi-scale fused feature map channel by channel to generate channel weighted feature map (size 28x28x1024). This highlights important channels and weakens noise channels (such as background regions).

[0091] Spatial feature extraction: Perform spatial pooling operation on channel weighted feature map: use max pooling to extract the maximum activation value at each position, and average pooling to extract the average activation value, generate two spatial feature matrices (size 28x28). This captures the distribution pattern of interlayers in spatial dimensions (such as crack extension direction).

[0092] Spatial attention weight generation: Concatenate spatial feature matrices of max pooling and average pooling outputs.

[0093] Input convolutional layer (kernel size 7x7) to generate initial spatial weight matrix.

[0094] Normalized by Sigmoid function, output spatial attention weight matrix (size 28x28). Weight values range from 0 to 1, high weight areas correspond to high probability positions of the interlayer (such as the centerline of the crack). For example, in the edge area of the interlayer, the weight value is close to 1, which enhances the segmentation accuracy.

[0095] Step 3.3. Generate interlayer geometric parameter map through pre-trained geological feature extractor and convert it into geological constraint factor.

[0096] The geological feature extractor is a pre-trained graph neural network (GNN), the interlayer geometric parameter map is a feature map containing interlayer geometric attributes (such as curvature and extension direction), and the geological constraint factor is a weight factor generated by normalization, which is used to integrate geological prior knowledge (such as interlayer morphology) into the attention mechanism.

[0097] Step 3.3.1. Extract interlayer curvature parameters and extension direction parameters from the input image to generate geological structure parameter tensor.

[0098] The interlayer curvature parameter describes the bending degree of the interlayer (the larger the value, the more obvious the bending), the extension direction parameter describes the strike of the interlayer (such as the angle value), and the geological structure parameter tensor is a feature map integrating these parameters (size 28x28xD, D is the parameter dimension).

[0099] Input the original image or the preprocessed image (resolution 224x224).

[0100] Use the pre-trained geological feature extractor (such as a GNN-based model): extract the interlayer curvature parameter (calculate the local curvature value, range 0~1) and the extension direction parameter (calculate the interlayer angle, range 0°~360°).

[0101] Output the geological structure parameter tensor (size 28x28x2), where the first channel is the curvature parameter and the second channel is the extension direction parameter. For example, in the curved crack area, the curvature value is greater than 0.5.

[0102] Step 3.3.2. Perform convolution operation on the geological structure parameter tensor to generate initial constraint factor.

[0103] The initial constraint factor is an intermediate weight matrix generated by convolution operation, which is used to preliminarily integrate geological parameters.

[0104] Input the geological structure parameter tensor (size 28x28x2).

[0105] Apply a convolution layer (kernel size 3x3) to generate an initial constraint factor matrix (size 28x28). The convolution operation integrates curvature and direction parameters to generate spatially dependent weight values (value range unrestricted). For example, in areas with consistent extension direction, the weight values are higher.

[0106] Step 3.3.3, Normalize the initial constraint factor by a Sigmoid function to generate a geological constraint factor.

[0107] The geological constraint factor is the normalized maximum weight matrix (size 28x28) with values in the range [0, 1] used to constrain attention weights.

[0108] Input the initial constraint factor matrix (size 28x28).

[0109] Apply the Sigmoid function for normalization to output the geological constraint factor matrix (size 28x28). High factor value areas reinforce geological priors (such as when the factor value is close to 1 when the curvature is large), ensuring that the model generalizes in data-scarce scenarios.

[0110] Step 3.4, Fuse the spatial attention weight and the geological constraint factor to generate an attention-weighted feature map.

[0111] The attention-weighted feature map is a feature map that applies fusion weights, fusing spatial attention weights (emphasizing key positions) and geological constraint factors (integrating geological priors), to improve the identification of complex interlayers.

[0112] Weight fusion: Input the spatial attention weight matrix (size 28x28) and the geological constraint factor matrix (size 28x28).

[0113] Perform an element-wise multiplication operation: generate a fusion weight matrix (size 28x28). This operation ensures that spatial attention is constrained by geological parameters (such as only enhancing weights in areas with consistent extension direction).

[0114] Feature weighting: multiply the fusion weight matrix with the channel-weighted feature map (size 28x28x1024) position by position to generate an attention-weighted feature map (size 28x28x1024).

[0115] Output retains key features (such as fracture edges), while suppressing irrelevant areas (such as background noise). For example, in low-contrast interlayer areas, the weighted feature values are significantly enhanced.

[0116] Step 4, Process the multi-scale fusion feature map through a geological prior guided attention network to generate an attention-weighted feature map. The geological prior guided attention network fuses channel attention weights, spatial attention weights, and geological structure parameter constraint factors.

[0117] The geology prior guided attention network (MAF+GFGA) is a multi-modal framework that combines channel attention weights, spatial attention weights, and geology structure parameter constraint factors to generate the final attention weighted feature map. This network uses geology knowledge to guide attention and improve model accuracy in data-scarce scenarios.

[0118] The network structure should: Input multi-scale fused feature map (size 28x28x1024).

[0119] Perform steps 3.1-3.4 in the complete process: first generate channel attention weights, then generate spatial attention weights, then generate geology constraint factors, and finally fuse the output attention weighted feature map.

[0120] Geology prior integration: In the fusion stage, the geology constraint factor (generated from step 3.3) dynamically adjusts the spatial attention weight (generated from step 3.2) to ensure that the attention distribution conforms to the characteristics of the interlayer geology (such as weight enhancement when curvature is large).

[0121] Output attention weighted feature map (size 28x28x1024), which strengthens the interlayer related areas (such as crack boundaries) and weakens the interference areas.

[0122] Step 5, process the attention weighted feature map based on the preset encoder-decoder structure to generate the initial segmentation prediction map.

[0123] The preset encoder-decoder structure is a neural network framework (such as the U-Net architecture), and the initial segmentation prediction map is the preliminary output of the model (size 224x224), which contains pixel-level prediction probabilities of the interlayer (values in the range [0, 1]).

[0124] Decoder structure processing: Input attention weighted feature map (size 28x28x1024).

[0125] Use transpose convolution layers to gradually upsample the feature map: first upsample to 56x56 resolution, then to 112x112, and finally to 224x224 resolution (same as the input image).

[0126] Introduce residual connections in the decoder: upsample the encoder high-level features (such as ) and add them to the current features to avoid information loss (formula: output feature = upsampled feature + decoded feature).

[0127] Prediction map generation: Output segmentation prediction with 1x1 convolutional layer: Generate initial segmentation prediction map (size 224x224). Each pixel value represents the probability of interlayer presence (0 for background, 1 for interlayer).

[0128] This output is used for subsequent loss calculation and optimization (step 6). For example, in the interlayer edge area, the prediction value is close to 1, ensuring clear boundaries.

[0129] In one specific example, a multimodal attention framework (MAF) is constructed, including channel attention (SE Block) and spatial attention (CBAM), and a constraint factor γ is generated through geological structure parameters (such as interlayer curvature, extension direction).

[0130] MAF: Channel attention: Channel weight generation: Perform global average pooling on the feature map to obtain channel features , and generate channel weights through a fully connected layer.

[0131] Feature weighting: Multiply α and F to enhance key channel features, .

[0132] Spatial attention: Spatial feature extraction: Perform max pooling and average pooling on to obtain spatial features .

[0133] Spatial weight generation: Generate spatial weights through a convolutional layer.

[0134] Feature weighting: Multiply β and to enhance key spatial regions, .

[0135] Geological structure parameter generation and constraint (GFGA): Geological feature extraction: Generate interlayer geometric parameters through a pre-trained geological feature extractor (GNN), where D is the parameter dimension (such as curvature, extension direction).

[0136] Constraint factor calculation: Input G into a convolutional layer to generate a constraint factor , which is used to constrain the attention weight distribution and improve the recognition ability of complex geological structures.

[0137] Final attention mask: Multiply the traditional multimodal attention result with the constraint factor γ to generate the final feature .

[0138] By introducing the interlayer curvature and extension direction parameters, the attention weight distribution is constrained, and the recognition ability for low-contrast areas is improved. For example, in the area where the interlayer extension direction is consistent with the main direction of the image, the attention weight is significantly enhanced.

[0139] Step 6, calculate the difference between the initial segmentation prediction map and the real label through the adaptive loss function module, and generate optimization gradient data.

[0140] The adaptive loss function module is a neural network component that quantifies the error between the model prediction (initial segmentation prediction map) and the real label (artificially labeled interlayer area), and generates optimization gradient data (gradient vector) for backpropagation to update the model parameters. This module solves the class imbalance (such as small interlayer pixel ratio) and boundary ambiguity problems in weak interlayer segmentation through dynamic weight distribution and edge enhancement.

[0141] Input data preparation: Initial segmentation prediction map: prediction probability map (size 224x224) generated from step 5, each pixel value represents the probability of interlayer existence (0 for background, 1 for interlayer).

[0142] Real label: artificially labeled binary mask map (size 224x224), soft interlayer area (such as cracks or faults) labeled by geologists.

[0143] These inputs ensure that error calculation is based on pixel-level differences.

[0144] Step 6.1, calculate the Dice loss value and Focal loss value of the initial segmentation prediction map and the real label.

[0145] The Dice loss value measures the overlap between the prediction and the real label (suitable for class imbalance scenarios), and the Focal loss value enhances the punishment for difficult samples (such as small target interlayers), avoiding model neglect of minority classes.

[0146] Dice loss calculation: Based on pixel-level comparison, the formula is: .

[0147] where is the prediction probability, is the real label, is the pixel index.

[0148] This emphasizes the overlap of the interlayer area and alleviates the problem of background pixel dominance (for example, in sparse interlayer areas, the loss value is higher to enhance learning).

[0149] Focal loss calculation: To alleviate the class imbalance issue and enhance the optimization of small targets (e.g., laminations), the formula is where α is the class weight, set to 0.25 (to balance positive and negative samples), and γ is the focal factor, usually set to 2 (to enhance the penalty on hard samples).

[0150] For example, when the predicted probability is close to 0 but the true label is 1 (indicating that the model incorrectly ignores laminations), the loss value significantly increases, enhancing sensitivity to small targets.

[0151] This step outputs two loss values, which are used for subsequent dynamic weight allocation.

[0152] Step 6.2, dynamically allocate Dice loss weight and Focal loss weight according to the lamination density parameter, and generate a weighted joint loss value.

[0153] The lamination density parameter represents the density of lamination pixels per unit area (the higher the value, the more dense the laminations), and the dynamic allocation generates weights through the Sigmoid function to ensure that the model adapts to different data regions (e.g., focusing on Dice loss in high-density areas and Focal loss in low-density areas). The weighted joint loss value is the total loss after fusing the weights, optimizing training stability.

[0154] Step 6.2.1, count the number of lamination pixels per unit area to generate the lamination density parameter.

[0155] The lamination density parameter is a scalar value that quantifies the density of lamination distribution, calculated as the total number of lamination pixels in the true label divided by the total area of the image.

[0156] Input the true label map (size 224x224).

[0157] Count the number of lamination pixels (the number of pixels with a true label value of 1).

[0158] Calculate the density parameter ρ: ρ = (number of lamination pixels) / (224x224). For example, in a sparse lamination area, ρ may be less than 0.1; in a dense area, ρ may be greater than 0.3.

[0159] Step 6.2.2, input the lamination density parameter into the Sigmoid function to generate the Dice loss dynamic weight.

[0160] The Dice loss dynamic weight is the weight value (range 0~1) output by the Sigmoid function, used to adjust the contribution proportion of Dice loss in the total loss, and the Sigmoid function ensures smooth transition of the weight.

[0161] Input the lamination density parameter ρ.

[0162] Step 6.2.3. Calculate the Focal loss dynamic weight according to the Dice loss dynamic weight.

[0163] The Focal loss dynamic weight is a complementary weight, calculated as 1 minus the Dice loss weight, ensuring the total loss weight sum is 1.

[0164] For example, in low density areas (ρlow), wFocal approaches 1, strengthening the Focal loss to handle difficult samples.

[0165] After outputting the dynamic weight, generate the weighted joint loss value: dynamic weight and Calculate by the interlayer density ρ (the number of interlayers per unit area), the formula is where k is the coefficient controlling the gradient, usually set to 5 (adjusting the smoothness of weight changes), is the interlayer density threshold for dynamic weight calculation; , improving the adaptability of the model to different density areas.

[0166] Step 6.3. Extract the edge mask of the real label through the edge detection algorithm, and calculate the edge-aware loss value.

[0167] The edge-aware loss value strengthens the segmentation accuracy of the interlayer boundary, generates a binary edge mask (only containing edge pixels) through Canny edge detection, and calculates the cross-entropy loss.

[0168] Step 6.3.1. Perform Canny edge detection operation on the real label to generate a binary edge mask.

[0169] The binary edge mask is an image (size 224x224) that identifies the interlayer boundary, with edge pixel values of 1 and non-edge values of 0. The Canny algorithm uses high and low threshold values to detect edges.

[0170] Input the real label image.

[0171] Apply Canny edge detection: first Gaussian filter for denoising, then calculate gradient amplitude and direction, and finally use double threshold (low threshold 0.1, high threshold 0.3) to generate a binary mask.

[0172] Output the mask image, for example, set the pixel value to 1 at the crack boundary.

[0173] Step 6.3.2. Extract the pixel prediction value of the edge region in the initial segmentation prediction image to generate an edge prediction subset.

[0174] The edge prediction subset is a set of pixel prediction values in the initial segmentation prediction image corresponding to the edge mask position.

[0175] Input initial segmentation prediction map and binary edge mask.

[0176] Traverse the mask map, when the mask value is 1, extract the prediction value pi in the same position of the prediction map.

[0177] Generate subset data, for example, prediction probability only containing boundary pixels.

[0178] Step 6.3.3, calculate the cross-entropy loss of the edge prediction subset and the real edge mask, generate the edge-aware loss value.

[0179] Edge extraction: Canny edge detection on the real label Y, strengthen the segmentation accuracy of the interlayer edge, generate the edge mask .

[0180] Edge loss calculation: add edge weight in cross-entropy loss , the formula is , where is the label and prediction value of the edge region, is the edge weight coefficient, usually set to 0.7 (balance edge and overall segmentation accuracy), is the real label and prediction value of the overall region.

[0181] Step 6.4, fuse the weighted joint loss value and the edge-aware loss value to generate the optimization gradient data.

[0182] Optimization gradient data is the gradient vector of the loss function, used for back propagation to update the model weight. Fusion is achieved by weighted sum.

[0183] Input weighted joint loss value and edge-aware loss value.

[0184] Calculate total loss: total loss = weighted joint loss + edge-aware loss.

[0185] Calculate the gradient by automatic differentiation (such as PyTorch or TensorFlow backward function) to generate the optimization gradient data.

[0186] For example, the gradient data includes the partial derivative of each model parameter, guiding the parameter update direction.

[0187] Step 7, based on the optimization gradient data, iteratively update the model parameters to generate the final semantic segmentation model.

[0188] Iterative update is to adjust the model weight by repeatedly applying the gradient data through the optimizer (such as AdamW); the final semantic segmentation model is the trained network that can accurately segment the weak interlayer.

[0189] Optimizer configuration: Using AdamW optimizer (combining adaptive learning rate and weight decay), initial learning rate set to , weight decay coefficient .

[0190] Iterative update: Input optimization gradient data (generated from step 6).

[0191] Perform backpropagation: calculate the gradient of model parameters (such as convolution kernel weights).

[0192] Apply gradient descent: update parameter formula parameter = parameter - learning rate × gradient.

[0193] Use learning rate scheduling: cosine annealing strategy, learning rate reduced by 10% every 5 epochs to avoid local optimum.

[0194] Training termination: Set early stopping mechanism (EarlyStopping) to monitor validation set loss (stop when there is no improvement for 10 consecutive epochs).

[0195] After training for 100 epochs, output the final semantic segmentation model (saved as.pt or.h5 file).

[0196] This ensures that the model generalizes in data-scarce scenarios, such as mIoU improvement on low-contrast images.

[0197] Step 8, based on the final semantic segmentation model, perform segmentation operation on the target soft interlayer image to generate interlayer segmentation result map.

[0198] The target soft interlayer image is a new image to be segmented (such as geological exploration or engineering detection image); the segmentation operation is the model inference process; the interlayer segmentation result map is a binary output map (same size as input), which identifies the interlayer region.

[0199] Input preparation: Get the target image (such as a satellite image of a certain mining area or a tunnel site image) and perform the same preprocessing as step 2 (normalize to [0, 1]).

[0200] Model inference: Load the final semantic segmentation model.

[0201] Input image to model: processed through encoder-decoder structure, output prediction probability map.

[0202] Result generation: Apply threshold (usually 0.5) to the prediction probability map: pixels with probability ≥ 0.5 are set to 1 (interlayer), otherwise 0 (background).

[0203] The interlayer segmentation result map (224x224 in size) can be directly visualized or used for risk assessment.

[0204] For example, the method is executed on an embedded device, and supports real-time applications such as engineering monitoring.

[0205] The above is a method embodiment of the present application. Based on the same inventive concept, the embodiments of the present application also provide a deep learning-based weak interlayer semantic segmentation device, the structure of which is shown in Figure 2 .

[0206] Figure 3 A deep learning-based weak interlayer semantic segmentation device provided by an embodiment of the present application is shown in Figure 3 . The device includes: at least one processor 301; and a memory 302 in communication connection with the at least one processor; The memory 302 stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor 301 to enable the at least one processor 301 to: obtain weak interlayer image data, the weak interlayer image data including geological exploration images, engineering detection images, and synthetic data; pre-process the weak interlayer image data to generate normalized image data and enhanced image data; construct a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate a multi-scale fusion feature map; the dynamic feature perception architecture includes an adaptive atrous convolution mechanism and a cross-scale feature interaction module; process the multi-scale fusion feature map through a geology-prior-guided attention network to generate an attention-weighted feature map; the geology-prior-guided attention network fuses channel attention weights, spatial attention weights, and geologic structure parameter constraint factors; process the attention-weighted feature map based on a preset encoder-decoder structure to generate an initial segmentation prediction map; calculate the difference between the initial segmentation prediction map and a real label through an adaptive loss function module to generate optimization gradient data; iteratively update model parameters based on the optimization gradient data to generate a final semantic segmentation model; and perform a segmentation operation on a target weak interlayer image based on the final semantic segmentation model to generate an interlayer segmentation result map.

[0207] Some embodiments of the present application provide a non-volatile computer storage medium for deep learning-based weak interlayer semantic segmentation corresponding to Figure 1 , which stores computer executable instructions, and the computer executable instructions are configured to: The soft interlayer image data includes geological exploration images, engineering detection images and synthetic data; the soft interlayer image data is preprocessed to generate normalized image data and enhanced image data; a dynamic feature perception architecture is constructed to process the normalized image data and the enhanced image data to generate multi-scale fusion feature maps; the dynamic feature perception architecture includes an adaptive hollow convolution mechanism and a cross-scale feature interaction module; an attention network guided by geological prior is used to process the multi-scale fusion feature maps to generate attention weighted feature maps; the attention network guided by the geological prior fuses channel attention weights, spatial attention weights and geological structure parameter constraint factors; a preset encoder-decoder structure is used to process the attention weighted feature maps to generate an initial segmentation prediction map; an adaptive loss function module is used to calculate the difference between the initial segmentation prediction map and a real label to generate optimization gradient data; model parameters are iteratively updated based on the optimization gradient data to generate a final semantic segmentation model; and the final semantic segmentation model is used to perform a segmentation operation on a target soft interlayer image to generate an interlayer segmentation result map.

[0208] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments focuses on the difference from other embodiments. In particular, the IoT device and medium embodiments are basically similar to the method embodiments, and thus are described simply. The relevant parts can be referred to the description of the method embodiments.

[0209] The system and medium provided by the embodiments of the present application are one-to-one corresponding to the method, and thus the system and medium have similar beneficial technical effects to the method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the system and medium will not be described here.

[0210] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0211] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof.

[0212] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof.

[0213] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof.

[0214] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0215] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, non-volatile memory, such as read-only memory (ROM), EPROM, and / or flash memory, etc. The memory is an example of computer-readable media.

[0216] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can implement information storage by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0217] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0218] The above only describes the embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A weak interlayer semantic segmentation method based on deep learning, characterized by: The method comprises: Acquiring weak interlayer image data, wherein the weak interlayer image data includes geological exploration images, engineering detection images, and synthetic data; Preprocessing the soft interlayer image data to generate normalized image data and enhanced image data; Constructing a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate a multi-scale fusion feature map; the dynamic feature perception architecture includes an adaptive hole convolution mechanism and a cross-scale feature interaction module; Processing the multi-scale fusion feature map through a geological prior-guided attention network to generate an attention-weighted feature map; the geological prior-guided attention network fuses channel attention weights, spatial attention weights, and geological structure parameter constraint factors; Processing the attention-weighted feature map based on a preset encoder-decoder structure to generate an initial segmentation prediction map; Calculate the difference between the initial segmentation prediction map and the true label through the adaptive loss function module to generate optimized gradient data; Iteratively update model parameters based on the optimized gradient data to generate a final semantic segmentation model; A segmentation operation is performed on the target soft interlayer image based on the final semantic segmentation model to generate an interlayer segmentation result image.

2. The method for semantic segmentation of weak interlayer based on deep learning according to claim 1, characterized in that: Preprocessing the soft interlayer image data to generate normalized image data and enhanced image data specifically includes: Extracting multi-level feature maps of the input image through the encoder to generate a first-level feature map, a second-level feature map, a third-level feature map, and a fourth-level feature map; Adopting an adaptive dilated convolution mechanism to process the fourth-level feature map to generate dynamic dilation rate parameters and multi-branch output features; The first-level feature map, the second-level feature map, the third-level feature map and the multi-branch output features are fused through a cross-scale feature interaction module to generate a multi-scale fused feature map; A residual connection is introduced in the decoder to process the fourth-level feature map and the multi-scale fusion feature map to generate a high-resolution feature map.

3. The method for semantic segmentation of weak interlayer based on deep learning according to claim 2, characterized in that: Adopting an adaptive dilated convolution mechanism to process the fourth-level feature map, generating dynamic dilation rate parameters and multi-branch output features, specifically including: Perform a global average pooling operation on the input feature map to generate a local feature vector; Input the local feature vector into the learnable parameter matrix and generate a dynamic void rate parameter through Softmax normalization; The receptive field of the convolution kernel is adjusted according to the dynamic dilation rate parameter, and a multi-branch dilation convolution operation is performed on the input feature map to generate multi-branch output features.

4. The method for semantic segmentation of weak interlayer based on deep learning according to claim 1, characterized in that: Constructing a dynamic feature perception architecture to process the normalized image data and enhanced image data to generate a multi-scale fusion feature map, specifically including: Extracting channel features of the multi-scale fusion feature map and generating channel attention weights through a fully connected layer; Extract the spatial features of the channel weighted feature map and generate spatial attention weights through the convolution layer; Generate interlayer geometric parameter maps through pre-trained geological feature extractors and convert them into geological constraint factors; The spatial attention weights are fused with the geological constraint factors to generate an attention-weighted feature map.

5. The method for semantic segmentation of weak interlayer based on deep learning according to claim 4, characterized in that: The interlayer geometry parameter map is generated by the pre-trained geological feature extractor and converted into geological constraint factors, including: Extract interlayer curvature parameters and extension direction parameters from the input image to generate geological structure parameter tensor; Performing a convolution operation on the geological structure parameter tensor to generate an initial constraint factor; The initial constraint factor is normalized by a Sigmoid function to generate a geological constraint factor.

6. The method for semantic segmentation of weak interlayer based on deep learning according to claim 1, characterized in that: The adaptive loss function module calculates the difference between the initial segmentation prediction map and the true label to generate optimized gradient data, specifically including: Calculate the Dice loss and Focal loss of the initial segmentation prediction map and the true label; Dynamically allocate Dice loss weight and Focal loss weight according to the interlayer density parameter to generate a weighted joint loss value; Extract the edge mask of the true label through the edge detection algorithm and calculate the edge perception loss value; The weighted joint loss value and the edge-aware loss value are fused to generate optimized gradient data.

7. The method for semantic segmentation of weak interlayer based on deep learning according to claim 6, characterized in that: The Dice loss weight and Focal loss weight are dynamically allocated according to the interlayer density parameter to generate a weighted joint loss value, specifically including: Count the number of interlayer pixels per unit area and generate interlayer density parameters; Input the interlayer density parameter into the Sigmoid function to generate a Dice loss dynamic weight; The Focal loss dynamic weight is calculated according to the Dice loss dynamic weight.

8. The method for semantic segmentation of weak interlayer based on deep learning according to claim 6, characterized in that: The edge detection algorithm is used to extract the edge mask of the true label and calculate the edge perception loss value, which includes: Perform Canny edge detection on the true label to generate a binary edge mask; Extract the pixel prediction values ​​of the edge area in the initial segmentation prediction image to generate an edge prediction subset; The cross entropy loss between the edge prediction subset and the true edge mask is calculated to generate an edge-aware loss value.

9. A weak interlayer semantic segmentation device based on deep learning, characterized in that: The device comprises: at least one processor; and, a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: Acquiring weak interlayer image data, wherein the weak interlayer image data includes geological exploration images, engineering detection images, and synthetic data; Preprocessing the soft interlayer image data to generate normalized image data and enhanced image data; Constructing a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate a multi-scale fusion feature map; the dynamic feature perception architecture includes an adaptive hole convolution mechanism and a cross-scale feature interaction module; Processing the multi-scale fusion feature map through a geological prior-guided attention network to generate an attention-weighted feature map; the geological prior-guided attention network fuses channel attention weights, spatial attention weights, and geological structure parameter constraint factors; Processing the attention-weighted feature map based on a preset encoder-decoder structure to generate an initial segmentation prediction map; Calculate the difference between the initial segmentation prediction map and the true label through the adaptive loss function module to generate optimized gradient data; Iteratively update model parameters based on the optimized gradient data to generate a final semantic segmentation model; A segmentation operation is performed on the target soft interlayer image based on the final semantic segmentation model to generate an interlayer segmentation result image.

10. A non-volatile computer storage medium for deep learning-based weak interlayer semantic segmentation, storing computer-executable instructions, characterized in that: The computer executable instructions are configured to: Acquiring weak interlayer image data, wherein the weak interlayer image data includes geological exploration images, engineering detection images, and synthetic data; Preprocessing the soft interlayer image data to generate normalized image data and enhanced image data; Constructing a dynamic feature perception architecture to process the normalized image data and the enhanced image data to generate a multi-scale fusion feature map; the dynamic feature perception architecture includes an adaptive hole convolution mechanism and a cross-scale feature interaction module; Processing the multi-scale fusion feature map through a geological prior-guided attention network to generate an attention-weighted feature map; The geological prior-guided attention network integrates channel attention weights, spatial attention weights, and geological structure parameter constraint factors; Processing the attention-weighted feature map based on a preset encoder-decoder structure to generate an initial segmentation prediction map; Calculate the difference between the initial segmentation prediction map and the true label through the adaptive loss function module to generate optimized gradient data; Iteratively update model parameters based on the optimized gradient data to generate a final semantic segmentation model; A segmentation operation is performed on the target soft interlayer image based on the final semantic segmentation model to generate an interlayer segmentation result image.

Images