Intelligent identification system and method for true and false cracks based on deep multi-network feature fusion

Abstract

This invention discloses an intelligent system and method for identifying true and false fractures based on deep multi-network feature fusion, belonging to the fields of geological exploration and artificial intelligence. It solves the problems of low identification efficiency and difficulty in distinguishing true and false fractures in complex backgrounds in existing systems. The system includes a data preprocessing module and a deep multi-network intelligent identification model. The preprocessing module standardizes core images and crops them to fit the image blocks. The identification model is based on an encoder-decoder architecture, integrating CNN-Transformer dual-branch encoding units and PMCA decoding units to achieve complementary local and global features and enhanced feature space perception. The identification method involves five steps: data preprocessing, model construction and training, fracture identification and segmentation, and result post-processing analysis, outputting pixel-level segmentation maps and statistically analyzing fracture parameters. This invention achieves automated and high-precision identification of true and false fractures. The model's performance outperforms mainstream segmentation networks, and its lightweight design balances accuracy and efficiency, providing reliable data support for geological exploration work such as oil and gas reservoir evaluation.

Description

Technical Field

[0001] This invention belongs to the fields of geological exploration engineering technology, computer vision and artificial intelligence technology, and in particular relates to a system and method for intelligent identification of real and fake cracks based on deep multi-network feature fusion. Background Technology

[0002] Core samples are the most direct and reliable data for understanding underground geological conditions. The characteristics of fracture development within the core samples are key to evaluating the permeability of oil and gas reservoirs, assessing wellbore stability, and formulating development plans. Fractures can be classified according to their origin into "true fractures" (naturally formed by geological tectonic movements, diagenesis, etc.) and "false fractures" (artificially created during core sampling, transportation, and preparation). Accurately identifying and distinguishing between these two types of fractures is crucial for the accuracy of geological interpretation.

[0003] Existing core fracture identification technologies suffer from three core defects, and there are currently no effective solutions: Manual identification methods: Traditional identification relies on the manual observation and description of geological experts, which has the disadvantages of low efficiency, strong subjectivity, poor consistency, and difficulty in processing massive amounts of core image data; Traditional image processing methods, such as automatic recognition methods based on edge detection (Sobel operator, Canny detection), threshold segmentation, and morphological operations, are sensitive to image quality and lighting conditions, have poor robustness under complex backgrounds and texture interference, and cannot distinguish between real and fake cracks with similar shapes and grayscale features. Existing deep learning methods, such as convolutional neural network (CNN) models like U-Net and DeepLab, have been applied to crack segmentation. However, CNNs have limited receptive fields, making it difficult to model long-distance contextual dependencies. They are not adaptable to multi-scale cracks and lack the ability to enhance spatially perceptual channel features, making them susceptible to false positives / false negatives due to interference from the core mineral background. While Transformer models possess a global self-attention mechanism, they have high computational complexity and weak ability to capture local details. The feature fusion strategies and detail recovery mechanisms of existing CNN-Transformer hybrid models still have room for optimization for the task characteristics of core cracks, which are characterized by "complex backgrounds, small targets, and subtle inter-class differences."

[0004] Therefore, developing an intelligent system capable of accurately and automatically identifying and distinguishing between real and false cracks in core images, and addressing the aforementioned shortcomings of existing technologies, has significant theoretical and engineering application value. Summary of the Invention

[0005] To address the shortcomings of existing core fracture identification methods, such as low efficiency of manual interpretation and insufficient ability of existing automatic identification models to distinguish between real and fake fractures in multi-scale and complex backgrounds, this invention provides an intelligent identification system and method for real and fake fractures based on deep multi-network feature fusion.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A real and fake crack intelligent identification system based on deep multi-network feature fusion includes a data preprocessing module and a deep multi-network intelligent identification model. The data preprocessing module is used to perform standardization and sliding window cropping on the input core images in sequence. The standardization process includes color normalization, contrast enhancement, noise reduction and size regularization. The sliding window cropping strategy is used to generate image patches that are adapted to the model input. The deep multi-network intelligent recognition model is based on an encoder-decoder architecture and is used to extract features and perform semantic segmentation on preprocessed core images, outputting pixel-level classification maps.

[0007] Furthermore, the deep multi-network intelligent recognition model includes a CNN-Transformer dual-branch encoding unit, a partition merge channel attention (PMCA) decoding unit, and a result output module; The CNN-Transformer dual-branch coding unit is used to fuse local detail features with global semantic information; The Partition Merge Channel Attention (PMCA) decoding unit is integrated at the decoder skip connection and is used to perform spatially aware weight calibration and enhancement of encoded features; The results output module is used to post-process the segmentation results output by the model, including contour optimization, result visualization, and to generate an analysis report containing information on the distribution of real and fake cracks.

[0008] Furthermore, the CNN-Transformer dual-branch coding unit includes a CNN branch, a Transformer branch, and a cross-branch feature fusion module: The CNN branch is based on an improved ResNet architecture and contains multiple convolutional modules with embedded SEM-B bottleneck units, which are used to extract local texture and edge features from core images in layers. The bottleneck unit integrates standard convolution, depthwise separable convolution, and dilated depthwise separable convolution, which achieves lightweight model while ensuring feature extraction capabilities. The Transformer branch is based on the SegFormer architecture and consists of an overlapping patch embedding module and multiple Transformer blocks, used to model long-range dependencies and global context of an image. The cross-branch feature fusion module is used to dynamically inject the multi-scale global features output by the Transformer branch into the corresponding layer features of the CNN branch after adaptive adjustment, so as to achieve complementary enhancement of local features and global semantics.

[0009] Furthermore, the core of the Partition Merging Channel Attention (PMCA) decoding unit lies in combining the channel attention mechanism with the spatial partitioning strategy, and its working process includes the following steps: (1) Spatial partitioning: The input feature map is evenly divided into K non-overlapping rectangular sub-regions, and global average pooling is performed independently on each region to obtain its representation vector; (2) Partition pooling and weight learning: A channel attention weight vector is generated for each region through a multilayer perceptron (MLP) with shared weights; (3) Weighted merging and feature enhancement: The pooling features of each sub-region are multiplied by their corresponding weights and summed to obtain a merged channel weight vector. This weight vector is then multiplied by the original input feature map channel by channel to output the enhanced feature map, thereby achieving differentiated enhancement of key feature channels in different spatial regions.

[0010] Furthermore, K in step (1) is determined through experimental optimization. When K=4, the model achieves the optimal balance between the average intersection-union ratio (mIoU) and the recall rate.

[0011] Furthermore, the training process of the deep multi-network intelligent recognition model employs a combined loss function, which is a weighted sum of cross-entropy loss and Dice loss, and its calculation formula is: L total =αL CE +βL Dice , where α and β are both non-negative adjustable weighting coefficients.

[0012] A method for intelligent identification of real and fake cracks based on deep multi-network feature fusion is implemented using a core image-based intelligent crack identification system, including the following steps: S1. Data Preparation and Preprocessing: Collect and label core photos, and construct a dataset containing real cracks, false cracks, and background categories; standardize and enhance the images, and divide the dataset into training, validation, and test sets; S2. Model Construction: Construct the deep multi-network intelligent recognition model described above; S3. Model Training: Train the model using the training set data, monitor the training process using the validation set and adjust the hyperparameters until the model converges. S4. Crack Identification and Segmentation: Input the core photo to be identified into the trained model, and the model automatically outputs a pixel-level segmentation map, in which each pixel is classified as background, real crack, or fake crack. S5. Post-processing and analysis of results: The result output module performs thresholding and post-processing on the probability map to generate the final segmentation mask and analysis results. Then, it performs morphological operations on the segmentation map output by the model to optimize the boundary, statistically analyzes the length, area, and density parameters of the cracks, and generates a visualization report.

[0013] Furthermore, in step S1, a sliding window cropping method is used to generate training samples from high-resolution core images. The window size is set to 256×256 pixels and the step size is 128 pixels to increase the number of samples and ensure the integrity of the fracture area.

[0014] Furthermore, in step S3, the AdamW optimizer is used in conjunction with a cosine annealing learning rate scheduling strategy for model training, with a total of 50-200 epochs for training.

[0015] Further, step S4 is as follows: In the encoding stage, the image extracts features through both CNN and Transformer branches, and achieves complementary advantages through a cross-branch fusion mechanism; In the decoding stage, after upsampling, the deep features are fused step by step with the corresponding shallow high-resolution features enhanced by the PMCA module to gradually restore spatial details. Finally, a 1×1 convolutional layer is used as the core execution unit to output a probability map of each pixel belonging to the background, real crack, or fake crack.

[0016] By adopting the above technical solution, the beneficial effects of the present invention are as follows: 1) The CNN-Transformer dual-branch coding network of the present invention captures both the fine texture and macroscopic morphological structure of cracks through a cross-branch feature fusion mechanism, which significantly improves the model's feature representation ability for multi-scale and complex cracks and solves the problem of imbalance in local / global feature capture in existing models. 2) The PMCA mechanism designed in this invention breaks through the limitation of traditional channel attention ignoring spatial information. By feature map partitioning and regional weight learning, it strengthens crack-related feature channels and suppresses interference from complex backgrounds in the core, thereby improving recall while ensuring recognition accuracy and solving the problem of weak background interference resistance of existing models. 3) This invention realizes automated and high-precision intelligent identification of true and false cracks in core images, which greatly improves the efficiency and objectivity of geological interpretation work and frees up manual labor. 4) Experimental verification shows that the model of this invention significantly outperforms mainstream segmentation networks such as U-Net, DeepLabv3+, and SegFormer in key metrics such as mean intersection-union ratio (mIoU), precision, and recall. Moreover, the lightweight design of the model achieves a good balance between accuracy and inference efficiency. 5) The identification results of this invention can accurately count parameters such as the length, area, and density of fractures, providing reliable data support for oil and gas reservoir evaluation, geostress analysis, wellbore stability assessment, and drilling engineering design, and have important engineering application value. Attached Figure Description

[0017] Figure 1 : Schematic diagram of the core image preprocessing stage analysis; including 1.1 original core image (RGB three channels, 800x652 pixels, used as the original input data for the model), 1.2 enhanced grayscale image (contrast enhanced by histogram equalization to improve crack visibility), 1.3 grayscale histogram analysis (the horizontal axis is the intensity of 0-255 pixels, and the vertical axis is the number of pixels, used to analyze image contrast and brightness distribution); Figure 2 A comparative analysis diagram of traditional edge detection methods; including 2.1 Sobel operator edge detection (first derivative operator, detects horizontal / vertical edges, fast calculation speed), 2.2 Canny edge detection (multi-stage algorithm, good detection effect but parameter sensitive), 2.3 adaptive threshold processing (adaptive Gaussian threshold, robust to illumination changes), and 2.4 comprehensive edge detection results (combination of multiple methods, connecting edges through opening and closing operations, used as input for traditional crack recognition). Figure 3 The system architecture diagram of this invention is shown in sequence, which includes the input core image, data preprocessing module, CNN-Transformer dual-branch encoding network, partition merge channel attention (PMCA) decoder, and result postprocessing module, and finally outputs a true / false fracture distribution map and analysis report. Figure 4 : Schematic diagram of CNN-Transformer dual-branch coding network structure; showing the overall structure of the input image (256×256×3) through the CNN module (embedded with SEM-B bottleneck structure, containing residual connections, outputting F1-F8 multi-scale features) and the Transformer module (containing Patch embedding, multi-head attention, feedforward layer, outputting corresponding scale features), achieving complementary feature enhancement through cross-branch fusion; Figure 5 : SEM-B bottleneck unit structure diagram; showing the parallel dual-branch structure (branch 1: 3×3 depth separable convolution, branch 2: 3×3 dilated depth separable convolution, dilation rate = 2), after feature concatenation and 1×1 convolution channel dimensionality reduction, the features are output through residual connection, reflecting the design features of expanding receptive field, lightweighting, and avoiding gradient vanishing. Figure 6 : A schematic diagram of the PMCA mechanism structure and principle; showing the core process of outputting an enhanced feature map after the input feature map is spatially partitioned, global average pooling, MLP weight learning, and weighted merging; Figure 7 A schematic diagram of the integration of a UNet-like decoder and PMCA; illustrating the correspondence between encoder downsampling and decoder upsampling. The PMCA module is integrated at the jump connections of each level, realizing the fusion of encoder feature enhancement with upsampled deep features. Figure 8 Example of recognition effect: Displays original core photos, edge detection results, and crack results automatically identified by the model, annotating the number of real cracks and their corresponding depths, demonstrating the model's recognition effect. Detailed Implementation

[0018] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0019] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and preferred embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0020] The intelligent crack identification system based on deep multi-network feature fusion constructed in this embodiment includes a data preprocessing module and a deep multi-network intelligent identification model. The specific implementation of each module is as follows: Data preprocessing module: The input core images are subjected to color normalization, contrast enhancement, Gaussian noise reduction and size regularization in sequence. Then, a sliding window cropping strategy with a 256×256 pixel window and a step size of 128 pixels is used to generate image blocks that are adapted to the model input. Deep multi-network intelligent recognition model: Based on encoder-decoder architecture, it consists of a CNN-Transformer dual-branch encoding unit, a PMCA decoding unit, and a result output module. The CNN-Transformer dual-branch encoding unit consists of two branches: the CNN branch, an improved ResNet architecture, with four convolutional modules, each embedding a SEM-B bottleneck unit, outputting multi-scale features {F5, F6, F7, F8}; and the Transformer branch, a SegFormer architecture, composed of an overlapping patch embedding module and four Transformer blocks, outputting multi-scale features {F1, F2, F3, F4}. The cross-branch feature fusion module dynamically injects the features from the Transformer branch, after resolution adjustment, into the corresponding layers of the CNN branch, achieving complementarity between local and global features. PMCA decoding unit: Integrated at the skip connections of each level of the UNet-like decoder, it evenly divides the input feature map into 4 non-overlapping rectangular sub-regions. After global average pooling, shared weight MLP weight learning, and weighted merging, it outputs an enhanced feature map to achieve differentiated enhancement of key feature channels. Results output module: Performs contour optimization and visualization processing on the model segmentation results, and generates an analysis report containing the distribution of real / fake cracks and crack parameters.

[0021] Model training uses a combined loss function: L total =0.6L CE +0.4L Dice (α=0.6, β=0.4 are experimentally selected non-negative weight coefficients). By combining cross-entropy loss and Dice loss, the model's segmentation effect on small target cracks is improved.

[0022] The method for identifying real and fake cracks in core images based on the above system is implemented through the following steps: S1: Data Preparation and Preprocessing Core scan images of different blocks and lithologies were collected and meticulously annotated at the pixel level by geological experts to distinguish between true fractures, false fractures, and background, thus constructing the original dataset. Figure 1 The preprocessing workflow performs grayscale enhancement and histogram equalization on the original image, and then generates training samples by cropping through a 256×256 pixel window with a step size of 128 pixels. The samples are divided into training set, validation set and test set in a ratio of 8:1:1.

[0023] S2: Model Building according to Figure 3 System architecture, Figure 4 Two-branch coding network Figure 5 SEM-B bottleneck unit, Figure 7 The PMCA and decoder integration method is used to build a deep multi-network intelligent recognition model with a total of approximately 5.1M parameters.

[0024] S3: Model Training The model was trained on an NVIDIA GPU training platform, using the AdamW optimizer (initial learning rate 1e-3), a cosine annealing learning rate scheduling strategy, and a combined loss function set to L. total =0.6L CE +0.4L Dice The total number of training iterations is 100 epochs (50-200 epochs is the preferred value). The training process is monitored in real time using a validation set, and hyperparameters are adjusted until the model converges.

[0025] S4: Crack Identification and Segmentation The core image to be identified is input into the trained model. In the encoding stage, the image features are extracted and fused by the CNN-Transformer dual branch. In the decoding stage, the features are enhanced by the PMCA module and then the upsampled deep features are fused step by step. Finally, the probability map of each pixel belonging to the background, real crack, and false crack is output through a 1×1 convolutional layer to generate a pixel-level segmentation map.

[0026] S5: Results Post-processing and Analysis The probability map output by the model is thresholded, and morphological operations such as small-region filtering and void filling are performed to optimize the crack boundaries. Parameters such as crack length, area, density, and orientation are statistically analyzed. Finally, a visual analysis report containing the distribution of true / false cracks and quantitative crack parameters is generated (e.g., ...). Figure 8 (As shown).

[0027] The system in this embodiment undergoes performance verification on an independent test set, and the key metrics are as follows: The average crossover ratio (mIoU) reached over 84.5%; The recall rate for true cracks exceeds 92%; The model has high inference efficiency, with only about 5.1M parameters, and supports rapid identification of massive amounts of core images.

[0028] Compared with existing mainstream segmentation networks (U-Net, DeepLabv3+, SegFormer), the model of this invention achieves significant improvements in mIoU, accuracy, and recall. It also has better anti-interference ability and recognition accuracy in complex mineral backgrounds and multi-scale crack scenarios, achieving a good balance between accuracy and inference efficiency.

[0029] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A real / fake crack intelligent identification system based on deep multi-network feature fusion, characterized in that: Includes a data preprocessing module and a deep multi-network intelligent recognition model; The data preprocessing module is used to perform standardization and sliding window cropping on the input core images in sequence. The standardization process includes color normalization, contrast enhancement, noise reduction and size regularization. The sliding window cropping strategy is used to generate image patches that are adapted to the model input. The deep multi-network intelligent recognition model is based on an encoder-decoder architecture and is used to extract features and perform semantic segmentation on preprocessed core images, outputting pixel-level classification maps.

2. The intelligent crack identification system based on deep multi-network feature fusion according to claim 1, characterized in that: The deep multi-network intelligent recognition model includes a CNN-Transformer dual-branch encoding unit, a partition merged channel attention (PMCA) decoding unit, and a result output module; The CNN-Transformer dual-branch coding unit is used to fuse local detail features with global semantic information; The Partition Merge Channel Attention (PMCA) decoding unit is integrated at the decoder skip connection and is used to perform spatially aware weight calibration and enhancement of encoded features; The results output module is used to post-process the segmentation results output by the model, including contour optimization, result visualization, and to generate an analysis report containing information on the distribution of real and fake cracks.

3. The intelligent crack identification system based on deep multi-network feature fusion according to claim 2, characterized in that: The CNN-Transformer dual-branch coding unit includes a CNN branch, a Transformer branch, and a cross-branch feature fusion module. The CNN branch is based on an improved ResNet architecture and contains multiple convolutional modules with embedded SEM-B bottleneck units, which are used to extract local texture and edge features from core images in layers. The bottleneck unit integrates standard convolution, depthwise separable convolution, and dilated depthwise separable convolution, which achieves lightweight model while ensuring feature extraction capabilities. The Transformer branch is based on the SegFormer architecture and consists of an overlapping patch embedding module and multiple Transformer blocks, used to model long-range dependencies and global context of an image. The cross-branch feature fusion module is used to dynamically inject the multi-scale global features output by the Transformer branch into the corresponding layer features of the CNN branch after adaptive adjustment, so as to achieve complementary enhancement of local features and global semantics.

4. The intelligent crack identification system based on deep multi-network feature fusion according to claim 2, characterized in that: The core of the Partition Merging Channel Attention (PMCA) decoding unit lies in combining the channel attention mechanism with the spatial partitioning strategy. Its working process includes the following steps: (1) Spatial partitioning: The input feature map is evenly divided into K non-overlapping rectangular sub-regions, and global average pooling is performed independently on each region to obtain its representation vector; (2) Partition pooling and weight learning: A channel attention weight vector is generated for each region through a multilayer perceptron (MLP) with shared weights; (3) Weighted merging and feature enhancement: The pooling features of each sub-region are multiplied by their corresponding weights and summed to obtain a merged channel weight vector. This weight vector is then multiplied by the original input feature map channel by channel to output the enhanced feature map, thereby achieving differentiated enhancement of key feature channels in different spatial regions.

5. The intelligent crack identification system based on deep multi-network feature fusion according to claim 4, characterized in that: K in step (1) is determined through experimental optimization. When K=4, the model achieves the optimal balance between the average intersection-union ratio (mIoU) and the recall rate.

6. The intelligent crack identification system based on deep multi-network feature fusion according to claim 1, characterized in that: The training process of the deep multi-network intelligent recognition model adopts a combined loss function, which is composed of a weighted sum of cross-entropy loss and Dice loss. Its calculation formula is: L total =αL CE +βL Dice , where α and β are both non-negative adjustable weighting coefficients.

7. A method for intelligent identification of true and false cracks based on deep multi-network feature fusion, implemented using the intelligent identification system for true and false cracks in core photographs as described in any one of claims 1-6, characterized in that: Includes the following steps: S1. Data Preparation and Preprocessing: Collect and label core photos, and construct a dataset containing real cracks, false cracks, and background categories; standardize and enhance the images, and divide the dataset into training, validation, and test sets; S2. Model Construction: Construct the deep multi-network intelligent recognition model described above; S3. Model Training: Train the model using the training set data, monitor the training process using the validation set and adjust the hyperparameters until the model converges. S4. Crack Identification and Segmentation: Input the core photo to be identified into the trained model, and the model automatically outputs a pixel-level segmentation map, in which each pixel is classified as background, real crack, or fake crack. S5. Post-processing and analysis of results: The result output module performs thresholding and post-processing on the probability map to generate the final segmentation mask and analysis results. Then, it performs morphological operations on the segmentation map output by the model to optimize the boundary, statistically analyzes the length, area, and density parameters of the cracks, and generates a visualization report.

8. The intelligent crack identification method based on deep multi-network feature fusion according to claim 7, characterized in that: In step S1, a sliding window cropping method is used to generate training samples from high-resolution core images. The window size is set to 256×256 pixels and the step size is 128 pixels to increase the number of samples and ensure the integrity of the fracture area.

9. The intelligent crack identification method based on deep multi-network feature fusion according to claim 7, characterized in that: In step S3, the AdamW optimizer and cosine annealing learning rate scheduling strategy are used to train the model, with a total of 50-200 epochs for training.

10. The intelligent crack identification method based on deep multi-network feature fusion according to claim 7, characterized in that: Step S4 is as follows: In the encoding stage, the image extracts features through both CNN and Transformer branches, and achieves complementary advantages through cross-branch fusion mechanism; in the decoding stage, the deep features are upsampled and then fused with the corresponding shallow high-resolution features enhanced by the PMCA module step by step to gradually restore spatial details. Finally, a 1×1 convolutional layer is used as the core execution unit to output the probability map of each pixel belonging to the background, real crack, or fake crack.

Images