Deep learning-based sedimentary rock rapid classification system and method

Abstract

The invention discloses a portable sedimentary rock classification system and method based on deep learning, and relates to the technical field of rock identification and geological survey. The system comprises a portable microscopic image acquisition module, a cloud deep learning model and a mobile terminal module. A user shoots macroscopic or microscopic images of sedimentary rocks through a mobile phone, the macroscopic or microscopic images are uploaded to a cloud end and then subjected to feature extraction and classification through the optimized ResNet50 model, and a classification result is fed back to a user end in real time. The innovation points comprise: (1) adopting a macroscopic and microscopic data set training model to significantly improve the classification precision; (2) designing a portable microscopic module to realize rapid acquisition of microscopic images; (3) mobile terminal identification service; and (4) the ResNet50 model based on residual network optimization supports high-precision fine-grained classification (sandstone, conglomerate, siltstone and clay rock). Experiments show that the system classification accuracy reaches 99% or above, and compared with a traditional laboratory method, the efficiency is greatly improved. The method can be widely applied to geological exploration, teaching and scientific research and science popularization scenes.

Description

(1) Technical Field

[0001] The present invention relates to the cross - field of computer vision and geological survey, and particularly to an intelligent classification system and method for sedimentary rocks integrating deep learning technology and portable hardware modules, which are applicable to field geological surveys, rock resource exploration, and popular science education scenarios. (2) Background Art

[0002] Traditional sedimentary rock classification mainly relies on laboratory analysis, which requires complex processes such as sampling, slice preparation, and microscopic observation, taking several hours to several days. The existing technologies have the following limitations:

[0003] 1. High degree of manual dependence: The classification results are significantly affected by the operator's experience. Reports from the International Society of Rock Mechanics indicate that the classification consistency of different experts for the same sample is only 68% - 75%.

[0004] 2. Bulky equipment: The weight of laboratory - grade microscopes usually exceeds 10 kg, making it difficult to meet the requirements of on - site real - time detection.

[0005] 3. Low degree of digitization: Existing mobile applications mostly implement simple classification based on rule engines and cannot handle complex texture features.

[0006] In recent years, although some scholars have tried to apply machine learning to rock identification (such as the sandstone classification method disclosed in CN109934486A), there are still the following problems:

[0007] - The training data is single, only using macroscopic images;

[0008] - The model generalization ability is insufficient, and the accuracy rate is generally lower than 90%;

[0009] - A complete end - to - end service system has not been formed. (3) Summary of the Invention

[0010] Based on this, it is necessary to provide a sedimentary rock identification system and method based on deep learning for the above - mentioned technical problems.

[0011] In the first aspect, the present application provides a sedimentary rock identification method based on deep learning, which is applied to a sedimentary rock identification system. The method includes:

[0012] The present invention constructs a high-quality macro-micro training dataset, which is the key foundation for achieving high-precision sedimentary rock classification. The dataset covers various typical sedimentary rock types such as sandstone, conglomerate, siltstone, and claystone, and contains rich macro and micro images. The macro images are collected from samples under different perspective conditions, including far perspectives and near perspectives, which can reflect the overall characteristics of the rocks; the micro images are collected by a portable microscopic module with a resolution greater than 1600×2157 pixels and less than 3840×2160 pixels, which can clearly show the microstructures of the rocks. The data collection sources are extensive, including laboratory sample libraries, field sampling, and public databases, ensuring the diversity and representativeness of the data. To further improve the generalization ability of the model, the present invention performs various preprocessing and data augmentation operations on the images, such as color correction, random cropping, horizontal flipping, vertical flipping, Gaussian noise injection, etc. These operations not only enrich the data diversity but also enhance the model's adaptability to different environments and conditions.

[0013] The present invention adopts a variety of advanced technical means in the image preprocessing stage. For micro images, in addition to performing conventional size normalization and standardization operations, local contrast enhancement technology is particularly introduced. By enhancing the local contrast of the images, the texture features of sedimentary rocks can be more clearly highlighted, enabling the model to more accurately identify and classify. For macro images, background segmentation processing is performed to remove non-rock regions, thereby reducing the interference of background information on the classification results and further improving the image quality and classification accuracy. These preprocessing operations not only enhance the recognizability of the images but also provide a better data foundation for subsequent feature extraction and classification, effectively improving the overall performance of the system.

[0014] The present invention adopts a deep learning model based on ResNet50 and optimizes and improves it. During the model training process, first, the dataset is carefully annotated, using a multi-label annotation method. Each image is annotated with the sedimentary rock category and the corresponding particle size range, ensuring the accuracy and integrity of the data annotation. The first 1000 rounds use all the data for basic training to ensure that the model can learn the basic features of the data; subsequent rounds focus on difficult samples, and misclassified samples are screened by algorithms. The cumulative number of training rounds is not less than 4000 times to further improve the stability and classification accuracy of the model. Through this optimized training strategy, the model can better learn the complex features of sedimentary rock images, thereby improving the accuracy and reliability of classification.

[0015] In order to further improve the performance of deep learning models, the present invention optimizes the ResNet50 model in various aspects. In terms of the model structure, a residual network structure is introduced. Through the design of residual blocks, the problem of gradient disappearance in the training of deep networks is effectively alleviated, enabling the model to better learn deep features. During the training process, data augmentation techniques are also adopted, such as random cropping, horizontal flipping, vertical flipping, Gaussian noise injection, etc. These techniques not only increase the diversity of data but also improve the generalization ability of the model, enabling it to better adapt to different input images. Through these optimization measures, the deep learning model performs excellently in the sedimentary rock classification task and can quickly and accurately identify different types of sedimentary rocks.

[0016] In a second aspect, the present application provides a sedimentary rock identification module based on deep learning.

[0017] The present invention designs a portable microscopic image acquisition module. This module combines a microscopic lens with a mainstream smartphone camera and can achieve 50-fold magnification, facilitating the standardized acquisition of microscopic images of sedimentary rocks by users in the field environment. This module is characterized by being light and portable and easy to operate. Its weight is only a fraction of that of traditional laboratory microscopes, making it convenient to carry and operate. Its design also takes into account the complexity of the field environment and has functions of waterproofing and dustproofing, enabling it to adapt to various harsh environmental conditions. In addition, the module has good compatibility with smartphones. Users can directly take pictures and preview images through the mobile phone, with simple and intuitive operations, greatly improving the efficiency and convenience of field acquisition. Through this portable microscopic image acquisition module, users can quickly obtain high-quality microscopic images of sedimentary rocks in the field, providing reliable data support for subsequent classification analysis.

[0018] In a third aspect, the present application provides a sedimentary rock identification method based on deep learning.

[0019] The mobile module of the present invention has a variety of practical functions, including user authentication, image acquisition and uploading, result display, storage of historical recognition records, and visualization of real-time classification probabilities, etc. The user authentication function connects to the MySQL database, supports one-key login via WeChat, and automatically obtains the user's nickname and avatar information, facilitating users to quickly log in and use. The image acquisition module calls the mobile phone camera, supports real-time preview and photo-taking functions. Users can switch between the front and rear cameras through the interface buttons to select a suitable shooting angle. After shooting, the image is automatically stored in the local album, and a preview thumbnail is displayed. The image upload module uploads the image data to the cloud server via the HTTPS protocol, and the upload progress is displayed in real-time, supporting the resume function to ensure the integrity of data in case of unstable network. The result display module receives the classification results returned by the cloud, and displays the recognition category, confidence level, and corresponding probability distribution in the form of a card. The card design is simple and intuitive, including the icon, name, and detailed description information of the rock category. Users can click on the card to view the historical recognition records and understand the past operation situations. The historical recognition record storage function stores the recognition records in the MySQL database. Users can view the historical data in the personal center, and the data includes information such as time, image, and classification results, etc. Through these functions, the mobile module provides users with a convenient and efficient operation experience, enabling users to classify sedimentary rocks anytime and anywhere, and conveniently manage and view the classification results.

[0020] The present invention constructs a portable sedimentary rock classification system based on deep learning. The system has a complete architecture, covering the entire process from image acquisition to classification result feedback. The system mainly includes a portable microscopic image acquisition module, a cloud server, a mobile module, and a macro-micro training model. The portable microscopic image acquisition module is responsible for collecting microscopic images of sedimentary rocks in the field; the mobile module provides a user-friendly operation interface. Users can take macro images with a mobile phone or upload the collected microscopic images, and upload the images to the cloud server; the cloud server deploys an optimized deep learning model to extract features and classify the uploaded images, and feedback the classification results to the user side in real-time. This system architecture not only realizes intelligent sedimentary rock classification, but also improves the overall performance and user experience of the system through the powerful computing power of the cloud server and the convenience of the mobile module. In addition, the system also has good scalability and compatibility, and can be easily integrated with other geological exploration equipment or data platforms, providing strong technical support for fields such as geological exploration, teaching, and scientific research.

[0021] Technical effects

[0022] Verified, the system meets the following indicators:

[0023] (IV) Description of the drawings

[0024] Figure 1 Flow schematic diagram of the portable sedimentary rock classification method based on deep learning provided for Example 1;

[0025] Figure 2 Flowchart for the identification and use of the portable sedimentary rock classification module provided for Example 1;

[0026] Figure 3 Schematic diagram of the identification result interface of the portable sedimentary rock classification module provided for Example 1. (V) Specific implementation manners

[0027] Example 1: System deployment

[0028] The following is the detailed content of Example 1:

[0029] 1. Hardware configuration

[0030] 1.1 Portable microscopic image acquisition module

[0031] - The portable microscopic module uses a portable microscopic lens, which is connected to the camera of a Huawei smartphone and magnifies 50 times;

[0032] 1.2 Mobile module

[0033] - Development framework: Developed based on the mobile module framework. The front end uses WeChat developer tools for page design, and the logic layer is implemented using PyCharm. The mobile end supports mainstream mobile phone models.

[0034] - Functional modules:

[0035] - Image acquisition module: Calls the mobile phone camera and supports real-time preview and photo-taking functions. Users can switch the front and rear cameras through the interface button and select a suitable shooting angle. After shooting, the image is automatically stored in the local photo album, and a preview thumbnail is displayed.

[0036] - Image upload module: After the user selects the image to be recognized, the mobile end uploads the image data to the cloud server through the HTTPS protocol. The upload progress is displayed in real time, and the breakpoint resume function is supported to ensure the integrity of the data in case of unstable network.

[0037] - Result display module: Receives the classification result returned by the cloud and displays the recognized category, confidence level, and corresponding probability distribution in the form of a card. The card design is simple and intuitive, including the rock category icon, name, and detailed description information. Users can click on the card to view the historical recognition records and understand the past operation situations.

[0038] - User Data Management Module: Integrates user authentication functions, supports one - click login via WeChat, and automatically obtains user nickname and avatar information. The recognition records are stored in a MySQL database, and users can view historical data in the personal center. The data includes information such as time, images, and classification results.

[0039] 2. Model Training

[0040] 2.1 Dataset Construction

[0041] - Sample Collection: Macroscopic image collection covers typical sedimentary rock types such as sandstone, conglomerate, siltstone, and claystone. The shooting scenes include field sites, laboratories, and mine sites to ensure data diversity. Microscopic images are collected through a portable microscopic module with a magnification of 50 times. The sample sources include laboratory sample libraries, field sampling, and public databases, with a total sample size of over 4000 images.

[0042] - Data Pre - processing: Perform color correction on the collected images and unify the color space to sRGB. The images are cropped to 224×224 pixels, and the scale adjustment maintains the aspect ratio unchanged. Data augmentation methods include random rotation (±30°), horizontal flipping, vertical flipping, and Gaussian noise injection to enhance the generalization ability of the model.

[0043] - Data Annotation: Adopt a multi - label annotation method. Each image is annotated with sedimentary rock categories and corresponding particle size ranges to ensure data accuracy.

[0044] 2.2 Model Optimization

[0045] - Network Structure Improvement: Insert an SE (Squeeze - and - Excitation) attention module between the 3rd and 4th residual blocks of the ResNet50 model to enhance the model's attention to key features. The module obtains channel - level features through global average pooling, generates attention weights through a fully - connected layer and an activation function, and then multiplies them with the original feature map channel - by - channel to achieve feature recalibration.

[0046] - Training Strategy: Adopt a progressive training strategy. In the first 1000 rounds, all data is used for basic training with a learning rate of 0.001. In subsequent rounds, focus on difficult samples. The misclassified samples are screened through an algorithm, and the learning rate is gradually reduced to 0.0001. The total number of training rounds is ≥4000 times.

[0047] 3. On - site Testing

[0048] 3.1 Testing Environment

[0049] - Testing Location: Geotechnical Laboratory of Chongqing University. Representative rock samples are selected, covering different sedimentary rock types.

[0050] - Test equipment: Portable microscopic image acquisition module, smartphone (Huawei), cloud server and mobile module. The equipment was debugged in advance to ensure the normal operation of all functions.

[0051] 3.2 Test process

[0052] The on-site test steps for classification and recognition are as Figure 1 shown.

[0053] - Sample collection: Sedimentary rock samples were collected, including types such as sandstone, conglomerate, siltstone, and claystone. The sample collection followed the principle of randomness to ensure the representativeness of the data. Information such as sample location, geological background, and environmental conditions was recorded during the collection process.

[0054] - Image acquisition S1: The portable microscopic image acquisition module was used to take microscopic images of the samples, with the magnification set to 50 times and the shooting angle kept vertical. At the same time, the smartphone was used to take macroscopic images of the samples, ensuring sufficient light and a simple background, as Figure 2 shown in step 1. For each sample, no less than 5 images were taken, covering different perspectives and lighting conditions.

[0055] - Image preprocessing and uploading S2: The images were cropped to 224×224 pixels, and the aspect ratio was kept unchanged during the scale adjustment. Then the taken images were uploaded to the cloud server through the mobile module, as Figure 2 shown in step 2. After the upload was completed, wait for the deep learning model to process.

[0056] - Deep learning model processing S3: The cloud-optimized ResNet50 model was used for feature extraction and recognition. During the recognition process, data such as upload time, recognition time, and network latency were recorded to evaluate the performance of the system.

[0057] - Classification result generation S4: The fully connected layer outputs a four-dimensional classification vector (sandstone, conglomerate, siltstone, claystone), and the Softmax function calculates the class probabilities to generate the classification results.

[0058] - Result display S5: Receive the classification results returned by the cloud, and display the category with the highest probability, as Figure 2 shown in step 2. Record the recognized category and the corresponding probability distribution. At the same time, record the operation of the system in different environments to evaluate its stability and reliability.

[0059] 3.3 Test results

[0060] - Classification accuracy: Statistically, the classification accuracy of the system reached 99.2%, and the recognition accuracy of types such as feldspar sandstone, calcareous conglomerate, and silty mudstone exceeded 98%. Typical results are as Figure 3As shown, the system has a good recognition effect on different sedimentary rock types and meets the actual application requirements.

[0061] - User experience: The testers feedback that the system is easy to operate, has a friendly interface design, and the recognition results are accurate and reliable. The portable microscopic image acquisition module is light and portable, making it easy to operate in the field. The mobile module has complete functions and meets the on-site rapid recognition requirements.

[0062] 4. Industrial applicability

[0063] 4.1 Application scenarios

[0064] - Geological exploration: The system can assist geologists in quickly identifying sedimentary rock types in the field, improving exploration efficiency and reducing labor costs. It has important application value in the fields of mineral resource exploration, oil and gas field development, etc.

[0065] - Teaching and research: The system can be used as a teaching tool for geology to help students intuitively understand the characteristics of sedimentary rocks and enhance their learning interest. Researchers can use the system to collect a large amount of sample data for sedimentary rock related research work.

[0066] - Popular science education: The system can popularize sedimentary rock knowledge to the public through the mobile module, improving the public's understanding of geological science. It has good promotion prospects in geological museums, science popularization exhibitions and other occasions.

[0067] 4.2 Economic benefits

[0068] - Cost savings: The equipment cost of the system is about 2800 yuan, which is about 98% lower than the cost of traditional laboratory equipment. In actual applications, the system can significantly reduce the costs of sampling, transportation and analysis, etc., and improve work efficiency.

[0069] - Market potential: With the increasing demand for sedimentary rock recognition in the fields of geological exploration, teaching and research, and popular science education, the system has broad market prospects. In the future, it can be further expanded to the fields of metamorphic rock and igneous rock classification to enhance market competitiveness.

[0070] 5. Conclusion

[0071] The present invention provides a portable sedimentary rock classification system and method based on deep learning. Through technological innovations such as macro-micro data fusion, portable hardware design and model optimization, rapid and high-precision classification of sedimentary rocks is achieved. The system has good field applicability, user-friendliness and economic benefits, and can be widely applied in the fields of geological exploration, teaching and research, and popular science education. In the future, we will further optimize the system performance and expand the application scope to contribute to the development of geological science.

Claims

1. A portable sedimentary rock classification system based on deep learning, characterized in that, Including: - A portable microscopic image acquisition module, which consists of an adjustable microscopic lens module and a mobile phone adapter interface, and supports the standardized acquisition of microscopic images of sedimentary rocks in the field environment; - A cloud server, which deploys a deep learning model optimized based on ResNet50, and is used for feature extraction and classification of uploaded sedimentary rock images; - A mobile module, with functions of image uploading, result display and user data management, and communicates with the server through an interface; - A macro-micro training data set, which contains macro and micro images of sandstone, conglomerate, siltstone and claystone, and is constructed after preprocessing such as color correction, random rotation and scale adjustment.

2. The system according to claim 1, wherein The portable microscopic image acquisition module combines a microscopic lens with a mainstream smart phone camera interface and has a magnification of 50 times.

3. The system according to claim 1, wherein The deep learning model is optimized in the following ways: The deep learning model adopts a residual network (ResNet50) structure, and the network depth is 50 layers; - The residual network structure includes multiple residual blocks, and each residual block adds the input directly to the output through a skip connection to alleviate the problem of gradient disappearance; - The training rounds of the deep learning model are not less than 4000 times to ensure the stability of the model; - The deep learning model adopts data augmentation techniques such as random cropping and horizontal flipping during the training process to improve the generalization ability of the model.

4. The system according to claim 1, characterized in that, The mobile module includes: - A user identity authentication module, which is connected to a MySQL database; - A historical recognition record storage function, which records the time, image and classification results; - A real-time classification probability visualization module, which displays the confidence distribution of each category.

5. The system according to claim 1, characterized in that, The macro-micro training data set is constructed in the following ways: - The acquisition sources include laboratory sample libraries, field sampling and public databases; - The resolution of the micro images is greater than 1600×2157 pixels and less than 3840×2160 pixels. The macro images contain samples under different perspective conditions, including far perspectives and near perspectives; - The data augmentation methods include random cropping and horizontal flipping.

6. A sedimentary rock classification method based on deep learning, characterized in that, Including the following steps: S1. Image acquisition: Obtain sedimentary rock images through the portable microscopic module or the mobile phone camera; S2. Image preprocessing and uploading: Preprocess the images, including size normalization and standardization, and upload them to the cloud; S3. Deep learning model processing: Input the processed images into the ResNet50 model, and extract multi-scale features through the convolutional layer; S4. Classification result generation: The fully connected layer outputs a four-dimensional classification vector (sandstone, conglomerate, siltstone, claystone), and the Softmax function calculates the category probability; S5. Result display: Return the highest probability category and the visualization result to the user side.

7. The method according to claim 6, wherein In the model inference process in step S3, it satisfies: - The processing time of a single image ≤ 500ms; - Support for concurrent processing ≥ 100 requests / second; - The server uses GPU acceleration.

8. The method according to claim 6, characterized in that, The preprocessing in step S2 also includes: - Perform local contrast enhancement on the micro images; - Perform background segmentation on the macro images and remove non-rock areas.

Images