CN116229458A - A method for detecting inclusions based on YOLOV5 - Google Patents

Abstract

The invention belongs to the field of oil and gas exploration. A detection method of a fluid inclusion based on YOLOV 5. The YOLOV5 model is improved aiming at the characteristics of the fluid inclusion, so that the method is suitable for detecting the fluid inclusion. The technical core of the invention comprises: collecting and creating a fluid inclusion dataset; the model adds CA attention standing mechanism; the PANet in the neg part of the YOLOV5 model is replaced with BIFPN; the model is added with a small target detection layer; better weights are trained for detection. The invention can carry out real-time intelligent analysis on the inclusion sheet under the microscope to judge whether the inclusion sheet is a fluid inclusion or not, and can realize rapid and effective identification of the fluid inclusion.

Description

A method for detecting inclusions based on YOLOV5

technical field

The invention relates to the field of oil and gas exploration, and relates to a method for detecting fluid inclusions based on YOLOV5.

Background technique

Inclusions captured in minerals are the most complete and direct original fluid (or melt) samples preserved so far, and temperature measurement of inclusions can provide quantitative basis for the determination of fluid properties, sources, diagenetic stages, and diagenetic environments. It has important guiding significance for oil and gas resource evaluation, reservoir geochemistry, fluid type, fluid source and exploration. The oil and gas charging time can be determined by using the homogeneous temperature of oil and gas inclusions, and the oil and gas accumulation periods can be directly divided according to the difference in oil and gas charging time. Combined with the geological structure characteristics of the sample output, the burial heating history of the reservoir, etc., the homogeneous temperature can also be used to estimate the accumulation time and paleogeothermal gradient, infer the history of diagenesis, and study the maturity of source rocks. In order to study inclusions, we must first find inclusions in minerals. At present, the commonly used method to find inclusions is to observe the inclusion sheet under an optical microscope, then circle the approximate range, and then conduct the next step of temperature measurement research.

The target detection task is to let the computer automatically detect the location and type of the target object in the image or video. It is a classic task in computer vision. At present, deep learning methods in the field of target detection are mainly divided into two categories: two-stage target detection algorithms; one-stage target detection algorithms. The former is to generate a series of candidate boxes as samples by the algorithm, and then classify the samples through the convolutional neural network; the latter does not need to generate candidate boxes, and directly converts the problem of target frame positioning into a regression problem. It is precisely because of the difference between the two methods that they are also different in performance. The former is superior in detection accuracy and positioning accuracy, and the latter is superior in algorithm speed. As a one-stage detector, YOLOv5 has the advantages of small amount of calculation and fast recognition speed. At present, it is widely used in target recognition, and the recognition effect is good.

Finding inclusions under a light microscope can also be viewed as a form of image processing, looking for patterns of specific shapes. Most of the fluid inclusions are oval, round and irregular, and a small amount of pure liquid inclusions are irregular. Fluid inclusions mainly include single-phase and two-phase inclusions, and the gas phase in most gas-liquid two-phase inclusions fluctuates violently. According to the characteristics of fluid inclusions, there is a phenomenon of violent jumping of the gas phase. There are obvious features in the instantaneous image, and the deep learning method can be used to identify the image of this special shape, that is, inclusion target detection. YOLOv5 is a high-performance, general-purpose target detection model that can complete the two tasks of target location and target classification at one time, so YOLOv5 is selected as the basic skeleton of target detection.

Because of the small size of inclusions, it is generally time-consuming and laborious to find inclusions under an optical microscope, and requires a certain amount of experience to find them. At the same time, manual searches will inevitably lead to missing and wrong finding. The invention improves the YOLOV5 model to improve the accuracy and effect of fluid inclusion detection.

Contents of the invention

The technical problem to be solved by the present invention is to propose a method for detecting fluid inclusions based on YOLOV5. The inclusion target is small. The present invention researches and improves the feature extraction network and feature fusion network of the YOLOV5 model, and at the same time adds a small target detection layer to improve the detection ability of fluid inclusions. The present invention overcomes the defects existing in manual search and realizes efficient and accurate identification of inclusions.

1. In order to achieve the above purpose, it is realized through the following technical solutions, and its specific steps are:

Step 1. Collection of images of inclusions: observe the inclusions under an optical microscope and take videos of the inclusions. Sampling the captured video, extracting frames at different times of the video, and obtaining pictures.

Step 2: Image annotation and data set division: Annotate the bounding box position and category of the inclusions in the obtained image, and divide the annotated data randomly into a training set and a test set at a ratio of 4:1.

Step 3: Image data enhancement: Use torchvision to process the training images, rotate and crop them, and increase the number of images in the training set to improve the recognition ability of the model.

Step 4, build the model: the detection network model of YOLOV5 consists of three parts: backbone backbone, Neck and output module output. The backbone backbone includes the BottleneckCSP module and the Focus module; the BottleneckCSP module is used to enhance the learning of the entire convolutional neural network Performance; the Focus module is used to slice the picture, expand the input channel to 4 times the original, and obtain the downsampling feature map after a convolution; the Neck uses a structure combining FPN and PAN, and the conventional The FPN layer of the FPN layer is combined with the feature pyramid from the bottom up, and the extracted semantic feature and position feature are fused, and the feature fusion of the backbone layer and the detection layer is performed at the same time, so that the model can obtain more abundant feature information; the output module output Predicts image features, outputting a vector with the target object's class probability, object score, and location of the object's bounding box.

Add attention mechanism: Introduce the coordinate attention (CA) attention mechanism in the backbone network of YoloV5. CA attention mechanism. The channel relationship and long-term dependence are encoded through precise location information. The specific operation is divided into two steps: Coordinate information embedding and Coordinate Attention generation. First, the global average pooling is decomposed into horizontal and vertical directions. Specifically, given an input X, we first encode each channel along the horizontal and vertical coordinates using a pooling kernel of size (H, 1) or (1, W), respectively. Therefore, the output of the cth channel with height h can be expressed as:

Similarly, the output of the cth channel of width w can be expressed as:

The above two transformations aggregate features along two spatial directions respectively to obtain a pair of direction-aware feature maps. The ability to encode horizontal and vertical location information into channel attention enables mobile networks to focus on a wide range of location information without causing too much computation. Not only the inter-channel information is obtained, but also the direction-related position information is considered, which helps the model to better locate and identify targets; it is flexible and lightweight enough to be easily inserted into the core structure of the mobile network; it is helpful for the package Body location recognition.

Optimize the Neck part of the model: replace the PANet of the original network with a BiFPN network to improve detection accuracy. The BiFPN network introduces learnable weight factors to represent the importance of different input features while repeatedly applying top-down and bottom-up multi-scale feature fusion. BiFPN uses cross-connection to remove nodes that contribute less to feature fusion in PANet, and adds a skip connection between the input node and output node of the same scale, which fuses more features without adding more costs. On the same feature scale, each bidirectional path is regarded as a feature network layer, and the same layer is repeatedly used to achieve higher-level feature fusion to improve detection accuracy. That is, the process of adding a 4-fold downsampling to the original input image. After the original image is down-sampled by 4 times, it is sent to the feature fusion network to obtain a feature map of a new size. The feature map has a small receptive field and relatively rich position information, which can improve the detection effect of small targets.

Adjust the detection scale: Similar to the traditional target detection network, the original YOLOv5s network also performs feature fusion from the third feature layer. The small target detection layer is to add the second feature layer to the feature fusion network, thereby improving the detection ability of the network for small targets. The present invention adds a small target detection layer on the basis of the original YOLOv5s algorithm to retain shallow semantic information. Add the 160×160 feature map that was not fused in the feature extraction network to the detection layer, and add an upsampling operation and a downsampling operation in the feature fusion network, so as to increase the final output detection layer to 4 layers. After adding the detection layer, the number of output prediction frames is correspondingly increased from 9 to 12, and the 3 added prediction frames are all with different aspect ratios and for small target detection.

Step 5. Train the model and adjust parameters to optimize the model: the training set that has been sampled and divided is trained based on the improved YOLOV5 model. The loss function is calculated for each iteration, and the parameter values are updated to minimize the value of the loss function until the model converges, while preventing overfitting.

Step 6. After completing the model training, save the model weight parameters and set the format to .pt format. Reload the weight file saved to the model, and use this weight file to detect inclusions, and detect whether there are inclusions in the image.

Description of drawings

Figure 1 is a picture of inclusions collected and processed;

Figure 2 uses Make Sense to mark the image;

Figure 3CA attention mechanism;

Figure 4 Improved YOLOv5 network model

Figure 5 result graph

Detailed ways

Step 1. Collection of images of inclusions: Since there is no data set about inclusions, the collection channel of the data set in the present invention is five dolomite slices in the laboratory. Observe the inclusions under an optical microscope and take a video of the inclusions. Then through the video sampling pictures, use OpenCV to sample the video at intervals, extract the frames of the video at different times, and obtain 500 pictures.

Step 2. Image annotation and data set division: Use Make Sense to mark the position and category of the object bounding box on the obtained picture, and then divide the data set into a training set and a verification set with a ratio of 4:1.

Step 3: Image data enhancement: Use torchvision to process the training images, rotate and crop them, and increase the number of images in the training set to improve the recognition ability of the model.

Step 4, build the model: the detection network model of YOLOV5 consists of three parts: backbone backbone, Neck and output module output. The backbone backbone includes the BottleneckCSP module and the Focus module; the BottleneckCSP module is used to enhance the learning of the entire convolutional neural network Performance; the Focus module is used to slice the picture, expand the input channel to 4 times the original, and obtain the downsampling feature map after a convolution; the Neck uses a structure combining FPN and PAN, and the conventional The FPN layer of the FPN layer is combined with the feature pyramid from the bottom up, and the extracted semantic feature and position feature are fused, and the feature fusion of the backbone layer and the detection layer is performed at the same time, so that the model can obtain more abundant feature information; the output module output Predicts image features, outputting a vector with the target object's class probability, object score, and location of the object's bounding box.

Add attention mechanism: Introduce the coordinate attention (CA) attention mechanism in the backbone network of YoloV5. CA attention mechanism. The channel relationship and long-term dependence are encoded through precise location information. The specific operation is divided into two steps: Coordinate information embedding and Coordinate Attention generation. First, the global average pooling is decomposed into horizontal and vertical directions. Specifically, given an input X, each channel is first encoded along the horizontal and vertical coordinates using a pooling kernel of size (H, 1) or (1, W), respectively. Therefore, the output of the cth channel with height h can be expressed as:

Similarly, the output of the cth channel of width w can be expressed as:

The above two transformations aggregate features along two spatial directions respectively to obtain a pair of direction-aware feature maps. The ability to encode horizontal and vertical location information into channel attention enables mobile networks to focus on a wide range of location information without causing too much computation. Not only the inter-channel information is obtained, but also the direction-related position information is considered, which helps the model to better locate and identify targets; it is flexible and lightweight enough to be easily inserted into the core structure of the mobile network; it is helpful for the package Body location recognition.

Optimize the Neck part of the model: replace the PANet of the original network with a BiFPN network to improve detection accuracy. The BiFPN network introduces learnable weight factors to represent the importance of different input features while repeatedly applying top-down and bottom-up multi-scale feature fusion. BiFPN uses cross-connection to remove nodes that contribute less to feature fusion in PANet, and adds a skip connection between the input node and output node of the same scale, which fuses more features without adding more costs. On the same feature scale, each bidirectional path is regarded as a feature network layer, and the same layer is repeatedly used to achieve higher-level feature fusion to improve detection accuracy. That is, the process of adding a 4-fold downsampling to the original input image. After the original image is down-sampled by 4 times, it is sent to the feature fusion network to obtain a feature map of a new size. The feature map has a small receptive field and relatively rich position information, which can improve the detection effect of small targets.

Adjust the detection scale: Similar to the traditional target detection network, the original YOLOv5s network also performs feature fusion from the third feature layer. The small target detection layer is to add the second feature layer to the feature fusion network, thereby improving the detection ability of the network for small targets. The present invention adds a small target detection layer on the basis of the original YOLOv5s algorithm to retain shallow semantic information. Add the 160×160 feature map that was not fused in the feature extraction network to the detection layer, and add an upsampling operation and a downsampling operation in the feature fusion network, so as to increase the final output detection layer to 4 layers. After adding the detection layer, the number of output prediction frames is correspondingly increased from 9 to 12, and the 3 added prediction frames are all with different aspect ratios and for small target detection.

Step 5. Train the model and adjust parameters to optimize the model: the training set that has been sampled and divided is trained based on the improved YOLOV5 model. The loss function is calculated for each iteration, and the parameter values are updated to minimize the value of the loss function until the model converges, while preventing overfitting.

Step 6. After completing the model training, save the model weight parameters and set the format to .pt format. Reload the weight file saved to the model, and use this weight file to detect inclusions, and detect whether there are inclusions in the image.

Claims (2)

1. This technology proposes a fluid inclusion detection method based on YOLOV5. The essence is to use the YOLOV5 algorithm for target recognition, make a data set according to the characteristics of fluid inclusions, improve the official YOLOV5 algorithm, and add a CA attention mechanism to the model. , Replace the PANet in the Neck part with BIFPN, and add a small target detection layer to make it more suitable for the detection of inclusions. The core elements of this technology are the data set creation of inclusions and the improvement of the YOLOV5 model. 2. A method for detecting fluid inclusions based on YOLOV5, characterized in that the following steps are adopted: (1) Observe the inclusions under an optical microscope and take a video of the inclusions. Sampling the captured video, extracting frames at different times of the video, and obtaining pictures; (2) Label the bounding box position and category of the inclusions in the obtained picture, and divide the labeled data randomly into a training set and a test set according to a ratio of 4:1; (3) Use torchvision to process the training images, rotate and crop them, and increase the number of pictures in the training set to improve the recognition ability of the model; 4) The YOLOV5 algorithm is improved, the official YOLOV5 algorithm is improved, the model adds a CA attention mechanism, replaces the PANet of the Neck part with BIFPN, and adds a small target detection layer. Make it suitable for fluid inclusion detection; Step 5. Train the model and adjust the parameters to optimize the model, calculate the loss function, and update the parameter value to minimize the value of the loss function until the model converges, and at the same time to prevent overfitting; Step 6. After the model training is completed, save the model weight file and reload it, and use this weight file to detect inclusions, and detect whether there are inclusions in the image.

Images