CN116229458A

CN116229458A - Method for detecting inclusion based on YOLOV5

Info

Publication number: CN116229458A
Application number: CN202310231853.5A
Authority: CN
Inventors: 王兴建; 文雪梅; 曹俊兴
Original assignee: Chengdu Univeristy of Technology
Current assignee: Chengdu Univeristy of Technology
Priority date: 2023-03-10
Filing date: 2023-03-10
Publication date: 2023-06-06

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

Method for detecting inclusion based on YOLOV5

Technical Field

The invention relates to the field of oil and gas exploration, in particular to a method for detecting a fluid inclusion based on YOLOV 5.

Background

The inclusion captured in the mineral is the most complete and direct original fluid (or melt) sample remained so far, and temperature measurement of the inclusion can provide quantitative basis for determining the fluid property, source, diagenetic period and diagenetic environment. Has important guiding significance for oil gas resource evaluation, reservoir geochemistry, fluid type, fluid source, exploration and the like. The oil gas filling time can be determined by utilizing the uniform temperature of the oil gas inclusion, and the oil gas storage period can be directly divided according to the difference of the oil gas filling time. By combining with the geological structure characteristics of the sample output, the buried heat history of the reservoir and the like, estimation of the time to be buried and the ancient geothermal gradient, inference of the diagenetic effect history, research on the maturity of the hydrocarbon source rock and the like can be performed by utilizing the uniform temperature. To conduct the inclusion study, first, the inclusion in the mineral is found. The method for finding inclusion is that the inclusion piece is observed under an optical microscope, then the approximate range is circled, and then the next temperature measurement study is carried out.

The target detection task is to enable a computer to automatically detect the position and the type of a target object of interest in an image or video, and is a classical task in computer vision. The current deep learning methods in the field of target detection are mainly divided into two categories: a target detection algorithm of the two stage; a one stage target detection algorithm. The former is that a series of candidate frames serving as samples are generated by an algorithm, and then the samples are classified by a convolutional neural network; the latter directly converts the problem of target frame positioning into regression problem processing without generating candidate frames. The two methods are different in performance, the former is superior in detection accuracy and positioning accuracy, and the latter is superior in algorithm speed. YOLOv5 as One-stage detector has the advantages of small calculation amount, high recognition speed and the like. The method is widely applied to target recognition at present, and the recognition effect is good.

Finding inclusions under an optical microscope can also be considered as an image processing effort to find patterns of specific shape. The fluid inclusion is mostly elliptical, circular and irregular, and a small amount of pure liquid phase inclusion is irregular. The fluid inclusion mainly comprises a single-phase inclusion and a two-phase inclusion, and the gas phase in most of the gas-liquid two-phase inclusion is violently jumped. According to the characteristics of the fluid inclusion, the phenomenon of gas phase violent jumping exists. The instantaneous image has obvious characteristics, and the image with special shape can be identified by a deep learning method, namely inclusion object detection. YOLOv5 is a high-performance and general target detection model, and can complete two tasks of target positioning and target classification at one time, so that YOLOv5 is selected as a basic skeleton for target detection.

Because the inclusion shape is small, it is time consuming and labor consuming to find the inclusion under the optical microscope generally, and it is necessary to rely on a certain experience to find, and meanwhile, the condition of finding leakage and finding error is unavoidable in manual finding. The method improves the YOLOV5 model, and improves the accuracy and effect of fluid inclusion detection.

Disclosure of Invention

The invention aims to solve the technical problem of providing a method for detecting a fluid inclusion based on YOLOV 5. The method has the advantages that the object of the inclusion is smaller, the feature extraction network and the feature fusion network of the YOLOV5 model are researched and improved, meanwhile, a small object detection layer is added, and the detection capability of the fluid inclusion is improved. The invention overcomes the defect of manual searching at present and realizes the efficient and accurate identification of inclusion.

1. In order to achieve the above purpose, the following technical solutions are adopted to achieve the purpose, and the specific steps are as follows:

step one, inclusion image collection: the inclusion was observed under an optical microscope and video of the inclusion was taken. And sampling the shot video, extracting frames of the video at different times, and obtaining pictures.

Step two, image annotation and data set division: and randomly dividing marked data into a training set and a testing set according to the ratio of 4:1 for the positions and the types of boundary boxes of inclusion bodies in the obtained picture mark graph.

Step three: image data enhancement: and (3) processing the training image by using the torchvision, rotating and cutting, and increasing the number of training set pictures so as to improve the recognition capability of the model.

Fourth, building a model: the detection network model of the YOLOV5 consists of a trunk backbone, neck and an output module output, wherein the trunk backbox comprises a Bottleneck CSP module and a Focus module; the Bottleneck CSP module is used for enhancing the learning performance of the whole convolutional neural network; the Focus module is used for slicing the picture, expanding an input channel to 4 times of the original input channel, and obtaining a downsampling feature map through one convolution; the Neck adopts a structure of combining FPN and PAN, combines a conventional FPN layer with a feature pyramid from bottom to top, fuses the extracted semantic features with the position features, and fuses the features of a trunk layer and a detection layer, so that the model obtains more abundant feature information; the output module predicts the image characteristics and outputs a vector with the class probability of the target object, the object score and the position of the object boundary box.

The mechanism of adding attention: a Coordinate Attention (CA) attention mechanism was introduced in the YoloV5 backbone network. CA attention mechanism. The channel relation and long-term dependence are encoded by accurate position information, and the specific operation is divided into 2 steps of embedding the Coordinatate information and generating Coordinate Attention. Global average pooling is first decomposed into two directions, horizontal and vertical. Specifically, given an input X, each channel is first encoded along the horizontal and vertical coordinates, respectively, using a scaling kernel of size (H, 1) or (1, W). Thus, the output of the c-th channel of height h can be expressed as:

likewise, the output of the c-th channel of width w can be expressed as:

the 2 transforms aggregate features along two spatial directions, respectively, to obtain a pair of direction-perceived feature maps. The ability to encode both lateral and longitudinal location information into channel attention enables the mobile network to focus on a wide range of location information without excessive computation. Not only is the information between channels acquired, but also the position information related to the direction is considered, so that the model is facilitated to better locate and identify the target; is flexible and lightweight enough to be simply inserted into the core structure of a mobile network; is favorable for positioning and identifying the inclusion.

Optimizing the Neck part of the model: and the PANet of the original network is converted into the BiFPN network, so that the detection precision is improved. The bippn network introduces a learnable weight factor to characterize the importance of different input features while iteratively applying top-down and bottom-up multi-scale feature fusion. The BiFPN adopts cross connection to remove the node with smaller contribution degree to feature fusion in the PANet, and a jump connection is added between the input node and the output node of the same scale, so that more features are fused while more cost is not increased. On the same characteristic scale, each bidirectional path is regarded as a characteristic network layer, and the same layer is repeatedly utilized for multiple times, so that higher-level characteristic fusion is realized, and the detection precision is improved. I.e. a 4-fold downsampling process of the original input picture. The original picture is sent to a feature fusion network after 4 times downsampling to obtain a feature picture with a new size, the feature picture has smaller receptive field and relatively rich position information, and the detection effect of detecting a small target can be improved.

And (3) adjusting the detection scale: similar to the traditional target detection network, the YOLOv5s original network also starts feature fusion from the layer 3 feature layer. The small target detection layer is formed by adding the layer 2 feature layer into a feature fusion network, so that the detection capability of the network on the small target is improved. 160×160 feature maps that were not originally fused in the feature extraction network are added to the detection layer, and 1 upsampling operation and downsampling operation are added to the feature fusion network, so that the final output detection layer is added to 4 layers. After the detection layer is added, the output prediction frames are correspondingly increased from 9 to 12, and the 3 added prediction frames are all different in aspect ratio and are detected for small targets.

Training a model and adjusting a parameter optimization model: the sample divided training set is trained based on a modified 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.

And step six, after model training is completed, saving model weight parameters, and setting the format as a. Pt format. Reloading the weight file stored in the model, detecting inclusion by using the weight file, and detecting whether the inclusion exists in the image.

Drawings

FIG. 1 is a view of the inclusion acquired and processed;

FIG. 2 marks an image using a Make Sense;

FIG. 3CA attention mechanism;

FIG. 4 improved YOLOv5 network model

FIG. 5 results

Detailed Description

Step one, inclusion image collection: since there is no dataset about inclusions, the collection channel for the dataset of the present invention is five dolomite flakes in the laboratory. The inclusion was observed under an optical microscope and video of the inclusion was taken. And then through the video sampling pictures, the OpenCV is used for sampling the shot video at intervals, and frames of the video at different times are extracted to obtain 500 pictures.

Step two, image annotation and data set division: and labeling the object boundary box positions and categories of the obtained pictures by using a Make Sense, and then dividing the data set into a training set and a verification set, wherein the ratio of the training set to the verification set is 4:1.

likewise, the output of the c-th channel of width w can be expressed as:

Claims

1. The technology provides a detection method of a fluid inclusion based on YOLOV5, which is essentially characterized in that a YOLOV5 algorithm is used for target identification, a data set is manufactured according to the characteristics of the fluid inclusion, the official YOLOV5 algorithm is improved, a CA attention standing mechanism is added to a model, and a BIFPN is used for replacing PANet of a Neck part and a small target detection layer is added, so that the detection method is more suitable for detection of the inclusion. The key elements of the technology are the data set preparation of inclusion and the improvement of a YOLOV5 model.

2. A detection method of a fluid inclusion based on YOLOV5 is characterized by comprising the following steps:

(1) The inclusion was observed under an optical microscope and video of the inclusion was taken. Sampling the shot video, extracting frames of the video at different times, and obtaining pictures;

(2) Randomly dividing marked data into a training set and a testing set according to the ratio of 4:1 for the positions and the types of boundary boxes of inclusion in the obtained picture marking graph; the method comprises the steps of carrying out a first treatment on the surface of the

(3) The torchvision is used for processing training images, rotating and cutting are carried out, and the number of training set pictures is increased, so that the recognition capability of a model is improved;

4) The yolo 5 algorithm is improved, the official yolo 5 algorithm is improved, the CA attention mechanism is increased by the model, the PANet of the Neck part is replaced by BIFPN, and a small target detection layer is added. Making it suitable for fluid inclusion detection;

training the model and adjusting the parameter to optimize the model, calculating a loss function, updating parameter values to minimize the value of the loss function until the model converges, and simultaneously preventing overfitting;

and step six, after model training is completed, saving a model weight file, reloading, detecting inclusion by using the weight file, and detecting whether the inclusion exists in the image.