Extinction angle feature recognition method based on orthoclase orthogonal polarization sequence image

Abstract

The invention discloses an extinction angle feature recognition method based on an orthoclase orthogonal polarization sequence image. Comprising the following steps: firstly, developing a stripe feature attention module to enhance the attention of a network on texture features; then, a 2D backbone network is constructed using the module to efficiently extract spatial features of the orthogonally polarized image. And then, the extracted sequence image features are fused, classification labels and position information are combined, and the fused sequence image features are sent to a customized spatial-temporal feature extraction module based on Transform, so that spatial-temporal features are further extracted, and accurate classification identification is realized. Compared with the prior art, the method provided by the invention has the advantages that the four classification evaluation indexes are obviously improved, the accurate identification of the extinction angle characteristics of the orthogonally polarized sequence image of the plagioclase mineral particles is realized, and the method has a wide application prospect in geological research and oil exploration.

Description

TECHNICAL FIELD

[0001] The application designs an extinction angle feature recognition method based on an orthographic polarized sequence image of plagioclase, and a new network model AFI-Net, relates to the problem of determining the lithological characteristics of underground rock layers in geological research and oil exploration, and belongs to the field of computer vision and intelligent information processing. BACKGROUND

[0002] Accurate identification of plagioclase particles with extinction angle features plays an important role in measuring the sodium and calcium content ratio in plagioclase, which is crucial for revealing the geological history and resource distribution of the earth. In the traditional manual determination method, geologists first identify plagioclase particles containing sodium plagioclase twins under an orthographic polarized microscope. Then, by rotating the polariscope clockwise and counterclockwise, the extinction change of the twin monomers is observed, which aims to determine whether the plagioclase particles show extinction angle features and calculate the extinction angle accordingly. Finally, the sodium and calcium ratio is determined by table lookup method. However, on the one hand, observing the extinction change of twin monomers has problems such as complex steps, low efficiency, and inability to analyze qualitatively and quantitatively, on the other hand, the identification results are greatly affected by subjective factors due to the limitation of the professional knowledge and work experience of the identification workers. With the in-depth development of geological exploration, it is urgent to introduce computer image automatic analysis technology to improve the work efficiency of extinction angle feature recognition.

[0003] The extinction angle feature in plagioclase particles plays an important role in many fields such as geology, materials science and environmental research. However, there is still a lack of work using image analysis technology to identify the extinction angle feature in these particles. With the continuous progress of action recognition technology, it is now possible to use deep learning methods to determine the extinction angle feature of plagioclase particles.

[0004] It is well known that in recent years, action recognition methods have been widely used in medical diagnosis, intelligent monitoring, autonomous driving and video classification. However, due to the particularity of the extinction angle feature of plagioclase mineral particles, general action recognition methods still face problems such as information loss, difficulty in identifying weak texture features and difficulty in extracting long-distance spatiotemporal features when processing them. Therefore, we need to explore the most suitable method to identify whether plagioclase mineral particles have extinction angle features. To address this challenge, we propose an extinction angle feature recognition method based on an orthographic polarized sequence image of plagioclase. SUMMARY

[0005] The application designs an extinction angle feature recognition method based on an orthographic polarized sequence image of plagioclase, referred to as AFI-Net, which can efficiently extract features from the polarized sequence image and effectively integrate, so as to accurately determine whether the plagioclase mineral particles have the extinction angle feature. Specifically, we first develop a stripe feature attention module, which aims to enhance the network's attention to texture features. Subsequently, we use the module to build a 2D backbone network to efficiently extract spatial features of the orthographic polarized image. Then, the extracted sequence image features are fused and combined with classification labels and position information, and are sent to our customized time-space feature extraction module based on Transformer to further refine the time-space features and realize accurate classification and recognition.

[0006] An extinction angle feature recognition method based on an orthographic polarized sequence image of plagioclase, comprising the following steps:

[0007] (1) a stripe feature enhancement module (SABlock) is constructed;

[0008] (2) a stripe feature attention module (SAModule) is constructed;

[0009] (3) a 2D backbone network (SANet) is constructed;

[0010] (4) a neck module (Neck) is constructed;

[0011] (5) a time-space feature extraction module (TTE) is constructed; BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A block diagram of the extinction angle feature recognition method based on the orthographic polarized sequence image of plagioclase of the application.

[0013] Figure 2 A structure diagram of SANet of the application.

[0014] Figure 3 A structure diagram of Neck, TTE and MLP Head of the application. DETAILED DESCRIPTION

[0015] The application will be further described below with reference to the accompanying drawings: Figure 1 , FIGS. Figure 2 , FIGS. Figure 3 , FIGS.

[0016] The AFI-Net model network structure and principle are as follows:

[0017] The AFI-Net proposed by the application is a neural network model for recognizing the extinction angle features of the orthographic polarized sequence image of plagioclase mineral particles. As shown in FIG.Figure 1 As shown in FIG. 1, AFI-Net is composed of four parts: SANet, Neck, TTE and MLPHead. SANet is used to extract the features of the image, Neck is used to linearly transform the features and add Cls Token and position encoding, TTE is used to extract the spatial and temporal features of the features of 42 images, and MLPHead is used to predict the category according to the spatial and temporal features. Next, we will introduce the technical details of each part in detail.

[0018] (1) SANet

[0019] SANet mainly consists of 5 layers, {L0, L1, L2, L3, L4}, in which L0 consists of a 7x7 convolutional layer and a Max pooling layer, mainly used to reduce the image size and primary features. L1-L4 are respectively composed of 3, 4, 6 and 3 SAModule cascaded to extract features of different scales.

[0020] SAModule is composed of a stripe feature enhancement block (SABlock) and a channel attention block (SEBlock) cascaded and combined with a skip operation to introduce more fine-grained and directional attention, so that the network can focus more on the processing of areas with texture and structure information.

[0021] SABlock first uses a convolution kernel set by specific parameters (sx, sy) to perform convolution processing on the feature map, aiming to extract its texture and structure information G x and G y in the horizontal and vertical directions respectively. Assuming is the input feature map, the specific expression is as follows:

[0022] G x = Convsx3(X), G y = Convsy3(X)

[0023] where sx and sy represent the horizontal and vertical Sobel operators respectively, Convsx3 represents a 3x3 convolution operation with a convolution kernel parameter of sx, and Convsy3 represents a 3x3 convolution operation with a convolution kernel parameter of sy.

[0024] Then, the features in the horizontal and vertical directions are respectively added after taking the absolute value, and the comprehensive feature G is obtained:

[0025] G = abs(G x ) + abs(G y )

[0026] where abs(·) represents the absolute value.

[0027] Finally, G is added to the original input feature map X to obtain a new enhanced feature X s :

[0028] X s = SABlock(X) = X + abs(G x ) + abs(G y ) = X + G

[0029] where SABlock(·) represents a stripe feature enhancement module.

[0030] SEBlock is mainly composed of two CBR modules and an SE unit; SEBlock first improves the feature expression ability, training stability and computational efficiency of the network by cascading two CBR modules, assuming that the input is X s , then the expression of the output X s2 is as follows:

[0031] X s2 = CBR 2 (X s ) = ReLU(BN(Conv3(X s )))

[0032] where CBR(·) = ReLU(BN(Conv3(·))), BN(·) represents a batch normalization operation, Conv3(·) represents a 3x3 convolution layer, ReLU(·) represents a rectified linear unit activation function, and CBR 2 (·) represents two CBRs in cascade;

[0033] Then, the SE module is used to dynamically adjust each channel of the feature, so that the network can better focus on important channels. The specific expression of SE is as follows:

[0034] SE(X s2 ) = X s2 ⊙ Sigmoid(FC(ReLU(FC(GAP(X s2 )))))

[0035] where GAP(·) represents a global average pooling operation, FC(·) represents a fully connected layer, and ⊙ represents element-wise multiplication.

[0036] From the above, the expression of SEBlock is as follows:

[0037] SEBlock(X s ) = ReLu(X s + SE(CBR 2 (X s )))

[0038] In summary, the expression of SAModule is as follows:

[0039] SAModule(X) = SEBlock(SABlock(X)) + X.

[0040] (2) Neck

[0041] The Neck module first processes the features extracted by SANet performs a Flatten operation, thereby mapping them to Then, the feature dimension of each image is reduced to

[0042]

[0043] Then, the token sequence passed to the Transformer encoder is:

[0044] Z = [z cls ,F 21 ,F 22 ,...,F 2T ]+p

[0045] where i = 1,...,T, denotes the i-th vector of F2, is a learnable classification token that is added to the beginning of the sequence; finally, z cls is used as a comprehensive feature representation to generate the prediction; is a learnable positional embedding code to preserve positional information.

[0046] (3) Spatio-temporal feature extraction module

[0047] It consists of three Transformer layers and an MLP head; each Transformer layer includes LN, MHSA, MLP, etc. sub-layers; given an input X, the definition of MHSA is as follows:

[0048] MHSA(X) = Concat(head1,...,head n )W O

[0049] where the calculation of each attention head head i is as follows:

[0050]

[0051] where j = 1,...,n, and denote the projection matrices of the j-th head's query (Q), key (K), and value (V), respectively, where dk is the dimension of the query / key, k = C), d j is set to C / J, Concat(·) is the concatenation operation

[12] , W O ∈ R C×C is the parameter matrix of the projection;

[0052] Suppose the input of the l-th layer of the TransformerLayer is Z l , then the expression of the TransformerLayer is as follows:

[0053] Z l = MHSA(LN(Z l ))+Z l

[0054] Z l+1 = MLP(LN(Z l ))+Z l

[0055] Bring Z l into Z l+1 to get the expression of the l-th layer of the TransformerLayer:

[0056] TransformerLayer(Z l ) = MLP(LN(MHSA(LN(Z l ))+Z l ))+MHSA(LN(Z l ))+Z l

[0057] where LN represents layer normalization, MLP consists of two linear projection layers and a GELU activation function, and the specific expression is defined as follows:

[0058] MLP(x) = FC(GELU(FC(x)))

[0059] As can be seen from the above, the expression of TTE is as follows:

[0060] TTE(x) = TransformerLayer 3 (x)

[0061] where TransformerLayer 3 (·) represents three TransformerLayer cascaded.

[0062] The AFI-Net of the application is implemented using the PyTorch 1.9 framework on a single RTX3090 GPU, and the related hyperparameter settings of the experiment are shown in Table 1. We also use data enhancement techniques such as random image flipping, rotation, cropping, and combinations of these techniques to help the method achieve better performance. The implementation codes of the related comparative methods can be downloaded from their respective websites, and are executed using the default settings in all tables.

[0063] According to the research needs, the artificial cropping technology of the interaction between the ellipse and the polygon is used to crop the orthographic polarized sequence images of the plagioclase mineral particles from the polarized sequence images of the rocks. According to the identification of the geologists and in combination with the computer resources, these images are divided into two categories according to whether they have the extinction angle characteristics, each category has 1920 groups, and 3840 groups of original data sets required by the experiment are constructed. The data set is divided into a training set, a validation set and a test set in the ratio of 6:2:2.

[0064] The application performs comparative tests of different algorithms on the mineral particle polarized image data set, and studies the influence of the network structure ablation on the classification performance of the AFI-Net, wherein Accuracy, Precision, Recall and F1score are used as evaluation indexes, and the larger the numerical value is, the better the classification effect is. The related experimental results are shown in Tables 2, 3 and 4.

[0065] Table 1 Training hyperparameter settings of AFI-Net

[0066]

[0067] Table 2 Comparison between AFI-Net and 2D-DNN-based classification model in four evaluation indexes on plagioclase mineral particle polarized image data set

[0068]

[0069] Table 3 Comparison between AFI-Net and 3D-DNN-based classification model in four evaluation indexes on plagioclase mineral particle image data set

[0070]

[0071] Table 4 Comparison between AFI-Net and pure-trandformer classification model in four evaluation indexes on plagioclase mineral particle image data set

[0072]

[0073] From the results of Table 2, Table 3 and Table 4, it can be shown that the classification effect of the AFI-Net is better than that of the single-mode, 3D-DNN-based and 2D-DNN-based classification models.

Claims

1. A method for identifying extinction angle features based on orthogonal polarization sequence images of plagioclase, characterized by the following steps: (1) A stripe feature enhancement module (SABlock) was constructed; (2) Construct a stripe feature attention module (SAModule); (3) A 2D backbone network (SANet) was constructed; (4) Constructed a neck module (Neck); (5) A spatiotemporal feature extraction module (TTE) was constructed.

2. The method according to claim 1, characterized in that The stripe feature enhancement module constructed in step (1) is constructed as follows: first, the convolution kernel with specific parameters (sx, sy) is used to convolve the feature map, aiming to extract its texture and structure information G in the horizontal and vertical directions respectively. x and G y , assuming To input the feature map, the specific expression is as follows: Where sx and sy represent the horizontal and vertical Sobel operators respectively, Convsx3 represents a 3x3 convolution operation with the convolution kernel parameter sx, and Convsy3 represents a 3x3 convolution operation with the convolution kernel parameter sy. Then, the absolute values ​​of the horizontal and vertical features are calculated and added together to obtain the comprehensive feature G: G=abs(G x )+abs(G y ) Where abs(·) means finding the absolute value; Finally, add G to the original input feature map X to get the new enhanced feature X s : X s =SABlock(X)=X+abs(G x )+abs(G y )=X+G Wherein, SABlock(·) represents the stripe feature enhancement module.

3. The method according to claim 1, characterized in that In step (2), a stripe feature attention module is constructed. The construction method is as follows: a stripe feature enhancement block (SABlock) and a channel attention block (SEBlock) are cascaded and combined with a jump operation to introduce more fine-grained and directional attention, so that the network can focus more on processing areas with texture and structure information; SEBlock is mainly composed of two CBR modules and one SE unit. SEBlock first cascades two CBR modules to improve the network's feature expression ability, training stability and computational efficiency. Assume that the input is X s , then the output is X s2 The expression is as follows: X s2 =CBR 2 (X s )=ReLU(BN(Conv3(X s ))) Among them, CBR(·)=ReLU(BN(Conv3(·))), BN(·) represents batch normalization operation, Conv3(·) represents 3x3 convolution layer, ReLU(·) represents rectified linear unit activation function, CBR 2 (·) indicates two CBRs are cascaded; Then, the SE module is used to dynamically adjust each channel of the feature so that the network can better focus on important channels. The specific expression of SE is as follows: SE(X s2 )=X s2 ⊙Sigmoid(FC(ReLU(FC(GAP(X s2 ))))) Among them, GAP(·) represents global average pooling, FC(·) represents fully connected layer, and ⊙ represents element-wise multiplication; From the above, we can see that the expression of SEBlock is as follows: SEBlock(X s )=ReLu(X s +SE(CBR 2 (X s ))) To sum up, the expression of SAModule is as follows: SAModule(X)=SEBlock(SABlock(X))+X Where X represents the input of SAModule.

4. The method according to claim 1, characterized in that In step (3), a 2D backbone network is constructed as follows: SANet mainly consists of 5 layers, {L0, L1, L2, L3, L4}, where L0 consists of a 7x7 convolutional layer and a Max pooling layer, which is mainly used to reduce the image size and primary features. L1 to L4 are composed of 3, 4, 6, and 3 SAModules cascaded respectively to extract features of different scales.

5. The method according to claim 1, characterized in that Step (4) constructs a neck module. The construction method is as follows: The Neck module first extracts the features of SANet Perform the Flatten operation to map it to Then, the feature dimension of each image is reduced to Then, the token sequence passed to the transformer encoder is: Z=[z cls ,F 21 ,F 22 ,...,F 2T ]+p Where, i=1,...,T, represents the i-th dimension vector of F2, is a learnable classification tag that is added to the beginning of the sequence; finally, z cls Used as a comprehensive feature representation to generate prediction results; is a learnable position embedding code to preserve position information.

6. The method according to claim 1, characterized in that Step (5) constructs the spatiotemporal feature extraction module. The construction method is as follows: it consists of three Transformer layers and a multi-layer perceptron (MLP) head; each Transformer layer includes LN, MHSA, MLP and other sub-layers; given the input X, the MHSA is defined as follows: MHSA(X)=Concat(head1,...,head n )W O Among them, the calculation of each attention head head i is as follows: Where, j=1,...,n, and denotes the projection matrices of the query (Q), key (K), and value (V) of the j-th head, respectively, where d k is the dimension of the query / key (d k =C), d j is set to C / J, Concat(·) is the concatenation operation, W O ∈R C×C is the parameter matrix of the projection; Assume that the input of the transformer in layer l is Z l , then the expression of TransformerLayer is as follows: Z l =MHSA(LN(Z l ))+Z l WITH l+1 =MLP(LN(Z l ))+Z l Z l Bring in Z l+1 Get the expression of TransformerLayer of layer l: TransformerLayer(Z l )=MLP(LN(MHSA(LN(Z l ))+Z l ))+MHSA(LN(Z l ))+Z l Among them, LN represents layer normalization, MLP consists of two linear projection layers and a GELU activation function. The specific expression is defined as follows: MLP(x)=FC(GELU(FC(x))) From the above, we can see that the expression of Temporal transformer-based encoder (TTE) is as follows: TTE(x)=TransformerLayer 3 (x) Among them, TransformerLayer 3 (·) indicates three TransformerLayer cascades.

Images