Deep learning enhanced intelligent sporopollen fossil microscopic image identification and screening method

Abstract

The invention relates to a deep learning enhanced intelligent sporopollen fossil microscopic image identification and screening method. The method comprises the following steps: 1, obtaining a training sample set and a test sample set; 2, performing data denoising, labeling and normalization processing on the acquired image data; 3, introducing a self-attention mechanism into a backbone network of YOLOv8 to enhance the feature extraction capability; 4, adaptively adjusting large and small features with different importance by adopting a weighted bidirectional feature pyramid; and 5, inputting the sporopollen fossil image into the trained model, and outputting a sporopollen fossil identification result. According to the method, the sporopollen fossil image features can be rapidly extracted, the recognition accuracy is high, and professionals can be helped to efficiently evaluate the geographical environment.

Description

Technical Field

[0001] The present invention mainly relates to the field of stratigraphy, and specifically to a deep learning enhanced intelligent spore and pollen fossil microscopic image recognition and screening method. Background Art

[0002] Pollen is the original material for the formation of petroleum. Pollen fossils can directly reflect the abundance, type, morphology and existence time of ancient vegetation, indicating geological age. At the same time, they are of great significance for paleoclimate reconstruction and prediction of future climate. In addition, the study of pollen fossils also plays a vital role in the comparative study of stratigraphic division. The identification of pollen fossils under the microscope is mainly based on the morphological characteristics of pollen and the structure and decoration of the pollen wall. At present, the identification of pollen fossils is still limited to naked eye observation, relying on complex equipment and a large number of manual operations, which is not only time-consuming but also has common problems such as false detection and missed detection.

[0003] There are two types of traditional target detection methods. One is to segment the image according to the features of the image and use the clustering method for target segmentation. The other is to use a sliding window to select the target and use a classifier for classification and recognition. The recognition efficiency and accuracy of these two methods are low and the robustness is poor. The target detection algorithm based on deep learning uses convolutional neural networks and has strong feature extraction capabilities. It can be divided into two categories according to whether a candidate box is generated. One is a two-stage target detection algorithm, which first uses a region proposal network to generate a candidate box, and then uses a classifier to identify and predict the features in the candidate box, including R-CNN, SPPNet, Fast R-CNN, Faster R-CNN, etc., which have higher accuracy. The other is a one-stage target detection algorithm, which no longer generates a candidate box, but directly predicts and regresses the target, including the YOLO series algorithm and the SSD series algorithm, which has faster recognition efficiency. The attention mechanism has a good performance in the field of computer vision. It processes the input features to focus the model's attention on more important areas and reduce the interference of irrelevant features, thereby improving the accuracy and performance of target detection tasks.

[0004] In recent years, with the development of deep learning and the improvement of computer processing performance, the use of target detection methods to identify and analyze pollen fossils has gradually become a new research direction for scientific researchers. Summary of the invention

[0005] In response to the above problems, the present invention uses sporopollen fossil thin section image data and deep learning target detection technology to identify six types of sporopollen fossils, providing a feasible solution for realizing multi-type sporopollen fossil image recognition. Different from traditional methods, this invention eliminates the interference of artificial factors through target detection methods, has higher objectivity and accuracy, and can significantly improve the speed and accuracy of image processing.

[0006] The objective of the present invention is achieved through the following technical solution: a deep learning enhanced intelligent sporopollen fossil microscopic image recognition and screening method, the method comprising the following steps:

[0007] Step 1: Based on obtaining the training sample set and the test sample set; collecting thin-section samples for the sporopollen fossil image under a high-resolution microscope, and then electronically processing the sporopollen fossil image slide;

[0008] Step 2: Based on the obtained thin section data, data denoising, labeling and normalization are performed to construct a target detection dataset of sporopollen fossil thin section images; denoising is to smooth the noise by Gaussian filtering; the constructed dataset is expanded by normalizing and labeling the sporopollen fossil thin sections;

[0009] Step 3: Feature extraction is performed based on the introduction of the self-attention mechanism into the backbone network of YOLOv8; its self-attention mechanism module is the Swin-Fa block;

[0010] Swin Transformer divides the feature map into non-overlapping windows of different sizes. The window-based multi-head self-attention module calculates self-attention within each window. The sliding window-based multi-head self-attention module calculates self-attention after the window slides to achieve cross-window connection global modeling. The calculation of two consecutive Swin Transformer Blocks is as follows:

[0011]

[0012] Among them, W-MSA represents a window-based multi-head self-attention module. It calculates self-attention within each window, and SW-MSA represents a sliding window-based multi-head self-attention module. It calculates self-attention after the window slides to achieve cross-window connection global modeling. and z l They represent the output features of the (S)W-MSA module and the MLP module of the lth Swin Transformer Block, respectively. MLP is a multi-layer perceptron, and LN(·) is a layer normalization operation.

[0013] Step 4: Based on the weighted bidirectional feature pyramid, the size features of different importance are adaptively adjusted, and cross-scale connections are used between input and output nodes to achieve higher-level feature fusion;

[0014] Step 5: Input the sporopollen fossil image into the trained model and output the sporopollen fossil recognition result. The output of the model includes the probability of the sporopollen fossil category in the thin slice image to be detected.

[0015] The two-dimensional Gaussian distribution function formula described in step 2 above is as follows:

[0016]

[0017] Among them, x and y represent the row and column distances from the center point of the Gaussian window to the current point, G(x,y) is the weight at the position (x,y) in the Gaussian window, σ 2 is the variance of the Gaussian filter, and e is a natural constant.

[0018] In the above step 2: the marked pollen fossil categories include Acostraca, Alsophila, Porites, Osmanthus, Cedar powder and Krasuo powder.

[0019] In the above step 2: the main purpose is to construct and expand the required target detection data set by Gaussian blur sharpening, cropping, and splicing the pollen fossil thin section image.

[0020] In the above step 3: the self-attention mechanism based on factor decomposition is introduced; in order to maintain the relatively strong extraction ability of Swin Transformer for local features, the Swin Transformer Block is replaced by the Factorization Block only in Stage 4 where the spatial information loss is the most serious. The feature map output by Stage 3 is input into the Factorization Block through the patch merging layer and is evenly divided into multiple groups of features. i is the feature map of the i-th group, Q, K, V are feature matrices, and each group of features is factored independently; the sliding window strategy is used to convert z i Uniformly divide it into multiple non-overlapping local windows, and use expansion sampling to uniformly sample a fixed number of points in each local window, using P i j Represents the sampling point of the jth local window in the i-th group; embed features for each sampling point, and then fuse the features of the sampling points at the same position in different windows across windows:

[0021] K i =maxpooling(W i K P i j )

[0022] V i =maxpooling(W i V P i j )

[0023] Among them, maxpooling represents the maximum pooling, W iK and W i V represents the learnable linear embedding of the i-th group achieved by two separate 1×1 convolutions.

[0024] In the above process, the full attention matrix is ​​decomposed into several sparse sub-attention matrices; rich global information is obtained by merging features evenly distributed across the entire feature map; the final output features are as follows:

[0025]

[0026] z′=δ(z i ′,|i=1,...,G)

[0027] Among them, Softmax represents the activation function, Q i , K i and V i Indicates characteristics, represents the scaling factor, δ(·) represents the concatenation operation along the feature channel dimension, z′ is the final output feature, and G represents that the feature is decomposed into G parts in total.

[0028] The specific correlation coefficient formula of the evaluation index of the model described in step 5 above is as follows:

[0029]

[0030] Among them, TP is the number of samples correctly predicted as positive samples; FP is the number of samples incorrectly predicted as positive samples; FN is the number of samples incorrectly predicted as negative samples; P is the proportion of samples predicted as positive samples that are actually positive samples, which measures the accuracy of predicting positive samples; R is the proportion of samples that are actually positive samples that are correctly predicted as positive samples, which evaluates the effectiveness of detection performance; AP k is the average prediction accuracy of each class being predicted; n is the number of classes; mAP is the average of the average prediction accuracy of all classes.

[0031] The technical effects produced by the present invention are as follows:

[0032] 1. By comparing with public data sets, the YOLOv8 model with better performance was selected as the model framework. The Swin-Fa module was proposed and used to improve the backbone network of YOLOv8, enhancing the global feature extraction capability, thereby improving the model's detection performance for blurred targets.

[0033] 2. In order to better learn multi-scale features, the feature pyramid structure of the neck network feature fusion module is replaced with the BiFPN structure, which improves the detection accuracy of small targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Attached Figure 1 It is a block diagram of the deep learning enhanced intelligent spore and pollen fossil microscopic image recognition and screening method of the present application;

[0035] Attached Figure 2 This is a comparison of the spore-pollen fossil image before and after preprocessing;

[0036] Attached Figure 3 It is the dataset annotation graph;

[0037] Attached Figure 4 This is a model diagram for identifying pollen fossils;

[0038] Figure 5 This is the output result diagram of pollen fossil identification. DETAILED DESCRIPTION

[0039] The present invention is further described below in conjunction with embodiments:

[0040] A deep learning enhanced intelligent microscopic image recognition and screening method for pollen fossils, the method comprising the following steps: Figure 1 This is a flowchart of the steps of this application. Follow the flowchart as follows:

[0041] Step 1: This example uses the pollen fossils of the high-quality source rock section in the Hailar Basin as the research object, and uses the LEICADMI6000B50*630X stereo microscope of Leica Instruments Co., Ltd. to extract pollen fossil images under the working conditions of power supply: 220V (±10%), 50Hz, 1Φ, temperature: 15℃-25℃, humidity: <60%. Select the rock sample of the target layer, extract the pollen fossils and make thin slices according to the requirements (industry standard SY / T 5915-2018). To ensure clear identification, the slide and cover glass need to be cleaned with anhydrous ethanol, the glycerol glue on the slide should be appropriate, and the excess glue should be wiped off with anhydrous ethanol after the slice is cooled.

[0042] Step 2: Regarding the complex background information of sporopollen fossil images, this application uses image standardization to process. In order to enhance the desired information features and suppress unwanted redundant features, the present invention uses image denoising and sharpening operations to pre-process the image. Since the acquisition of image data is subject to external and subjective influences of the shooting environment, image denoising is conducive to producing a less interfering and clearer data set, thereby improving the performance of the algorithm.

[0043] This embodiment uses Gaussian filtering to achieve the noise reduction process of pollen fossil thin sections. Gaussian filtering can combine the uncertainty involved in noise detection and filter out low-frequency noise, thereby smoothing the image and making the edges clearer. It is obtained by weighted average of each pixel in the image and other pixels in the adjacent position. Gaussian filtering smoothes the noise and can well suppress the noise that obeys the normal distribution. Since the images are two-dimensional, what we need in image processing is a two-dimensional Gaussian distribution function, where the two-dimensional Gaussian distribution function formula is as follows:

[0044]

[0045] Among them, x and y represent the row and column distances from the center point of the Gaussian window to the current point, G(x,y) is the weight at the position (x,y) in the Gaussian window, σ 2 is the variance of the Gaussian filter, e is a natural constant, σ determines the amount of smoothing, and the Gaussian kernel coefficient depends on the value of σ. When σ becomes larger, more variance is allowed around the mean, and when σ becomes smaller, less variance is allowed around the mean. Therefore, a large filter variance is effective for smoothing an image, but at the same time it destroys the edges and important structures of the image.

[0046] Due to the lighting environment and human error factors when capturing the sporopollen fossil image, the quality of the obtained thin-section image is sometimes not ideal, which will affect the model performance. In order to compensate for this effect, this application adopts the method of increasing the high-frequency component to suppress the low-frequency component, enhance the details and contours of the edges of the rock particles in the thin-section image, and improve the edge clarity of the image, thereby achieving the purpose of image sharpening, and gaining some detail information and edge features of the image. The formula is as follows:

[0047]

[0048] Among them, F(x,y) represents the input image, (μ,v) represents each position of the image, and H(μ-x+1,v-y+1) is called the impulse response.

[0049] The spore and pollen fossil images obtained after the above preprocessing are as follows: Figure 2 As shown in the figure, it can be seen that the edge features are clearer and the edges of the target objects are distinct. Figure 3 shown.

[0050] Step 3: YOLOv8 is one of the most representative examples of one-stage target detection algorithms. It integrates and optimizes the essence of the YOLO series of algorithms, which not only greatly improves the training speed, but also improves the detection accuracy to a certain extent. Its network structure mainly consists of three parts: backbone network, neck network and detection head. The backbone network is responsible for feature extraction. Its main function is to extract relevant information from the image, which can then be used for further processing and analysis by subsequent networks or modules. The neck network is located between the backbone network and the detection head, and fuses the features extracted by the backbone network. The neck network part of YOLOv8 adopts the idea of ​​path aggregation feature pyramid. The detection head recognizes the features finally input by the neck network.

[0051] This embodiment introduces the self-attention mechanism into the backbone network of YOLOv8 to enhance the feature extraction capability. CNN uses convolution kernels to extract local information, lacks understanding of the global information of the image, and cannot establish dependencies between features. The last C2f layer of the backbone network of YOLOv8 is replaced with the Swin-Fa module combined with the factor decomposition self-attention mechanism to optimize the limitations of YOLOv8 in capturing global and contextual information, realize global feature extraction, and enhance the model's small target detection capability. Swin Transformer divides the feature map into non-overlapping windows of different sizes. Two consecutive Swin Transformer Blocks are calculated as follows:

[0052]

[0053] Among them, W-MSA represents a window-based multi-head self-attention module. It calculates self-attention within each window, and SW-MSA represents a sliding window-based multi-head self-attention module. It calculates self-attention after the window slides to achieve cross-window connection global modeling. and z l They represent the output features of the (S)W-MSA module and the MLP module of the lth Swin Transformer Block, respectively. MLP is a multi-layer perceptron, and LN(·) is a layer normalization operation.

[0054] However, the local window-based self-attention still limits the ability to model global dependencies, and has a weak feature extraction capability for blurry and small-sized targets in pollen fossil images. Therefore, this paper introduces a self-attention mechanism based on factorization, Factorization Self-Attention, which decomposes the attention matrix into multiple sparse self-attention matrices and makes the superposition of sub-attention matrices effectively approximate the original attention matrix to model stronger global dependencies. In order to maintain the relatively strong local feature extraction capability of Swin Transformer, the Factorization Block is used to replace the Swin Transformer Block only in Stage 4 where spatial information loss is the most serious. The feature map output by Stage 3 is input into the Factorization Block through the patch merging layer and is evenly divided into multiple groups of features. i is the feature map of the i-th group, Q, K, V are feature matrices, and each group of features is factored independently. Use the sliding window strategy to factor z i Uniformly divide it into multiple non-overlapping local windows, and use expansion sampling to uniformly sample a fixed number of points in each local window, using P i j Represents the sampling point of the jth local window in the i-th group. Feature embedding is performed for each sampling point, and then the features of the sampling points at the same position in different windows are fused across windows:

[0055] K i =maxpooling(W i K P i j )

[0056] V i =maxpooling(W i V P i j )

[0057] Among them, maxpooling represents the maximum pooling, W i K and W i V represents the learnable linear embedding of the i-th group achieved by two separate 1×1 convolutions.

[0058] In the above process, the full attention matrix is ​​decomposed into several sparse sub-attention matrices. Rich global information is obtained by merging features evenly distributed across the entire feature map. The final output features are as follows:

[0059]

[0060] Among them, Softmax represents the activation function, Q i , K i and V i Indicates characteristics, represents the scaling factor, δ(·) represents the concatenation operation along the feature channel dimension, z′ is the final output feature, and G represents that the feature is decomposed into G parts in total.

[0061] Step 4: Adopt weighted bidirectional feature pyramid to adaptively adjust the size features of different importance. The Neck network of YOLOv8 adopts the idea of ​​PAFPN for feature fusion. PAFPN combines feature pyramid network (FPN) and path aggregation network (PANet), which can more fully fuse multi-scale features, but its essence is still the simple addition of features of different sizes, resulting in more large-scale features being used, while ignoring the use of small-scale features. To solve this problem, the algorithm of the present invention adopts BiFPN structure in the Neck network for multi-scale feature fusion. BiFPN can make features propagate bidirectionally between different resolution levels, promote the target detection model to learn targets of different scales more comprehensively, enhance the contextual understanding of the target, improve the detection performance of the model and reduce the phenomenon of false detection and missed detection.

[0062] BiFPN removes the unidirectional input nodes that do not participate in cross-scale feature fusion, simplifies the network structure, adds skip connections, adopts cross-scale connections and uses weights of different sizes for feature fusion, thereby fusing shallower features and improving the detection performance of small targets. Its calculation is as follows:

[0063]

[0064] Among them, Conv represents the convolution operation, is the intermediate feature of the i-th layer in the top-down path, is the output feature of the i-th level in the bottom-up path, is the input feature of the i-th level in the bottom-up path, is the input feature of the i+1th level in the bottom-up path, is the output feature of the i-1th level in the bottom-up path, w i is a weight parameter and ε is a hyperparameter.

[0065] The final identification model of pollen fossils is as follows Figure 4As shown. After determining the target price measurement model of pollen fossils, the present invention randomly divides the constructed pollen fossil thin section data set into three subsets, namely 80% training set, 10% validation set and 10% test set, with a ratio of 8:1:1. The Windows 11 operating system is used, the GPU is Nvidia RTX 4090, the video memory is 24GB, the programming language is Python3.10.14, the training framework is PyTorch2.2.2, and CUDA12.1 is used for acceleration.

[0066] In order to evaluate the quality of the model results, this application uses precision P, recall R, average precision mAP and F1 score as evaluation indicators as the evaluation indicators of this model.

[0067] The precision rate P is the proportion of samples predicted as positive samples that are actually positive samples, which measures the accuracy of predicting positive samples. The recall rate R is the proportion of samples that are correctly predicted as positive samples among samples that are actually positive samples, which evaluates the effectiveness of detection performance and is defined as follows:

[0068]

[0069]

[0070] Among them, TP is the number of samples correctly predicted as positive samples, FP is the number of samples incorrectly predicted as positive samples, and FN is the number of samples incorrectly predicted as negative samples.

[0071] The average precision mAP is obtained by summing the average precision AP of each category and then dividing it by the number of categories. It is the average of the AP values ​​of all categories and reflects the overall detection performance of the model. It is defined as follows:

[0072]

[0073] Among them, AP k is the average prediction accuracy of each class being predicted, and n is the number of classes.

[0074] In order to verify the effectiveness of the algorithm proposed in this application, the algorithm proposed in this application was compared with mainstream target detection algorithms such as SSD, Faster-RCNN, YOLOX, YOLOv7, and YOLOv8 based on the sporopollen fossil dataset under consistent experimental indicators and conditions. The results are shown in Table 1.

[0075] Table 1 Comparison of detection results of different algorithms

[0076]

[0077] From the results of horizontal comparison between the algorithm of this application and network models such as SSD, Faster-RCNN, YOLOX, YOLOv7, and YOLOv8, the detection performance of YOLOv8 is significantly higher than that of other mainstream target detection algorithms, with an accuracy of 79.1%, a recall of 74.4%, and an mAP50 of 78.3%. The improved algorithm of YOLOv8 proposed in this paper has better accuracy, recall, and average precision, which are 1.4 percentage points, 2.2 percentage points, and 1.9 percentage points higher than YOLOv8, respectively.

[0078] In order to verify the effectiveness of each part proposed in the improvement of the YOLOv8 algorithm by the algorithm of this application, an ablation study was conducted based on YOLOv8. The control variable method was used to ensure that the only difference between the experimental group and the control group was the tested module. The experimental results are shown in Table 2.

[0079] Table 2 Ablation experiment results

[0080]

[0081]

[0082] The experimental results show that compared with the original YOLOv8, the introduction of the Swin-Fa self-attention module in the backbone network of YOLOv8 increases the mAP50 by 0.9 percentage points. Continuing to perform pyramid improvements on the neck network, the mAP50 increases by 1.0% percentage points, thus verifying the effectiveness of each step of optimization of the YOLOv8 model in this paper.

[0083] Step 5: Input the pollen fossil image into the trained model and output the pollen fossil recognition result. The final pollen fossil recognition output result is shown in the figure below. Figure 5 As shown. The sporopollen fossils to be identified and their species can be marked from microscope images containing various impurities, providing a reference for relevant researchers, saving time for manual identification of fossils, and reducing the false detection rate and missed detection rate of manual identification.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit it; under the idea of ​​the present invention, the technical features of the above embodiments can also be combined; the present invention is described in detail, and ordinary technicians in non-stratigraphic fields can also operate it.

Claims

1. A deep learning enhanced intelligent sporopollen fossil microscopic image recognition and screening method, the method comprising the following steps: Step 1: Based on obtaining the training sample set and the test sample set; collecting thin-section samples for the sporopollen fossil image under a high-resolution microscope, and then electronically processing the sporopollen fossil image slide; Step 2: Based on the obtained thin section data, perform data denoising, labeling and normalization to construct a sporopollenized fossil thin section image target detection dataset; Denoising is to smooth the noise by Gaussian filtering; the data set is expanded and constructed by normalizing and labeling the spore and pollen fossil thin sections; Step 3: Feature extraction based on introducing the self-attention mechanism into the backbone network of YOLOv8; Its self-attention mechanism module is the Swin-Fa block; Swin Transformer divides the feature map into non-overlapping windows of different sizes. The window-based multi-head self-attention module calculates self-attention within each window. The sliding window-based multi-head self-attention module calculates self-attention after the window slides to achieve cross-window connection global modeling. The calculation of two consecutive Swin Transformer Blocks is as follows: Among them, W-MSA represents a window-based multi-head self-attention module. It calculates self-attention within each window, and SW-MSA represents a sliding window-based multi-head self-attention module. It calculates self-attention after the window slides to achieve cross-window connection global modeling. and z l They represent the output features of the (S)W-MSA module and the MLP module of the ,th Swin Transformer Block, respectively. MLP is a multi-layer perceptron, and LN(·) is a layer normalization operation. Step 4: Based on the weighted bidirectional feature pyramid, the size features of different importance are adaptively adjusted, and cross-scale connections are used between input and output nodes to achieve higher-level feature fusion; Step 5: Input the sporopollen fossil image into the trained model and output the sporopollen fossil recognition result. The output of the model includes the probability of the sporopollen fossil category in the thin slice image to be detected.

2. The deep learning-enhanced intelligent sporopollen fossil microscopic image recognition and screening method according to claim 1, characterized in that: The two-dimensional Gaussian distribution function formula described in step 2 is as follows: Among them, x and y represent the row and column distances from the center point of the Gaussian window to the current point, G(x, y) is the weight at the position (x, y) in the Gaussian window, σ 2 is the variance of the Gaussian filter and P is a natural constant.

3. The deep learning-enhanced intelligent sporopollen fossil microscopic image recognition and screening method according to claim 1, characterized in that: In step 2: the marked pollen fossil categories include Acostraca, Alsophila, Porites, Osmanthus, Cedar powder and Krasuo powder.

4. The deep learning-enhanced intelligent sporopollen fossil microscopic image recognition and screening method according to claim 1, characterized in that: In step 2: the main purpose is to construct and expand the required target detection data set by Gaussian blur sharpening, cropping, and splicing the pollen fossil thin section images.

5. The deep learning-enhanced intelligent sporopollen fossil microscopic image recognition and screening method according to claim 1, characterized in that: In step 3: the self-attention mechanism based on factor decomposition is introduced. In order to maintain the relatively strong extraction ability of Swin Transformer for local features, the Swin Transformer Block is replaced by the Factorization Block only in Stage 4 where the spatial information loss is the most serious. The feature map output by Stage 3 is input into the Factorization Block through the patch merging layer and is evenly divided into multiple groups of features. i is the feature map of the i-th group, Q, K, V are feature matrices, and each group of features is factored independently; the sliding window strategy is used to convert z i Uniformly divide it into multiple non-overlapping local windows, and use expansion sampling to uniformly sample a fixed number of points in each local window, using P i j represents the sampling point of the ,th local window in the ,th group; The features of each sampling point are embedded, and then the features of the sampling points at the same position in different windows are fused across windows: K i =maxpooling(W i K P i j ) V i =maxpooling(W i V P i j ) Among them, maxpooling represents the maximum pooling, W i K and W i V represents the learnable linear embedding of the i-th group achieved by two separate 1×1 convolutions. In the above process, the full attention matrix is ​​decomposed into several sparse sub-attention matrices; rich global information is obtained by merging features evenly distributed across the entire feature map; the final output features are as follows: z′=δ(z′ i ,|i=1,...,G) Among them, Softmax represents the activation function, Q i , K i and V i Indicates characteristics, represents the scaling factor, δ(·) represents the concatenation operation along the feature channel dimension, z′ is the final output feature, and G represents that the feature is decomposed into G parts in total.

6. The deep learning-enhanced intelligent sporopollen fossil microscopic image recognition and screening method according to claim 1, characterized in that: The evaluation index of the model described in step 5, the specific correlation coefficient formula is as follows: Among them, TP is the number of samples correctly predicted as positive samples; FP is the number of samples incorrectly predicted as positive samples; FN is the number of samples incorrectly predicted as negative samples; P is the proportion of samples predicted as positive samples that are actually positive samples, which measures the accuracy of predicting positive samples; R is the proportion of samples that are actually positive samples that are correctly predicted as positive samples, which evaluates the effectiveness of detection performance; AP k is the average prediction accuracy of each class being predicted; n is the number of classes; mAP is the average of the mean prediction accuracy across all classes.

Images