A method and device for classifying graptolite fossils

Abstract

The present application relates to a kind of graptolite fossil image classification method, comprising the following steps: preparing graptolite fossil image dataset;Build and train the embedding classification model of metadata, embedding classification model of metadata includes two parts of embedding model and classification layer model;Build and train graptolite fossil monomer image classifier;Graptolite fossil monomer image and corresponding metadata in test set are respectively input graptolite fossil monomer image classifier and metadata embedding classification model, and the output of fusion is obtained graptolite fossil monomer image class prediction result.The present application also simultaneously gives the graptolite fossil image classification device for realizing the above method steps and the computer readable storage medium of storage program instruction.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and specifically to a method and apparatus for classifying images of graptolite fossils. Background Technology

[0002] Graptolites are a group of extinct pelagic animals that lived 300 million years ago, encompassing multiple families, genera, and species. The distribution of some of these species is closely linked to the distribution of shale gas—a clean and efficient energy source. Accurate identification of shale gas indicator species is crucial for shale gas resource exploration. However, identifying graptolite fossil images requires extensive expertise and is extremely time-consuming. Furthermore, training prospectors to identify graptolite fossils is highly inefficient due to the lack of biological characteristics and the incompleteness and blurred boundaries of individual specimens. Millions of dollars are spent daily maintaining drilling rigs while awaiting paleontologists' advice. Therefore, the oil industry needs reliable and efficient graptolite fossil image classification technology to minimize potential oil exploration costs, making the development of automated graptolite fossil classification technology a priority for both the paleontological community and the oil and gas industry.

[0003] Automatic classification technology for graptolite fossil images needs to achieve species-level category prediction to determine whether they belong to an indicator category of oil and gas resources. Currently, classification technologies related to graptolite fossils are almost nonexistent, with most biological classification models relying entirely on images. However, for graptolite fossils, due to severe damage to some individuals, the information contained in the images themselves is limited. Therefore, utilizing information such as the fossil's excavation location, geological age, and phylogenetic relationship with other surrounding fossils is necessary to improve classification accuracy. Since existing deep learning-based image classifiers cannot effectively utilize metadata information beyond the image itself, their robustness and reliability need further improvement when applied to the classification of graptolite fossil images.

[0004] A literature search of existing technologies revealed that patent CN115546134A provides a semantic segmentation and classification method for CT images of microscopic fossils. This method first preprocesses the fossil images using reconstruction and super-resolution techniques, and then uses a deep learning-based approach for semantic segmentation and classification. While this method also treats fossils as classification objects, preprocessed microscopic fossils are often separable from the surrounding rock and have a relatively simple background. However, as macroscopic fossils, graptolites not only cannot exist independently of the surrounding rock, but also exhibit rich categorization and complex, diverse rock textures that can interfere with classification. Therefore, this method is not effectively applicable to the classification of graptolite fossils.

[0005] Further research revealed that patent CN104820718B provides an image classification and retrieval method based on geographic location features and global visual features. This method extracts Gist features from images as global visual features and uses the K-means algorithm to cluster these features based on location information, using the features at the cluster centers as class features. When a user uploads an image, the possible class rankings are obtained by comparing each image individually. However, the Gist feature extraction method used in this patent is relatively coarse and unlikely to be effective in fine-grained image classification. Furthermore, this method cannot utilize the correlation between species as a classification criterion. Summary of the Invention

[0006] This invention addresses the aforementioned shortcomings of existing technologies by providing a fossil image classification method based on metadata priors and deep learning. Based on the correlation between fossil species distribution and geographical location, geological age, and inter-species correlations, this invention proposes an embedding classification model for individual metadata, utilizing metadata corresponding to individual graptolite fossil images to improve classification accuracy and robustness. This invention is achieved through the following technical solution, specifically:

[0007] A method for classifying images of graptolite fossils, characterized by comprising the following steps:

[0008] Step 1: Prepare the graptolite fossil image dataset, as follows:

[0009] 1) Collect original specimens of graptolite fossils by photographing;

[0010] 2) Use the positioning system to obtain the geographical location of the excavation site of the original graptolite fossil specimen; determine the geological age of the original graptolite fossil specimen; use the geographical location and geological age of the original graptolite fossil specimen as the corresponding metadata of the original graptolite fossil specimen to construct a graptolite fossil image dataset;

[0011] 3) Divide the graptolite fossil image dataset into a training set and a test set, ensuring that the test set contains at least one original graptolite fossil specimen;

[0012] 4) Segment each individual graptolite image from the original graptolite fossil specimen image, so that each individual graptolite fossil image inherits the category, metadata, and dataset partitioning of the original graptolite fossil specimen;

[0013] 5) Label the category of individual graptolite fossil images, including three levels: family, genus, and species, forming a data pair: {individual graptolite fossil image, metadata, category};

[0014] Step 2: Build and train the metadata embedding and classification model. The metadata embedding and classification model consists of two parts: an embedding model and a classification layer model. The method is as follows:

[0015] 1) Construct an embedding model. The embedding model is a neural network composed of multiple fully connected layers with different output dimensions. Its input is the geographical location of the original graptolite fossil specimen, and its output is the embedding vector of the geographical location of the original graptolite fossil specimen.

[0016] 2) Based on the geological age and category corresponding to individual images of graptolite fossils, a regularization term is constructed using co-occurrence relationships to adjust the Euclidean distance between embedding vectors;

[0017] 3) Build a classification layer model. The input of the classification layer model is the embedding vector of the geographic location, and the output is the prediction results at the three levels of family, genus and species. Calculate the cross-entropy loss for the prediction results at the three levels of family, genus and species respectively.

[0018] 4) Minimize the weighted sum of cross-entropy loss and regularization term using gradient descent to obtain parameter estimates for the embedding model and the classification layer model;

[0019] Step 3: Build and train a single image classifier for graptolite fossils;

[0020] Step 4: Input the individual images of graptolite fossils and their corresponding metadata from the test set into the individual image classifier and the metadata embedding classification model, respectively, and then fuse the two outputs to obtain the category prediction results of the individual images of graptolite fossils.

[0021] Furthermore, between step one and step two, augmentation processing is performed on the individual images and metadata of graptolite fossils in the training set.

[0022] Furthermore, the individual images of graptolite fossils in the training set are subjected to color augmentation processing with random floating.

[0023] Furthermore, the individual images of graptolite fossils in the training set are augmented by rigid body transformation or affine transformation.

[0024] Furthermore, augmentation processing with line optical distortion, elastic distortion, or mesh distortion is performed on the individual images of graptolite fossils in the training set.

[0025] Furthermore, cutout augmentation is performed on the individual images of graptolite fossils in the training set.

[0026] Furthermore, noise is added to the geographic location data in the metadata for augmentation.

[0027] Furthermore, Gaussian noise is added to the individual images of graptolite fossils in the training set for augmentation.

[0028] The present invention also provides a classification device for graptolite fossil images that implements the above method steps and a computer-readable storage medium for storing program instructions.

[0029] The beneficial effects of this invention are as follows: This invention utilizes the corresponding metadata information of fossil images to train a classifier, and the result is combined with a classifier for individual graptolite fossil images, improving prediction accuracy. Simultaneously, it utilizes species phylogenetic relationships and co-occurrence relationships to regularize the embedding classification model, improving the robustness of the metadata classifier. Compared with existing technologies, the meta-prior data classifier has lower requirements for the quality of the fossil images themselves, and has a wider range of applications and greater reliability. Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention.

[0031] Figure 2 Image augmentation process.

[0032] Figure 3 An embedded classification model for metadata.

[0033] Figure 4 A classifier for individual images of graptolite fossils.

[0034] Figure 5 This describes the overall process for predicting the individual categories of graptolite fossils during the testing process.

[0035] Figure 6 This invention presents the accuracy achieved by the method using different metadata information and compares it with pure image classification methods.

[0036] Figure 7 The hardware structure diagram of the identification device provided by the present invention. Detailed Implementation

[0037] The present invention will be further described in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0038] This invention is achieved through the following technical solution, as shown in the flowchart. Figure 1 As shown, the specific steps of this invention are as follows:

[0039] Step 1: Prepare the graptolite fossil image dataset. Construct data pairs of {graptolite fossil individual image, metadata, category} for subsequent training of the graptolite fossil individual image classifier and metadata embedding classification model. The specific steps are as follows:

[0040] 1) Take photos to collect original specimens of graptolite fossils.

[0041] 2) The geological age is determined based on the strata of the core from which the original graptolite fossil specimen was found. Geological ages refer to different periods in Earth's history during which specific geological and biological evolutionary events occurred, leaving behind geological records and fossil evidence that help geologists and paleontologists understand the evolutionary process of Earth's history. Simultaneously, GPS information, including latitude and longitude, is obtained from the excavation site of the original graptolite fossil specimen. However, due to geological movements, the GPS information at the excavation site cannot reflect the true spatial relationships between species within that geological age. Therefore, the PointTracker tool is used to convert the GPS information at the excavation site into paleoGPS information, referred to as geographical location in this patent. The geographical location and geological age of the original graptolite fossil specimen are used as corresponding metadata to construct a graptolite fossil image dataset.

[0042] 3) Divide the graptolite fossil image dataset into a training set and a test set, ensuring that the test set contains at least one graptolite fossil specimen. The training set is used to train the neural network, and the test set is used to observe the training effect and determine the training termination time.

[0043] 4) A primary graptolite specimen may contain multiple individual graptolite fossils. To reduce classification difficulty and ensure that the images of individual graptolite fossils are at a uniform scale, it is necessary to manually segment each individual graptolite fossil image from the primary specimen. The manually segmented individual graptolite fossil images are then divided into a test set and a training set based on the primary specimen. Dividing the specimens by primary specimen ensures that individual graptolite fossils from the same primary specimen do not end up in separate training and test sets.

[0044] 5) Label the category of individual graptolite fossil images, including three levels: family, genus, and species. Individual graptolite fossil images segmented from the specimen inherit metadata information from the original graptolite fossil specimen, forming a data pair: {individual graptolite fossil image, metadata, category}, represented as: tuple D. i =(X i T i I i C i Where i represents the tuple index, I i This is a single image of the corresponding graptolite fossil, X i =[X lon X lat [] represents geographical location, with latitude and longitude represented by X. lon X lat It means that T i =[T f T t [This refers to the geological age of graptolite fossils, with the start and end times denoted by T.]f T t Indicates. C i =[y f y g y s [] represents the classification name of a single graptolite fossil at different biological taxonomic levels. The graptolite fossil image dataset contains categories spanning different families, genera, and species, denoted as y. f y g y s .

[0045] Step 2: Augmentation of individual images and metadata of graptolite fossils, the specific steps are as follows:

[0046] Data augmentation aims to improve the robustness of a model when data is limited. The graptolite fossil image augmentation process uses individual graptolite fossil images from the training set as input images, and then processes them into a... Figure 2 The augmentation process is shown.

[0047] 1) The input image is subjected to random color fluctuation or no change at a certain ratio. Random color fluctuation refers to randomly superimposing a random number on the three RGB channels.

[0048] 2) Perform rigid body transformation, affine transformation, or no transformation on the single image of graptolite fossil output in step 1) according to a certain proportion.

[0049] 3) Cut out or leave unchanged the single image of graptolite fossil output in step 2) according to a certain proportion. Cutout refers to randomly occluding some parts of the image with black cubes to improve the generalization ability of the image classifier.

[0050] 4) Add Gaussian noise to the single image of the graptolite fossil output in step 3).

[0051] 5) Noise is superimposed on the geographic locations in the metadata to generate augmented data near the original metadata points. With latitude and longitude normalized in the geographic locations, the noise is taken from a normal distribution with a mean of 0 and a variance of 0.01.

[0052] Step 3: Build and train the metadata embedding and classification model. The specific steps are as follows:

[0053] Metadata embedding classification models such as Figure 3 As shown.

[0054] 1) Construct an embedding model. The purpose of an embedding model is to map a variable to an embedding space, so that the distance between variables in the embedding space reflects the relationships between them. Geographic location X iGraptolite fossils are continuous variables on a two-dimensional plane, and Euclidean distance can be used to reflect the geographical distance between two points. However, due to the influence of geological time spans and other complex factors, the species correlation of graptolite fossils cannot be simply reflected using distance on a two-dimensional plane. Therefore, it is necessary to consider geographical location X... i Mapping to an embedding space allows the distances between variables in the embedding space to more accurately reflect the species relationships of graptolite fossils by incorporating other factors. The mapping of this embedding model is defined as a function f(), which is a neural network composed of eight fully connected layers. The output dimensions of each layer are 16, 16, 32, 32, 64, 64, 128, and 128, respectively. f() will normalize the geographic location X... i Mapped to an embedding vector f(X) of length 128 i ).

[0055] 2) Constructing regularization terms using co-occurrence relationships. The co-occurrence relationship of graptolites refers to the fact that two graptolite groups exist in the same geological period T. i Does an intersection exist? Since the probability of finding graptolites from the same paleoecological community increases significantly when a particular graptolite is discovered, and this probability decreases as the time interval between the two species increases, this condition is used as a regularization term to constrain the training of the embedding model f(). Metadata is paired and labeled, and y is set... tco For each tuple D, the co-occurrence label is used. i Select one from the dataset that corresponds to its geological age T i Overlapping tuples D p As a positive sample, assign this positive sample the label y. tco =1. Simultaneously, randomly select a tuple D whose geological age does not overlap. n As a negative sample, this negative sample is also labeled y. tco =0. The embedding vector f(X) output by the embedding model is used. i ), calculate the regularization function L tco =y tco *||f(X1)-f(X2)|| 2 Where X1 and X2 are geographic locations in the paired metadata. When the pairings have a co-occurrence relationship (i.e., y...), ... tco =1), by using gradient descent to reduce the regularization function, the Euclidean distance between the two embedding vectors is reduced. However, when paired samples have no co-occurrence relationship (i.e., y = 1), the Euclidean distance between their embedding vectors is reduced. tco =0), L tco =0 This constraint is invalid.

[0056] 3) Construct the classification layer model. The classification layer model consists of three parts, which are represented by O(n). k In this representation, k∈{f,g,s} represents the hierarchy of the classification model, where f,g,s represent family, genus, and species, respectively; Ok They are 128 lines N k The column weight matrix. Where N k This represents the number of categories at level k. k Each column of weights represents the embedding vector for that class, and these weights are obtained through training. When the geographic location X... i When the embedding vector has a high affinity with species c, O k Column c O k(：，c) With the embedding vector f(X) i It should have a large interior area, i.e., the geographical location X. i The probability of species c appearing will be related to the inner product f(X). i )*O k(：，c) Positive correlation. Therefore, the Sigmod() nonlinear function is used to transform f(X) into a nonlinear function. i )*O c The range is mapped to the 0-1 range, representing X. i The posterior probability P(y|X) of species c existing at location c. i During training, the cross-entropy loss function and gradient descent method are used to obtain the embedding model f() and the classification layer model O(). k Parameter estimation. Calculate the classification loss at each of the three levels. in For k-level category prediction, y k Let be the truth value of the category at level k.

[0057] Using a hierarchical species label structure to train the metadata classification layer model has the advantage that closely related species (such as different species within the same genus) will have a higher affinity. When there are few fossil samples in the training set, the excavation location of a single fossil as input can indicate the possibility of multiple related fossil discoveries, thus improving the robustness of the model.

[0058] 1) Joint training of classification embedding models. The weighted sum L of the cross-entropy loss function and the regularization term is minimized using gradient descent. taxa +γL tco , where γ is the weighting coefficient. Training is stopped and the classification embedding model weights are saved when the loss function on the validation set no longer decreases significantly.

[0059] Step 4: Build and train a single image classifier for graptolite fossils, such as... Figure 4 As shown. The specific steps are as follows:

[0060] 1) Since the sample size of the graptolite fossil image dataset is small, the individual image classifier of graptolite fossils needs to be pre-trained on a large-scale dataset (such as ImageNet). In this example, the neural network ResNet50 is used as the individual image classifier of graptolite fossils.

[0061] 2) Figure 4 This describes the training process of a classifier for individual graptolite fossil images. Based on a pre-trained classifier, training is performed to classify graptolite fossil images at the "species" level. Training stops when the accuracy on the validation set does not show a significant improvement, and the weights of the classifier are saved.

[0062] Step 5: Test the classification layer model for individual graptolite fossil images and obtain the category prediction results for individual graptolite fossil images. The specific steps are as follows:

[0063] 1) Take images of the individual graptolite fossils to be classified and obtain their corresponding geographical locations using a positioning system. If the image contains multiple individual graptolite fossils, manual segmentation is required to obtain individual images of the graptolite fossils.

[0064] 2) Input the single image of the graptolite fossil into the single image classifier trained in step four to obtain the category prediction result P(y|I) of the single image of the graptolite fossil. i ).

[0065] 3) Input the metadata corresponding to the single image of the graptolite fossil into the metadata embedding classification model to obtain the category prediction result P(y|X) of the metadata. i );

[0066] 4) The final prediction is obtained by fusing the results of the metadata embedding classification model and the graptolite fossil individual image classifier. Common sense dictates that, in most cases, species category and the observed geographical location of an individual organism are not precisely correlated. Therefore, it can be assumed that the species' visual characteristics and the geographical location where they were discovered are approximately conditionally independent. According to Bayes' theorem: Since P(y|I,X) is proportional to P(y|I)P(y|X), the final species probability ranking can be obtained by simply multiplying the results of the two classifications. Figure 5 As shown, during the testing process, a classifier for individual graptolite fossil images was used to obtain classification results at the "species" level using individual graptolite fossil images. Using a metadata embedding classification model, we obtain classification results at the "species" level of metadata. The prior classification result of the metadata for the final graptolite fossil image is obtained by multiplying the two results.

[0067] The simulation experiment of the method of the present invention is as follows:

[0068] This experiment selected 113 classes from the graptolite fossil image dataset, totaling over 1100 fossil specimens and 34,000+ individual fossil images along with their corresponding metadata. 31,562 images and their corresponding metadata were used as the training set, and the remaining 3,051 images were used as the test set. Top-1, Top-3, and Top-5 accuracies were used to test the performance of the metadata-prior graptolite fossil individual image classifier. ResNet50 was used as the graptolite fossil individual image classifier, and the training of the classification model was conducted with different data constraints on metadata embedding. The changes in classification accuracy were observed, and the experimental results are as follows: Figure 3 As shown, AUG represents metadata augmentation, LOC represents adding geolocation to predict the category of graptolite fossil images, CO represents adding co-occurrence relationships of graptolite fossil species as regularization, and F, G, and S represent using family, genus, and species-level labels to constrain the training of the metadata embedding classification network, respectively. The improvement in classification accuracy of graptolite fossil images under different metadata embedding constraints was compared. Only family-level label constraints had no significant effect on improving classification accuracy. The metadata prior method significantly improved accuracy compared to graptolite fossil image classification using only image information. This is because using metadata information beyond the image further narrows the possible category range. Simultaneously, when performing geolocation embedding, the inter-species relationships were considered, improving the model's robustness when the number of specimens is insufficient. Finally, compared to classification using only graptolite fossil images, the Top-1 accuracy of graptolite fossil image classification was improved by 9.8%.

[0069] As another corresponding embodiment, and a corresponding device embodiment, its hardware structure is as follows: Figure 7 As stated above.

[0070] Figure 7 In this embodiment, the execution entities of processor 1 and memory 2 can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementation, this embodiment of the invention does not limit the execution entity and selects it according to the needs of actual application.

[0071] Data signals are transmitted between memory 2 and processor 1 via bus 3, which will not be described in detail in this embodiment of the invention.

[0072] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0073] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0074] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0075] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0076] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0077] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a USB flash drive, external hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0078] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0079] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of classifying a graptolite fossil image, characterized by, The method comprises the following steps: Step one: preparing the graptolite fossil image dataset, the method is as follows: 1) Collect graptolite fossil original specimens by taking pictures; 2) Obtain the geographical position of the graptolite fossil original specimen excavation site by using a positioning system; Determine the geological era of the graptolite fossil original specimen; Take the geographical position and geological era of the graptolite fossil original specimen as the corresponding metadata of the graptolite fossil original specimen, and construct the graptolite fossil image dataset; 3) Divide the graptolite fossil image dataset into a training set and a test set, so that the test set contains at least one graptolite fossil original specimen; 4) Segment each graptolite fossil monomer image in the graptolite fossil original specimen image, so that each graptolite fossil monomer image inherits the category, metadata and dataset division of the graptolite fossil original specimen; 5) Label the category of the graptolite fossil monomer image, including three levels of family, genus and species, to form a data pair: {graptolite fossil monomer image, metadata, category}; Step two: building and training the metadata embedding classification model, the metadata embedding classification model comprises an embedding model and a classification layer model, and the method is as follows: 1) Build the embedding model, the embedding model is a neural network composed of multiple fully connected layers with different output dimensions, and the input of the embedding model is the geographical position of the graptolite fossil original specimen and the output is the embedding vector of the geographical position of the graptolite fossil original specimen; 2) Based on the geological era and category corresponding to the graptolite fossil monomer image, a regularization term is constructed by using the co-occurrence relationship to adjust the Euclidean distance between embedding vectors; 3) Build the classification layer model, the input of the classification layer model is the embedding vector of the geographical position, and the output is the prediction result of the three levels of family, genus and species; the cross-entropy loss of the prediction results of the three levels of family, genus and species is calculated respectively; 4) The parameter estimation of the embedding model and the classification layer model is obtained by minimizing the weighted sum of the cross-entropy loss and the regularization term by using the gradient descent method; Step three: building and training the graptolite fossil monomer image classifier, the method is as follows: 1) Pre-training is performed to obtain a pre-trained graptolite fossil monomer image classifier; 2) Based on the pre-trained graptolite fossil monomer image classifier, the classification of the "species" level of the graptolite fossil monomer image is iteratively trained, and the weight of the iteratively trained graptolite fossil monomer image classifier is saved; Step four: input the graptolite fossil monomer image in the test set and the corresponding metadata into the graptolite fossil monomer image classifier and the metadata embedding classification model respectively, and fuse the outputs of the two to obtain the graptolite fossil monomer image category prediction result.

2. The method of classifying a graptolite fossil image according to claim 1, wherein, After step one and before step two, the graptolite fossil monomer image and metadata in the training set are also subjected to augmentation processing.

3. The method of classifying a graptolite fossil image according to claim 2, wherein, The graptolite fossil monomer image in the training set is subjected to color random floating augmentation processing.

4. The method of classifying a graptolite fossil image according to claim 2, wherein, The graptolite fossil monomer image in the training set is subjected to rigid body change or affine transformation augmentation processing.

5. The method of classifying a graptolite fossil image according to claim 2, wherein, The graptolite fossil monomer image in the training set is subjected to row optical distortion, elastic distortion or grid distortion augmentation processing.

6. The method of classifying a graptolite fossil image according to claim 2, wherein, The graptolite fossil monomer image in the training set is subjected to Cutout augmentation processing.

7. The method of classifying a graptolite fossil image according to claim 2, wherein, Noise is added to the geographical position in the metadata for augmentation processing.

8. The method of classifying a graptolite fossil image according to claim 2, wherein, Gaussian noise is added to the graptolite fossil monomer image in the training set for augmentation processing.

9. A device for classifying a graptolite fossil image, characterized by, The apparatus comprises a processor and a memory having program instructions stored therein, the processor invoking the program instructions stored in the memory to cause the apparatus to perform the method steps of any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program comprising program instructions, the program instructions being executed by a processor to cause the processor to perform the method steps of any one of claims 1-8.

Images