Multi-tiered marine organism classification method

Abstract

The application discloses a marine organism classification method based on multiple hierarchies, and adds a multiple hierarchy classification system with biological prior knowledge to a network model. The method mainly comprises three parts of hierarchy division, a multiple hierarchy network model and multiple hierarchy classification. Marine organisms are processed in hierarchies through hierarchy division, details of features of the same hierarchy are extracted by using the multiple hierarchy network, and the extracted hierarchical information is combined with a risk minimization strategy to optimize a classification result. The multiple hierarchy network model mainly designs a C-MBConv module and a fine classification module, so that the network model is more suitable for a multiple hierarchy classification task. In order to reduce error transmission caused by high-order categories, a risk minimization strategy is designed in the multiple hierarchy classification part, so that the possibility of error transmission is greatly reduced. The method can effectively improve the classification precision and has good application effect on data with multiple hierarchical categories.

Description

Technical Field

[0001] This invention relates to the field of marine biological classification, and more specifically to a multi-level marine biological classification method. Background Technology

[0002] With the continuous advancement of marine development, the significance of marine biological classification is becoming increasingly apparent. Manually distinguishing fish requires extensive and tedious inspection work, demanding skilled marine biologists. Due to the open ocean environment, complex backgrounds, and low light levels, marine observation is a challenging task. Utilizing expert resources is costly, and the large workload increases the risk of misidentification. Automating marine biological classification and identification through technology will contribute to the further development of marine biological science; automated systems can effectively assist marine biologists in classifying various marine organisms.

[0003] Currently, there are few applications capable of real-time monitoring of marine life and classifying images captured in the ocean. Such systems could be used for marine monitoring activities, such as assessing population sizes, classifying common local marine species, and tracking fish migrations. Accurate identification of marine species would be extremely helpful to researchers, marine scientists, and biologists, and could also aid in determining marine biomass levels and geological changes.

[0004] The poor quality, low brightness, and complex backgrounds of videos taken in the ocean make the task of classifying marine life habitats challenging. Currently, marine populations are subject to significant changes due to environmental and human factors, such as global warming, sea-level rise, marine pollution, overfishing, and overexploitation of marine resources.

[0005] These factors have further driven the development of a standardized, cost-effective, and reliable method for monitoring marine life across the entire ocean habitat. With the advancement of deep learning, image classification has long been a hot research area, and convolutional neural networks do not require explicit feature extraction methods. Many excellent convolutional neural networks perform exceptionally well in classification tasks, such as EfficientNetV2.

[0006] Most techniques in marine organism classification tasks are consistent with general classification, completing the task through a single classification. However, marine organisms are diverse and difficult to identify, so a single classification is insufficient for accurate classification. Summary of the Invention

[0007] This invention addresses the problem of existing algorithms neglecting species hierarchy rules by proposing a multi-level marine organism classification method based on biological prior knowledge. By adding a multi-level category system of biological prior knowledge to the network model, hierarchical rules are learned to improve classification accuracy. To better suit neural networks for general classification tasks and improve recognition accuracy, the invention optimizes the network by designing a new module for multi-level classification tasks, called the C-MBConv block, and a new network component for second-order classification, the fine-grained classification module. Combining the proposed modules and components, the network architecture C-EfficientNetV2 for this multi-level classification task is generated.

[0008] Multi-level classification tasks differ significantly from traditional classification tasks. Traditional classification tasks require only a single classification operation, but multi-level classification necessitates multiple classifications. This introduces significant risk into multi-level classification; if the initial classification is incorrect, subsequent classifications will build upon the previous error, leading to inaccurate predictions. In other words, multi-level classification inherently carries risk, which propagates to subsequent classifications. To minimize the risk of misclassification in subsequent decision-making tasks, a risk-minimizing model is designed.

[0009] This invention is achieved using the following technical solution:

[0010] Based on a multi-level marine organism classification method, the following steps are performed to obtain a model for image-based marine organism classification:

[0011] 1) Collect image data of intertidal marine organisms, classify them into hierarchical labels according to categories, and obtain a labeled sample dataset {image data, hierarchical labels}; the hierarchical labels are first-order categories and second-order categories;

[0012] 2) Establish a multi-level backbone C-EfficientNetV2 feature network to extract features from the sample dataset and output a hierarchical information feature map representing the correlation between categories;

[0013] 3) The hierarchical information features are classified sequentially using a multi-level classification network to obtain the accurate classification result of the sample ocean image; the multi-level classification network includes a first-order primary category discrimination unit, a minimum risk strategy unit, and a second-order fine category discrimination unit;

[0014] 4) Iterate repeatedly using the sample dataset to train the multi-level backbone C-EfficientNetV2 feature network and the multi-level classification network to obtain the trained and optimized biological category discrimination model.

[0015] The categorization of hierarchical labels involves using prior biological knowledge to divide different hierarchies and using clustering metrics to ensure sample balance for each first-order category, as well as category balance.

[0016] The multi-level backbone C-EfficientNetV2 feature network includes sequentially connected convolutional layers, several C-MBConv module groups, and fully connected layers; the C-MBConv module groups consist of unequal numbers of C-MBConv modules and are used to extract correlations between categories level by level.

[0017] The C-MBConv module structure includes a 1x1 convolutional layer, an attention layer, a depthwise separable convolutional layer, a 3x3 convolutional layer, and a 1x1 convolutional layer connected in sequence, which extracts channel features and spatial features related to the categories.

[0018] The first-order primary category discrimination unit is used to perform first-order category classification by utilizing the fine-scale features extracted by the multi-level network, and to generate first-order categories of species.

[0019] The minimum risk strategy unit uses a Bayesian method to remodel the output of the backbone network, estimates the confidence level of the prediction results, and calculates the risk minimization decision parameters:

[0020]

[0021] in, For the network's possible decisions for all first-order categories, For the Bayesian posterior probabilities of all possible decisions, and This represents the corresponding loss value.

[0022] The second-order fine-scale classification unit consists of a fine-scale classification module and a second-level classifier. The fine-scale classification module combines the feature maps in the backbone network and uses the first-order categories generated by optimization decisions as an aid to generate fine-scale feature maps for second-order classification. The second-level classifier performs second-order classification on the biological fine-scale features extracted from the current image based on the second-order label categories.

[0023] The fine category classification module includes:

[0024] Let the data set be Where x represents the image of a marine organism, and y represents the category label, with the superscript indicating the category level. Let N be the first-order category of the i-th image. 1 and N 2 These represent the number of first-order and second-order categories, respectively; f i For image x iImage features extracted by the backbone network C-EfficientNetV2;

[0025] Define parameters E i The weights for predicting the second-order class of the i-th image; when the first-order class of the i-th image is predicted to be k. Its definition is as follows:

[0026]

[0027] Among them, W cls1 For the weights of the first-order classifier, the image features f i Further reasoning leads to the Bayesian posterior probability, obtaining a confidence estimate for the first-order category, and thus deriving the risk-minimizing decision.

[0028] use Image features f i and The weights F for entering the fine classification module are obtained by combining these factors. i :

[0029]

[0030] Among them, f i j The superscript represents the weight corresponding to the first-order category, and α is the weight for minimizing risk;

[0031] The fine-grained classification module will further process the features F that have already acquired first-order classification knowledge. i Further combining coarse and fine granularity characteristics, F is obtained after C-MBConv processing. i ′, using E i With F i The final output D is obtained. i :

[0032]

[0033] The D i The data is fed into a second-order classifier to generate second-order categories, thus completing the multi-level classification task.

[0034] The trained and optimized biological category discrimination model is used to process images of marine organisms to be identified and automatically output the hierarchical categories of organisms in the current image.

[0035] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0036] This scheme designs a multi-level classification method, aiming to add a multi-level category system of biological prior knowledge to the network model, thereby mimicking the human process of species identification and making the classification results more accurate. Furthermore, the involved C-MBConv module and fine-grained classification help the network extract hierarchical information from images, making the network more adaptable to multi-level tasks. The multi-level marine organism classification method proposed in this scheme can effectively improve the classification performance of marine organisms and also shows good application results in datasets with multiple hierarchical levels. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the hierarchical division of the marine biological dataset proposed in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the C-MBConv module proposed in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of a multi-layered network proposed in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram illustrating the multi-level classification of an embodiment of the present invention; Detailed Implementation

[0041] To better understand the above-described objects, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention; however, the present invention may be practiced in other ways than those described herein, and therefore, the present invention is not limited to the specific embodiments disclosed below.

[0042] This proposal suggests a multi-level marine organism classification method. For complex marine organism datasets, it adds a multi-level category system based on prior biological knowledge to the network model, enabling the network to learn the hierarchical rules of the dataset. The described classification method mainly includes hierarchical partitioning, a multi-level network, and a multi-level classification component. Hierarchical partitioning categorizes marine organisms into different levels, and the multi-level network extracts detailed feature relationships within the same category. The extracted hierarchical information is combined with a risk minimization strategy to optimize the classification results, thereby improving classification accuracy.

[0043] This invention enhances the learning of similarities and differences between species through a multi-level training scheme. The original linear classification task is designed into two sub-tasks: first-order class classification and second-order class classification. The multi-level classification method also conforms to human object identification rules; for example, prior knowledge of a fish in water leads to posterior knowledge of a shark. The C-MBConv module can more effectively learn the correlations and differences between categories and extract hierarchical category features more effectively. A risk minimization strategy uses Bayesian methods to revise the first-order classification decision, minimizing the propagation of errors caused by first-order decision mistakes and making the classification results more accurate. The fine-grained classification module optimizes the first-order decision and combines coarse-grained and fine-grained features to complete the final second-order class classification.

[0044] Specifically, the following section provides a detailed introduction to the principles of each module in the multi-level marine organism classification method and the method of this scheme, mainly including the following steps:

[0045] Step A: Classify the collected dataset of intertidal marine life from the Nanji Islands by adding hierarchical categories on top of the original basic categories. The hierarchical categories are divided into first-order and second-order.

[0046] Specifically, the main purpose of this process is to simulate the human brain's rules for species recognition. In marine organisms, there is a clear problem of strong inter-class similarity within the same phylum and significant differences between different phyla. The hierarchical classification task in step A refers to grouping marine organisms with strong inter-class similarity into one group, such as squid and cuttlefish, while groups with greater differences will exist in different groups, such as squid and lobster. Each group will be called a first-order category, and the subcategories forming a first-order category will be called a second-order category. After the classification, some typical hierarchical clustering problems may be encountered, so further processing is required. After using prior biological knowledge to classify different levels, clustering indices are used to ensure sample balance within each first-order category and category balance.

[0047] Class balance refers to ensuring a balance in the number of second-order classes among the first-order classes. Class imbalance leads to severe data imbalance problems, thus affecting the performance of the final classifier. To avoid class imbalance, parameter B is defined. class To evaluate class balance, its definition is shown in formula (2). Furthermore, it is necessary to ensure not only class balance but also sample balance. If the number of samples in each first-order class differs significantly, it will severely affect the classifier's training and classification performance. The parameter B is defined as follows: sample Used to assess sample balance, its definition is shown in formula (3). B class With B sample The balance parameter B(c) is generated to ensure both sample balance and class balance. The definition of B(c) is as follows (1):

[0048]

[0049] The smaller the value of B(c), the better the balance of the samples, meaning the generated hierarchical categories are reasonable. In this invention, a threshold for B(c) is defined; when the threshold requirement is met, the categories are considered reasonable.

[0050]

[0051] Where c represents the number of first-order categories, b a b represents the average number of second-order categories among all first-order categories being evaluated. i B represents the total number of second-order categories in the i-th first-order category. class This represents the class balance of the first-order categories; the larger this parameter is, the more unbalanced the class is.

[0052]

[0053] Among them, s i Let s be the number of samples in a first-order class. a This represents the average number of samples in the first-order class. To ensure both sample balance and class balance, these need to be combined to generate the balance parameter B(c). An example of a partitioning diagram for a marine biological dataset is shown below. Figure 1 As shown.

[0054] Step B: Extract features from the hierarchical dataset using a multi-level network model, and obtain the hierarchical information features of the image using C-MBConv.

[0055] Specifically, since general network models are not suitable for this multi-level task, this method improves EfficientNetV2, which has excellent performance in many classification tasks, by designing the C-MBConv module to improve the network into C-EfficientNetV2, making it more suitable for multi-level classification tasks.

[0056] The design of the C-MBConv module is as follows: Figure 2As shown, the detailed structure is described below: After entering the C-MBConv module, the features first undergo a 1x1 convolution to expand the feature map along the channel dimension. The features then enter the SE channel attention module to enhance channel features, learn the relationships between different channels, and improve feature representation. Subsequently, they undergo depthwise separable convolution, specifically a depthwise convolution that performs independent convolution operations on each channel, with pointwise convolutions weighting the combinations along the depth direction. This significantly enhances the model's ability to learn inter-channel dependencies and spatial encoding, allowing for deeper correlations between channels. The feature maps of different channels are readjusted and combined to extract both channel and spatial features, enhancing the model's expressive power. Next, a 3x3 convolution simultaneously learns spatial and channel features, extracting features and further enhancing the model's generalization ability. Finally, a 1x1 convolution scales the feature map along the channel dimension, completing feature extraction and outputting fine-scale features extracted from the biological data in the image.

[0057] Based on the proposed C-MBConv block, the EfficientNetV2 network architecture within the NAS framework was adjusted, and the revised network was named C-EfficientNetV2. The C-EfficientNetV2 network consists of the following... Figure 3 As shown, C-EfficientNetV2 will mainly consist of C-MBConv blocks. First, the feature map will be expanded by a regular convolution, and then it will be passed through six stages composed of different numbers of C-MBConv blocks in succession, with the number of blocks being 2, 4, 4, 6, 9, and 15 respectively. The last stage is a fully connected layer (FC).

[0058] Step C utilizes hierarchical information features to perform multi-level classification, classifying sequentially according to the multi-level subcategories.

[0059] Step C1, First-order primary category discrimination unit: First-order category classification is performed using the fine-scale features extracted by the multi-level network (these features are used to index into the clearly defined first-order category in the hierarchical category) to generate the species' first-order category. The first-order classifier performs first-order classification based on the biological fine-scale features extracted from the current image according to the first-order label category.

[0060] Step C2, Minimum Risk Strategy: Combining the generated first-order categories, the minimum risk strategy is used to optimize the first-order decision classification. The optimized decision results are combined with the fine-scale features extracted by the multi-level network to form more detailed features with hierarchical information.

[0061] Specifically, because multi-level classification requires multiple classification steps, it carries significant risk. If the initial classification is incorrect, subsequent classifications will build upon the previous error, leading to errors in the overall prediction. For example, a sample's second-order category might be "purple sea urchin," but its actual first-order category is Echinoderms. However, if it is incorrectly classified as Arthropoda, subsequent classification tasks will determine its second-order category within Amputus. In other words, multi-level classification inherently carries risk, and this risk is propagated to subsequent classifications.

[0062] To minimize the risk of misclassification in subsequent decision-making tasks, a risk-minimizing model was designed. A confidence estimate for hierarchical classification was added to the designed C-EfficientNetV2 network. A Bayesian method was used to remodel the output of the new network, providing a confidence estimate of the prediction results instead of a single value. In summary, the definition of risk-minimizing decision-making is given by formula (4):

[0063]

[0064] in For the network's possible decisions for all first-order categories, For the Bayesian posterior probabilities of all possible decisions, and This represents the corresponding loss value. A first-order classification decision is made using the designed network model, and then Bayesian inference is used to extend the model's decision results, achieving risk minimization control of the decision.

[0065] Step C3, Second-order fine-grained category discrimination unit: It receives fine features with hierarchical information and uses coarse-scale features and fine-scale features to make a decision to generate the final second-order category, thus completing the multi-level classification task.

[0066] Specifically, the main task of the fine-grained classification feature module is to combine the feature maps in the backbone network and use the first-order categories generated by optimization decisions as an aid to generate fine-grained feature maps for second-order classification. This further improves the accuracy of second-order category classification. The fine-grained classification feature module consists of a deconvolution layer for concatenating coarse and fine-grained features, a C-MBConv, and a second-order category classifier. The feature maps are finally processed by the second-order category classifier to generate second-order final categories. The second-order category classifier performs second-order classification on the biological fine-scale features extracted from the current image based on the second-order label categories.

[0067] The detailed data flow is as follows, assuming the dataset is... Where x represents the image of marine life, and y represents the category label, with the superscript indicating the category level, such as... Let N be the first-order category of the i-th image.1 and N 2 These represent the number of first-order and second-order categories, respectively. i For image x i Image features extracted by the backbone network C-EfficientNetV2. Parameters defined. E i The weights used to predict the second-order class for the i-th image are primarily generated by the first-order classifier, and their main function is to strengthen the weights belonging to the same first-order class. When the predicted first-order class for the i-th image is k, Its definition is as follows:

[0068]

[0069] Among them, W cls1 For the weights of the first-order classifier, the image features f i Further reasoning leads to the Bayesian posterior probability, obtaining a confidence estimate for the first-order category, and thus deriving the risk-minimizing decision. use Image features f i and Combined to obtain F i F i For the weights to enter the fine classification module, F i The definition is as follows:

[0070]

[0071] Among them, f i j The superscripts represent the weights corresponding to the first-order categories, and α is the weight for minimizing risk, which determines its impact on the feature map. The fine-grained classification module will further process the features F that have already acquired first-order classification knowledge. i Further combining coarse and fine granularity characteristics, F is obtained after C-MBConv processing. i ′, using E i With F i The final output D is obtained. i D i The definition is as follows:

[0072]

[0073] D i The data is fed into a second-order classifier to generate second-order categories, completing the multi-level classification task. The specific multi-level classification process is as follows: Figure 4 As shown.

[0074] Experimental verification:

[0075] The multi-level marine organism classification method trained in this embodiment can efficiently classify marine organism datasets, and its training results are compared with those of several mainstream model frameworks:

[0076] This embodiment primarily conducts experiments on two datasets. The first dataset is the Nanji Islands intertidal marine life dataset, provided by the Institute of Oceanology, Chinese Academy of Sciences. It contains 342 marine species (second-order categories), which can be categorized into 9 major classes (first-order categories), totaling approximately 34.2 million marine life images. The dataset structure is as follows... Figure 2 As shown, the training and test sets were divided in a 4:1 ratio. To verify the scalability of this method, the second dataset used was the public dataset CIFAR-100, which contains 100 categories and a total of 60M images. The 100 categories in CIFAR-100 are further divided into 20 superclasses, consistent with the multi-level classification setting of this experiment. The CIFAR-100 dataset is divided as shown in Table 1. It is important to note that for the CIFAR-100 test, we did not use pre-training on any other datasets for any of the models; instead, we trained them directly.

[0077] Our experiments resized the images to 256×256 and cropped them to 224×224 for training. We used several data augmentation methods, including Mixup, CutMix, label smoothing, and RandAugment. Mixup proportionally adds two samples to generate new samples and labels, with the labels modified proportionally. CutMix crops a random rectangular region from one image onto another to generate a new image, again using proportional methods to determine the new label's proportion. Labelsmoothing adds noise to the labels, reducing the weight of the true label category in the calculation.

[0078] The training batch size was set to 128, with 512 training epochs, and optimized using AdamW with an initial learning rate of 1x10⁻¹⁰. -3 Furthermore, cosine scheduling is used, with weight decay set to 0.05.

[0079] First, benchmark tests were conducted against the baseline model EfficientNetV2 to check the effectiveness of the simulation method. Table 1 shows the experimental results comparing the original network model with the network model applying the multi-level classification method. C-EfficientNetV2 is an improved network model using the designed C-MBConv module. Based on the experimental results of the Nanji Islands intertidal marine organism dataset, the improved model, with nearly one-third fewer parameters, achieved an accuracy of 95.14%, further demonstrating the effectiveness and efficiency of the module designed in this invention. Observing the results in the table, it can be seen that CM-EfficientNetV2 achieved the highest accuracy of 96.47%, an improvement of 1.5% over the original baseline network model. CM-EfficientNetV2 is the complete model designed and proposed in this scheme using a multi-level marine organism classification method, proving that a multi-level marine organism classification method can effectively improve the performance of the original baseline network.

[0080] Table 1. Experimental results comparing the proposed solution with the baseline network model.

[0081]

[0082] To further verify the effectiveness of the method of this invention, it was compared with several advanced methods to obtain more reliable comparison results: 1) Swin Transformer, which introduces locality of convolution operations into the Transformer through a sliding window operation, focusing attention operations within the window. 2) CoAtNet, which efficiently combines depthwise convolution and attention mechanisms, achieving excellent performance. 3) iFormer, which continues to address the problem of insufficient ability of the Transformer to capture local information by splitting the channel and frequency ramp structures to improve performance and efficiency. Table 2 shows the comparison results of the proposed solution.

[0083] Table 2 Comparison results of the present invention's solution with four types of advanced methods

[0084]

[0085]

[0086] As can be seen from the results in Table 2, the present invention achieves the best performance compared with the four advanced methods, and the performance improvement is very significant, which more convincingly proves the reliability of the method of the present invention.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multi-level marine organism classification method, characterized in that, The following steps are performed to obtain a model for image-based marine organism classification: 1) Collect image data of intertidal marine life, classify them into hierarchical labels according to categories, and obtain a labeled sample dataset {image data, hierarchical labels}; the hierarchical labels are first-order categories and second-order categories; 2) Establish a multi-level backbone C-EfficientNetV2 feature network to extract features from the sample dataset and output a hierarchical information feature map representing the correlation between categories; the multi-level backbone C-EfficientNetV2 feature network includes sequentially connected convolutional layers, several C-MBConv module groups, and fully connected layers; the C-MBConv module groups are composed of unequal numbers of C-MBConv modules and are used to extract the correlation between categories level by level; 3) A multi-level classification network is used to classify the hierarchical information features sequentially to obtain accurate classification results for the sample ocean images; the multi-level classification network includes a first-order primary category discrimination unit, a minimum risk strategy unit, and a second-order fine category discrimination unit; the second-order fine category discrimination unit consists of a fine category classification module and a secondary classifier. The fine category classification module combines the feature maps in the backbone network and uses the first-order categories generated by optimization decisions as an aid to generate fine feature maps for second-order classification; the secondary classifier performs second-order classification on the biological fine-scale features extracted from the current image based on the second-order label categories; the fine category classification module includes: Let the data set be Where x is an image of a marine organism, y represents the category label, and the superscript indicates the category level. Let i be the first-order category of the i-th image; and These represent the number of first-order and second-order categories, respectively. For image Image features extracted by the backbone network C-EfficientNetV2; Define parameters , , The weights for predicting the second-order class of the i-th image; when the first-order class of the i-th image is predicted as j=k. When the first-order category of the i-th image is predicted to be... hour Its definition is as follows: ； in, For the weights of the first-order classifier, image features Further reasoning leads to the Bayesian posterior probability, obtaining a confidence estimate for the first-order category, and thus deriving the risk-minimizing decision. ; use Image features and Combined, the weights for entering the fine classification module are obtained. : ； in, The superscript indicates the weight corresponding to the first-order category. To minimize risk weights; The fine-grained classification module will further process the features that have already acquired first-order classification knowledge. Further combining coarse and fine granularity characteristics, after C-MBConv processing, the following was obtained: ,use and Get the final output : ； The The data is fed into a second-order classifier to generate second-order categories, thus completing the multi-level classification task. 4) Iterate repeatedly using the sample dataset to train the multi-level backbone C-EfficientNetV2 feature network and the multi-level classification network to obtain the trained and optimized biological category discrimination model.

2. The multi-level marine organism classification method according to claim 1, characterized in that, The categorization of hierarchical labels involves using prior biological knowledge to divide different hierarchies and using clustering metrics to ensure sample balance for each first-order category, as well as category balance.

3. The multi-level marine organism classification method according to claim 1, characterized in that, The C-MBConv module structure includes a 1x1 convolutional layer, an attention layer, a depthwise separable convolutional layer, a 3x3 convolutional layer, and a 1x1 convolutional layer connected in sequence, which extracts channel features and spatial features related to the categories.

4. The multi-level marine organism classification method according to claim 1, characterized in that, The first-order primary category discrimination unit is used to perform first-order category classification by utilizing the fine-scale features extracted by the multi-level network, and to generate first-order categories of species.

5. The multi-level marine organism classification method according to claim 1, characterized in that, The minimum risk strategy unit uses a Bayesian method to remodel the output of the backbone network, estimates the confidence level of the prediction results, and calculates the risk minimization decision parameters: ； in, Represents the input image of marine life; Represents the number of first-order categories; Represents category labels; For the network's possible decisions for all first-order categories; For the Bayesian posterior probabilities of all possible decisions, and This represents the corresponding loss value.

6. The multi-level marine organism classification method according to any one of claims 1-5, characterized in that, The trained and optimized biological category discrimination model is used to process images of marine organisms to be identified and automatically output the hierarchical categories of organisms in the current image.

Images