Marine organism classification method based on hierarchical neural network

Abstract

The application discloses a marine organism classification method based on a hierarchical neural network model, which automatically extracts high-quality features of data through the construction of the hierarchical neural network model, and solves the problem of marine organism recognition; the hierarchical neural network model comprises a visual tree construction module, a full-channel learning variable-dimension partial convolution module and a relative attention module; first, two hierarchical structures are constructed based on unsupervised learning and prior knowledge, and feature relationships and biological relationships are fitted respectively; a backbone network vertically designs a hierarchical neural network based on the full-channel learning variable-dimension partial convolution module and the relative attention module according to the complexity of features, coarse classification assists fine classification, marine organism classification is realized, and the scheme can improve the marine organism classification accuracy, and meanwhile, the balance of the accuracy and the parameter quantity and other network performances can be ensured.

Description

Technical Field

[0001] This invention relates to the field of marine organism classification, and specifically to a marine organism classification method based on hierarchical neural networks. Background Technology

[0002] Marine biological classification networks are important tools in biological research, enabling in-depth understanding of marine biodiversity, ecosystem health, and environmental monitoring. They are of great significance for promoting marine ecological research, protection, and management.

[0003] Most existing methods for classifying marine organisms simply apply traditional image classification techniques, neglecting the biological relationships inherent in the organisms themselves. Experts typically consider multiple factors when classifying marine organisms, such as hierarchical structure and morphological characteristics. While some algorithms attempt to incorporate these rules, the classification rules for marine organisms are extremely complex, involving both biodiversity and environmental specificity.

[0004] When constructing a marine life classification network, careful consideration must be given to how to embed these classification rules into the algorithm. This may require fine-tuning and optimization of the algorithm to adapt to different types of marine life and image conditions. Furthermore, it is necessary to select an appropriate artificial intelligence algorithm model for classification and optimize the model to improve classification accuracy and efficiency. This means in-depth research into the applicability of the dataset and the generalization ability of the model is needed to ensure that the algorithm can accurately classify marine life images under various conditions.

[0005] In addition to considering external factors, it is also necessary to pay attention to the complexity and robustness of the algorithm itself. In the marine environment, images may be affected by factors such as light and water quality, so the algorithm needs to have a certain level of complexity and robustness to work accurately under various conditions.

[0006] Of particular note is the higher complexity and robustness requirements of algorithms in marine organism classification tasks. These algorithms need to accurately classify marine organisms within complex biodiversity contexts, taking into account potential ecosystem changes and uncertainties. Existing classification methods struggle to meet these requirements, thus necessitating the development of novel algorithms to achieve accurate marine organism image classification and provide robust support for marine biological research. Summary of the Invention

[0007] The Marine Organism Hierarchical Classification Network is an innovative technology designed to address the problem of existing algorithms neglecting biological relationships in marine organism image classification. This network, based on a conventional neural network structure, mimics the rules used by the public when observing marine organism images to improve classification performance. This invention applies a hierarchical structure to solve the problem of traditional algorithms ignoring biological relationships in marine organism image classification. The classification task is defined as three sub-tasks: phylum, species, and feature layers. The corresponding phylum, species, and features are based on the rules and hierarchical thinking used by humans when observing marine organisms. A full-channel learning variable-dimensional partial convolutional module and a relative attention module are allocated based on the difficulty of feature extraction to weight the classification results at each level. Coarse-class results are used to assist fine-classification, achieving the fusion of biological relationships and features to improve the model's accuracy and robustness, thus enhancing classification performance. Based on the feature extraction modules at each level, multiple categories assist each other, and the classification results of the three sub-tasks at their respective levels are fed back to each other to achieve marine organism classification.

[0008] This invention is achieved using the following technical solution:

[0009] A marine organism classification method based on hierarchical neural networks obtains a model for marine organism classification by performing the following steps:

[0010] Step 1) Collect marine life image data and perform image processing; establish a sample dataset;

[0011] Step 2) Generate a classification visual tree using prior knowledge of marine biological relationships and unsupervised learning methods;

[0012] Step 3) Based on the visual tree, construct the EAHNet backbone neural network, which includes a shared layer, independent layers, and a classifier, to classify the processed marine organisms using biological relationships as an aid. The shared layer sub-task classification is used to further extract coarse-scale features from the processed image. The independent layer sub-task classification includes: a phylum layer, a feature layer, and a species layer. Each independent layer takes the shared coarse-scale features as input and extracts the corresponding fine-scale features after passing through each layer. The classifier weighted sums of the fine-scale features from each layer to generate complete classification labels.

[0013] Step 4) Iterate repeatedly using the sample dataset to train the EAHNet backbone neural network and obtain the trained and optimized marine organism classification model.

[0014] The image processing includes image noise reduction and size normalization.

[0015] The biological classification visual tree is a hierarchical structure defined by hierarchical clustering; the unsupervised learning method uses a convolutional neural network to extract biological feature relationships from image data, calculates the optimal number of clusters, and divides the hierarchical structure according to the optimal number of clusters.

[0016] The optimal number of clusters, δ(nc)×BICScore, is calculated by combining the confidence upper bound algorithm and the Bayesian information criterion to achieve the balance of the tree structure.

[0017] BIC Score:

[0018] bic_score = 2loss × N val +nc×logN val (4)

[0019] in, n is the number of samples, x i It is the i-th original data. It is x i After self-encoding reconstruction, p i N represents the probability that the current data belongs to the i-th category. val This represents the number of samples in the validation set, and nc represents the number of clusters.

[0020] The equilibrium parameters are:

[0021]

[0022] Where, parameter nc represents the number of clusters in the clustering result, r E r represents the average number of clusters in the current clustering result. k m represents the number of subclasses contained in the k-th coarse class. E m represents the average number of samples in all coarse clusters in the clustering results. k This represents the number of samples contained in the k-th cluster in the current clustering result.

[0023] The shared layer is the full-channel learning variable-dimensional partial convolutional module EAConv, which includes a partial convolutional module, an SE module, and a pointwise convolutional module, extracting coarse-scale features and inputting them into each independent layer.

[0024] The independent layer employs a relative attention module to extract fine-scale features from coarse-scale features, thereby capturing biological characteristics.

[0025] In the classifier structure, the fine-scale features generated by each independent layer are focused on and weighted according to the biological information structure of each layer, and are used to finally summarize the biological information of different structures to generate complete classification labels, thereby realizing the classification of marine organisms.

[0026] The classifier is a softmax classifier;

[0027] After generating fine-scale features in each independent layer, the results are fed into Softmax to obtain the classification results for each layer. Based on the biological information structure of each layer, the three results are weighted and focused to generate the final classification labels, thus achieving marine organism classification. The focus and weighting formulas are as follows:

[0028] F = αF1 + βF2 + γF3

[0029] Where F represents the final classification result, F1, F2, and F3 represent the classification results of the phylum, fine layer, and coarse layer, respectively, and α, β, and γ represent the weights assigned to each layer.

[0030] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0031] This scheme constructs a hierarchical neural network model to achieve accurate classification from images of various marine organisms. Inspired by key rules in artificial marine organism classification annotations, this model simulates the entire marine organism classification process by mimicking these rules. Furthermore, a full-channel learning variable-dimensional partial convolutional module is proposed to achieve efficient feature extraction from marine organism data. Additionally, a novel optimal clustering calculation method is proposed to address the problem of imbalanced hierarchical structure construction and to reasonably reflect biological relationships. This scheme can improve the performance of marine organism classification and achieves good application results. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the marine organism classification method described in an embodiment of the present invention;

[0033] Figure 2 This refers to the imbalance that occurs during the construction of the hierarchical structure in the embodiments of the present invention.

[0034] Figure 3 This is a schematic diagram illustrating the principle of the hierarchical neural network described in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram illustrating the hierarchical neural network and its modules as described in an embodiment of the present invention; Detailed Implementation

[0036] 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.

[0037] This proposal suggests a marine organism classification method based on a hierarchical neural network model. The model is inspired by key rules in artificial marine organism classification annotations, simulating the entire classification process by mimicking these rules. The hierarchical structure is constructed using prior knowledge and unsupervised learning based on hierarchical Bayesian information criteria. The hierarchical neural network model includes a full-channel learning variable-dimensional partial convolutional module and a relative attention module. The hierarchical structure utilizes biological relationships to assist and resolve data imbalance. The full-channel learning variable-dimensional partial convolutional module extracts coarse-scale features for subsequent task sharing. By introducing a relative attention mechanism, subsequent tasks are classified into three independent layers: phylum, species, and feature layers. Each layer is responsible for the categories at these three levels, achieving joint classification of coarse and fine categories, improving the model's accuracy and robustness, and enhancing classification performance. Based on the feature extraction modules at each level, multi-category mutual assistance is achieved, with the classification results of the three sub-tasks feeding back to each other to realize marine organism classification.

[0038] Specifically, the principles of each module of the hierarchical neural network model and the method of this scheme will be introduced in detail below, combined with, for example... Figure 1 As shown, it includes the following steps:

[0039] Step A: The raw data was processed to ensure the consistency and quality of the input data. Specific steps included:

[0040] Denoising, resizing, normalization, etc.;

[0041] Step B: Using prior knowledge of marine biological relationships and unsupervised learning methods, two hierarchical structures are generated. The unsupervised learning method uses a convolutional neural network to extract feature relationships from image data. The features are used to calculate the optimal number of clusters. The hierarchical structure is constructed through hierarchical clustering, that is, clustering is performed according to the optimal number of clusters. The number of clusters obtained by clustering is equal to the optimal number of clusters. Then, the optimal number of clusters for each cluster is calculated. The above operation is repeated until all clusters contain only one fine class, and finally the hierarchical structure is constructed.

[0042] Step C: After processing in Step A, the size of the marine biological data is uniformly set to 224×224, and the internal pixel values ​​are uniformly set to the interval [0, 1]. This data is then classified, with the classification task defined as a shared layer subtask and an independent layer subtask. The corresponding subtasks are responsible for data transfer between the shared and independent layers to achieve the purpose of biological relationship-assisted classification, including:

[0043] Step C1: For the shared layer subtask, this subtask mainly uses the processed marine life image, and utilizes two consecutive full-channel learning variable-dimensional partial convolution modules. After pointwise convolution, SE module, and regular convolution, coarse-scale features are extracted for subsequent independent layer sharing.

[0044] Step C2: For the independent layer subtask, the independent layers are divided into a phylum layer, a feature layer, and a species layer. Each independent layer shares coarse-scale features. After the feature map passes through each layer, fine-scale features are extracted. The fine-scale features from each layer are weighted and summed to generate complete classification labels. The coarse-scale features extracted from the shared layer are input into each independent layer. In each independent layer, the coarse-scale features are further processed using a full-channel learning variable-dimensional partial convolutional module or a relative attention module, depending on the difficulty and requirements of the task, to extract fine-scale features from the coarse-scale features, capturing more details and differences in biological features.

[0045] The identification difficulty of the species layer is significantly higher than that of other independent layers. The phylum layer has the fewest categories and the lowest identification difficulty. To balance the recognition performance and the number of parameters, we designed the independent layers in order of difficulty from hardest to easiest: "relative attention module + relative attention module", "full-channel learning variable-dimensional partial convolution module + relative attention", and "full-channel learning variable-dimensional partial convolution module + full-channel learning variable-dimensional partial convolution module".

[0046] Step D: After generating fine-scale features for each independent layer, the data enters the Softmax layer to obtain the classification results. Based on the biological information structure of each layer, the three results are weighted and considered to ensure that biological information with different structures is appropriately taken into account, and the final classification labels are generated, achieving marine organism classification. The weighting and consideration formulas are as follows:

[0047] F = αF1 + βF2 + γF3

[0048] Where F represents the final classification result, and F1, F2, and F3 respectively represent Figure 3 The classification results of the phylum layer, fine layer (i.e., feature layer), and coarse layer (i.e., species layer) are shown. α, β, and γ represent the weights assigned to each layer, and the specific values ​​are determined based on the experimental results in Table 2.

[0049] For the marine organism images to be processed, a series of standardized preprocessing steps are required to ensure image consistency and quality before inputting them into the network. This embodiment defines the marine organism classification task as three independent layers, such as... Figure 3 (a) Each independent layer is responsible for different levels of classification: species classification, phylum classification, and classification by characteristics. Unlike experts who typically classify the three levels one by one in a specific order, the marine biological classification model simulates a hierarchical classification approach that allows each independent layer to classify according to its own rules, while the species layer makes full use of information from other independent layers.

[0050] Specifically, as described below, this is the hierarchical classification principle to be simulated:

[0051] (1) First, based on the fine-scale characteristics, we will classify the phyla of organisms and identify and classify the phyla to which the marine organisms belong.

[0052] (2) At the same time, individuals are classified according to their characteristics.

[0053] (3) Based on the classification of biological phyla, marine organisms are further subdivided into different species by combining the characteristic relationships between organisms.

[0054] Therefore, this embodiment constructs three independent biological taxonomic layers, such as Figure 3 (c) After passing through a shared feature extraction layer, marine biological data enters three separate classification layers. Each classification layer is specifically responsible for processing a certain biological classification level (species, phylum, classification by feature), thus successfully mimicking the rules of hierarchical classification.

[0055] In step B, due to the specific characteristics of the dataset, some categories exhibit strong inter-class similarity (e.g., oyster shell and sea lettuce, prawn and tiger prawn) or weak intra-class similarity (e.g., white-spined sea urchin, giant horseshoe snail), which typically results in imbalance in the feature space. We specifically categorize the factors leading to visual tree imbalance into class imbalance and data imbalance.

[0056] Based on the above analysis, this embodiment proposes a Hierarchical Bayesian Information Criterion (HBIC) module. HBIC is an unsupervised learning method that aims to determine the optimal number of clusters by combining a confidence upper bound algorithm and the Bayesian information criterion. The learning process loss function includes reconstruction loss and entropy loss, such as... Figure 3 (b) This parameter is used to measure data dissimilarity and uncertainty. To address class imbalance, HBIC introduces a class balance parameter that considers the distribution of the number of subclasses within each class. It also addresses the data imbalance problem by considering the differences in sample size between different coarse classes to improve data balance. Finally, HBIC integrates these parameters to generate a balance parameter δ, which constrains the clustering process to obtain a high-quality and balanced visual tree.

[0057] The principle of the Hierarchical Bayesian Information Criterion (HBIC) is explained below:

[0058] The optimal number of clusters was determined by combining the Confidence Upper Bound (UCB) algorithm with the Bayesian Information Criterion (BIC). Compared to calculating the CVI, the UCB algorithm explores by selecting the number of clusters with the highest confidence upper bound. This method finds the optimal number of clusters in fewer iterations. The Bayesian Information Criterion (BIC) is a metric that comprehensively considers the balance between the number of clusters, the number of samples, and model parameters, effectively evaluating the complexity of clustering.

[0059] When using UCB for unsupervised learning, we employ a loss function that is the sum of the image reconstruction loss and the entropy loss. The reconstruction loss and entropy loss are defined as follows:

[0060]

[0061]

[0062] Where n is the number of samples, x i It is the i-th original data. It is x i After self-encoding reconstruction, p i This represents the probability that the current data belongs to the i-th category. Then, the final loss function of the UCB algorithm is defined as:

[0063]

[0064] This loss will also be applied to the evaluation, and the resulting loss value will be used to calculate the BIC Score. The BIC Score calculation formula is as follows:

[0065] bic_score = 2loss × N val +nc×logN val (4)

[0066] Where, N val This represents the number of samples in the validation set, and nc represents the number of clusters.

[0067] Furthermore, regarding the imbalance problem, during the clustering process, the algorithm only selects the optimal number of clusters without considering whether the number of finer clusters contained in each coarse cluster is balanced. This leads to class imbalance in the visual tree. To address this issue, we calculate the class balance parameter for each clustering result, defined as:

[0068]

[0069] The parameter nc represents the number of clusters in the clustering result, r E r represents the average number of clusters in the current clustering result. k This represents the number of subcategories contained in the k-th coarse category.

[0070] Besides the class imbalance problem, the hierarchical structure also suffers from data imbalance. The number of samples in each sub-class ranges from 50 to 2033, showing significant differences, which can lead to issues such as… Figure 2 The problem with the right branch. Both coarse class 1 and coarse class 2 contain 4 fine classes, but there is a significant difference in the amount of data between the two coarse classes. This can seriously affect the recognition accuracy. Therefore, we improve the learning of the visual tree by evaluating data balance, and the evaluation formula is defined as:

[0071]

[0072] Where, m E m represents the average number of samples in all coarse clusters in the clustering results. k This represents the number of samples in the k-th cluster of the current clustering result. To address both class imbalance and data imbalance during the visual tree learning process, we combine these two parameters to generate a new balance parameter δ, defined as:

[0073] δ = category score +data score (7)

[0074] Finally, considering the significant differences in data volume among the superclasses in the marine biology dataset we used, and thus focusing more on addressing the data imbalance problem, the final balance parameter is defined as:

[0075]

[0076] The parameter δ(nc) restricts each clustering, ensuring the constructed visual tree is effective for marine organism identification. A smaller value for this parameter results in a better-balanced visual tree. Since a smaller BICS score is better, we multiply this parameter by the BIC score to obtain the hierarchical Bayesian information criterion, denoted as δ(nc)×BIC score.

[0077] This parameter is designed to determine the appropriate number of clusters in hierarchical clustering, addressing the problem of unbalanced structure in visual tree construction.

[0078] In step C, the hierarchical neural network model EAHNet described in this embodiment employs an innovative feature extraction strategy, such as... Figure 4 (a) includes a full-channel learning variable-dimensional partial convolution module (EAConv), which combines partial convolution operations (PConv), a compression and activation module (SE), and pointwise convolution (Conv1X1) to efficiently and accurately capture local features in the image. (b) introduces a relative attention mechanism that uses relative position encoding (Bias), combining the advantages of convolution and self-attention mechanisms while maintaining the model's translation invariance and dynamic weight allocation.

[0079] The principle behind the full-channel learning variable-dimensional partial convolution module is explained below:

[0080] EAConv is an innovative convolutional operation designed specifically for processing complex marine life images. The complexity of marine life images necessitates high-dimensional feature representations to better capture and express the intricate features within the images. EAConv combines partially convolutional modules, a Squeeze-and-Excitation (SE) module, and pointwise convolution to accurately capture local features early in the feature extraction process and introduces a channel attention mechanism to enhance interactions between features. It efficiently extracts spatial features by reducing redundant computation and memory access, while using the SE module to learn the weights of each channel to highlight useful features and suppress responses to useless features. The introduction of pointwise convolution increases the number of channels in the feature map, improving its expressive power, and reduces computational complexity and the number of parameters through an inverted bottleneck structure, thus maintaining the module's efficiency. EAConv's design enables the network to better extract key channel information without increasing computational cost, providing richer feature information for subsequent modules and effectively addressing the challenges of processing complex marine life images.

[0081] Experimental verification:

[0082] The hierarchical neural network trained in this embodiment, which considers biological and feature relationships, can classify marine organisms. The training results were compared with those of several mainstream model frameworks.

[0083] This embodiment of the experiment was conducted on a privately owned large-scale intertidal marine organism dataset of the Nanji Islands. The dataset consisted of images collected and identified by a single institution at various survey sites, including Shangma'an, Xiama'an, Jianyu, Xiaochaiyu, Dalei Island, Houji, Poyu, Sanpanwei, Longchuanjiao, and Dashanjiao. Training samples were labeled as biological species. The Nanji Islands intertidal marine organism dataset comprises 342 classes, with each class containing at least 50 images and a maximum of 2033 images.

[0084] To facilitate comparison with previous work, this embodiment uses the widely adopted Top-1 accuracy as the evaluation metric. Since the data collection process did not involve splitting the dataset, this embodiment will divide the training and validation sets in an 8:2 ratio.

[0085] The classification network was trained for 300 epochs, with randomly sampled patch sizes of 224×224 and batch sizes of 64. Typically, an initial learning rate of 1×10⁻⁶ was used. -3 The AdamW optimizer is reduced with cosine annealing decay, momentum of 0.9, mini-batch size of 128, and weight decay of 5 × 10⁻⁶. -2Training was stopped when the validation loss did not improve over the past 30 epochs. Several data augmentation methods were considered during training: random cropping, random augmentation, random horizontal flipping, label smoothing regularization, mixup, cutMix, and random erasure as data augmentations. All experiments were based on a large private dataset of marine life from the intertidal zone of the Nanji Islands (342 categories, approximately 93,000 images), with 80% of the dataset randomly used for training and 20% for validation, implemented in PyTorch on 1×NVIDIA 3090Ti GPU and 1×NVIDIA 4090 GPU.

[0086] First, experiments were conducted with the balancing parameters to determine whether different weight parameters corresponding to imbalanced categories and imbalanced data resulted in the optimal number of clusters. The weight parameters were set to [0.1, 0.9], [0.2, 0.8], [0.3, 0.7], [0.4, 0.6], and [0.5, 0.5]. The experimental results are shown in Table 1. The weight parameter [0.4, 0.6] resulted in the highest number of clusters and thus the highest accuracy.

[0087] Table 1. Impact of the hierarchical structure constructed based on the balance parameter on classification accuracy.

[0088]

[0089]

[0090] Table 2. The combined effect of the two hierarchical structures on classification accuracy.

[0091]

[0092] We then verified whether various hierarchical structures positively impacted the neural network, and combined these hierarchical structures to jointly improve the network's classification performance. To improve classification accuracy, we needed to experimentally determine suitable loss weight parameters. The training data was divided into two parts: a training set of 45,000 samples and a validation set of 5,000 samples. Table 2 shows the experimental results. We tried seven different sets of weight parameters from coarse to fine: [0, 0, 1], [0, 0.1, 0.9], [0.1, 0, 0.9], [0.1, 0.1, 0.8], [0.2, 0.2, 0.6], and [0.2, 0.3, 0.5]. The first three sets used at most one hierarchical structure, disregarding the influence of loss weights; the sole purpose was to verify the effectiveness of the two visualization trees. Based on the experimental results, we selected [0.1, 0.1, 0.8] as the optimal weight parameters.

[0093] Furthermore, using a convolutional neural network as the backbone network, and comparing it with convolutional neural networks and three other hierarchical classification networks, the results are shown in Table 3. This method achieved the best results on the large-scale intertidal marine organism dataset of the Nanji Islands.

[0094] Table 3 Comparison with other hierarchical classification methods

[0095]

[0096]

[0097] EAHNet's architecture is a hybrid structure of convolutional and attention modules. The method of this invention was compared with three state-of-the-art classification methods to obtain more comprehensive comparison results: 1) a single-module-based method that uses only convolutional modules for classification; 2) a single-module-based method that uses only attention modules for classification; and 3) an ensemble-based method that integrates multiple modules to obtain more robust and broader results.

[0098] Table 4. Comparison of the proposed method with three advanced methods on a large-scale marine biological dataset in the intertidal zone of the Nanji Islands.

[0099]

[0100] The method of this invention achieves considerable overall performance in terms of accuracy, parameter count, FLOPs, and latency. Specifically, it achieves the highest accuracy, while having at least 35MB fewer parameters compared to other networks with accuracy exceeding 95%. Compared to the highest-accuracy convolutional module-based method, it is similar in performance aspects such as parameter count, but its accuracy is approximately 2% higher than the best single-convolutional module method. Compared to the highest-accuracy attention module-based method, the method of this invention has 0.9% higher accuracy, but approximately 50MB more parameters and approximately 12ms higher latency. Compared to methods based on the fusion of convolution and attention, the method of this invention also surpasses the former in overall performance, convincingly demonstrating the efficiency of the method.

[0101] 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 marine organism classification method based on hierarchical neural networks, characterized in that, The following method steps are used to obtain a model for marine organism classification: Step 1) Collect marine life image data and perform image processing; establish a sample dataset; Step 2) Generate a classification visual tree using prior knowledge of marine biological relationships and an unsupervised learning method; the classification visual tree is a hierarchical structure defined by hierarchical clustering; the unsupervised learning method involves using a convolutional neural network to extract biological feature relationships from image data, calculating the optimal number of clusters, and dividing the hierarchical structure based on the optimal number of clusters; the optimal number of clusters is calculated using a combination of the confidence upper bound algorithm and the Bayesian information criterion. ×BIC Score is used to achieve balance in the tree structure. BIC Score: ； in, , It is the sample size. It is the i-th original data. yes Data after self-encoding reconstruction This indicates that the current data belongs to the first... The probability of each category, Indicates the number of samples in the validation set. Represents the number of clusters; The equilibrium parameters are: ； in, This represents the average number of clusters in the current clustering result. This represents the number of subcategories contained in the k-th coarse category. This represents the average number of samples in all coarse clusters in the clustering results. This represents the number of samples contained in the k-th cluster in the current clustering result; Step 3) Construct the EAHNet backbone neural network based on the classification visual tree, which includes a shared layer, independent layers, and a classifier. This network uses biological relationships to assist in the classification of processed marine organisms. The shared layer sub-task classification is used to further extract coarse-scale features from the processed image. The independent layer sub-task classification includes: a phylum layer, a feature layer, and a species layer. Each independent layer takes the shared coarse-scale features as input and extracts corresponding fine-scale features after passing through each layer. The classifier weighted sums of the fine-scale features from each layer to generate complete classification labels. Step 4) Iterate repeatedly using the sample dataset to train the EAHNet backbone neural network and obtain the trained and optimized marine organism classification model.

2. The marine organism classification method based on hierarchical neural networks according to claim 1, characterized in that, The image processing includes image noise reduction and size normalization.

3. The marine organism classification method based on hierarchical neural networks according to claim 1, characterized in that, The shared layer is the full-channel learning variable-dimensional partial convolutional module EAConv, which includes a partial convolutional module, an SE module, and a pointwise convolutional module, extracting coarse-scale features and inputting them into each independent layer.

4. The marine organism classification method based on hierarchical neural networks according to claim 1, characterized in that, The independent layer employs a relative attention module to extract fine-scale features from coarse-scale features, thereby capturing biological characteristics.

5. The marine organism classification method based on hierarchical neural networks according to claim 1, characterized in that, In the classifier structure, the fine-scale features generated by each independent layer are focused on and weighted according to the biological information structure of each layer, and are used to finally summarize the biological information of different structures to generate complete classification labels, thereby realizing the classification of marine organisms.

6. The marine organism classification method based on hierarchical neural networks according to claim 1 or 5, characterized in that, The classifier is a softmax classifier; After generating fine-scale features in each independent layer, the results are fed into Softmax to obtain the classification results for each layer. Based on the biological information structure of each layer, the three results are weighted and focused to generate the final classification labels, thus achieving marine organism classification. The focus and weighting formulas are as follows: ； Where F represents the final classification result, , , These represent the classification results at the phylum, characteristic, and species levels, respectively. , , This indicates that each layer is assigned its own weight.

Images