A biological image recognition method and system based on dynamic hierarchical semantic routing

Abstract

The present application relates to the technical field of biological image recognition, and particularly relates to a biological image recognition method and system based on dynamic hierarchical semantic routing; spatial geometric topological information of a target biological individual is extracted by a target detection network, and a region of interest is extracted according to the information; meanwhile, reasoning is performed from a high level to a low level in stages according to a biological classification hierarchical topological structure, and a dynamic hierarchical semantic routing mechanism is introduced, so that the model can automatically select a proper recognition granularity under different sample conditions, the problem of hierarchical logical conflicts caused by independent prediction of each level in the traditional method is avoided, and the structural self-consistency and reliability of the recognition result on the biological classification tree are ensured; a next-level legal candidate class set is dynamically generated according to a last-level prediction result, and a classification logic value is subjected to mask constraint, so that the occurrence of cross-level semantic conflicts is fundamentally eliminated; meanwhile, an elastic rollback mechanism is designed, so that the problem of cumulative propagation of fine-grained classification errors is effectively inhibited.

Description

Technical Field

[0001] This invention relates to the field of biological image recognition technology, and in particular to a biological image recognition method and system based on dynamic hierarchical semantic routing. Background Technology

[0002] With the rapid development of computer vision and deep learning technologies, automated biological classification methods based on image recognition have been widely applied in agriculture, ecological monitoring, and biodiversity research. Biological identification tasks are typically characterized by a large number of categories, subtle morphological differences, and high data acquisition costs, especially in highly diverse groups such as insects. Traditional biological classification mainly relies on expert observation of morphological features (such as antennae, wings, legs, and abdominal structures), which suffers from low efficiency, strong reliance on experience, and difficulty in scaling. In recent years, deep learning methods have been introduced into biological image recognition tasks, achieving end-to-end classification through convolutional neural networks (CNNs) or visual Transformers. Some schemes further introduce multi-level classification architectures, setting independent classification branches at different semantic levels such as "order-family-genus-species" to utilize the hierarchical structural information in the biological classification system.

[0003] However, existing methods lack effective path constraint mechanisms between classification branches at different levels during multi-level inference, leading to potential logical conflicts in prediction results across different levels. Specifically, each classification branch typically predicts independently, with each level performing global Softmax normalization based on the entire set of categories at that level to select the category with the highest probability. This means that the prediction at the next level is not limited to the subclass of the prediction at the previous level in the classification tree, and the predictions between levels are essentially decoupled. Taking insect image recognition as an example, the model might predict that the target belongs to "Callic Beetles" at the family level, but at the subsequent genus level, it might predict a genus that does not belong to Callic Beetles taxonomically, or even a genus-level category that has no subordinate relationship with Calicutidae. While such prediction results may have high single-level confidence in a statistical sense, they constitute a clear logical contradiction at the biological taxonomic level. That is, any complete path on the classification tree should be unique and definite, and there should be a strict parent-child inclusion relationship between higher-level categories (families) and lower-level categories (genuses, species). Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a biological image recognition method and system based on dynamic hierarchical semantic routing. This solves the technical problem that existing technologies, when predicting biological categories through multi-level classification architectures, lack an effective path constraint mechanism between classification branches at each level, which may lead to logical conflicts in prediction results at different levels.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a biometric image recognition method based on dynamic hierarchical semantic routing, comprising the following steps: Acquire images of the organism to be identified; The biometric image to be identified is input into a pre-trained biometric model that includes at least an object detection network and a multi-level classification network to obtain the biometric result. The specific steps include: Spatial geometric topological information of target biological individuals in biological images to be identified is extracted using an object detection network; Extracting the region of interest (ROI) of the target in the biological image to be identified based on spatial geometric topological information; The target region of interest is input into a multi-level classification network. The multi-level classification network performs reasoning step by step from the higher level to the lower level according to the hierarchical order of the pre-constructed biological classification hierarchical topology. During the step-by-step reasoning process, a dynamic hierarchical semantic routing mechanism is introduced to adaptively adjust the classification depth according to the data distribution characteristics and model uncertainty, and the final biometric identification result is output.

[0006] Preferably, the spatial geometric topology information includes at least a set of target bounding boxes and corresponding location confidence scores; the step of extracting the target region of interest in the biological image to be identified based on the spatial geometric topology information specifically includes: For each bounding box in the target bounding box set, an expansion coefficient is determined based on its corresponding location confidence score, wherein the expansion coefficient is negatively correlated with the location confidence score; The bounding box is scaled and its position updated according to the dilation coefficient to obtain the corrected bounding box. The target region of interest is extracted from the biological image to be identified based on the corrected bounding box.

[0007] Preferably, the dynamic hierarchical semantic routing mechanism includes: Joint decision-making should be made at least based on a dynamic routing decision function constructed from sample size, class imbalance, and network confidence. The classification depth is determined by comparing the calculation results of the dynamic routing decision function with a preset threshold; the degree of class imbalance is determined based on the ratio of the maximum number of samples in all classes to the number of samples in the current class.

[0008] Preferably, the multi-level classification network includes a shared feature extraction backbone network and multiple hierarchical classification branches, wherein the shared feature extraction backbone network is used to extract the deep feature representation from the target region of interest. Each of the hierarchical classification branches is used to implement conditional probability modeling at each level; The multi-level classification network employs a hierarchical conditional probability chain modeling approach to model the multi-semantic hierarchical classification process, transforming predictions at different semantic levels into conditional probability decomposition forms, thus achieving step-by-step reasoning from higher to lower levels; its joint probability distribution is expressed as: Among them, when When the initial conditional probability is defined as: In the above formula, Indicates the first The prediction results for each semantic level, which, from high to low, includes at least order, family, subfamily, genus, and species.

[0009] Preferably, when the multi-level classification network reasones from high to low levels, the prediction result of the current level is used to constrain the candidate category space of the next level, specifically including: Based on the prediction results of the current level and the biological classification hierarchical topology, determine the set of legal candidate categories for the next level. The set of legal candidate categories consists of all subclass nodes corresponding to the prediction results of the current level in the biological classification hierarchical topology. In the next level of classification calculation, the classification logic value is processed by mask constraint to obtain the corrected classification logic value. The classification logic value corresponding to the category that does not belong to the set of legal candidate categories is assigned to negative infinity, and the classification logic value corresponding to the category that belongs to the set of legal candidate categories is mapped according to the preset rules. The corrected classification logistic values ​​are normalized using Softmax to obtain the probability distribution of the next level, and the final prediction result and confidence level are calculated.

[0010] Preferably, when the prediction confidence of any level in the step-by-step reasoning process is lower than a preset threshold, an elastic backoff mechanism is triggered to terminate the reasoning to a lower level, and backoff step by step according to the hierarchical order of the biological classification hierarchical topology until the prediction confidence of the target level meets the preset condition or backoffs to the highest level; the prediction result of the target level that meets the preset condition is output as the final biometric result.

[0011] Preferably, the biometric model is trained end-to-end using a joint loss function, which is at least a weighted sum of object detection loss, hierarchical consistency loss, classification loss, and fine-grained loss. The target detection loss is used to optimize the target localization and classification accuracy, and consists of bounding box regression loss, target confidence loss and classification loss in the detection stage. The hierarchical consistency loss is used to constrain the prediction consistency between different semantic levels, so that the prediction results of higher levels can guide the learning of lower levels. The classification loss is used to constrain the accuracy of classification predictions at each level; The fine-grained loss is used to enhance the classification ability at the lowest level; The weighting of the fine-grained loss is determined based on the number of class samples, and the weighting is inversely correlated with the number of class samples.

[0012] The present invention also provides a biological image recognition system, including a processor and a memory, wherein the memory is used to store a computer program, and the computer program, when executed by the processor, implements the biological image recognition method.

[0013] By employing the above technical solution, the present invention provides a biological image recognition method and system based on dynamic hierarchical semantic routing, which has at least the following beneficial effects: 1. This invention extracts the spatial geometric topology information of target biological individuals through a target detection network and extracts the region of interest based on this information to eliminate background interference in complex natural scenes. At the same time, it infers from high level to low level according to the topological structure of biological classification and introduces a dynamic hierarchical semantic routing mechanism during the inference process, so that the model can automatically select the appropriate recognition granularity under different sample conditions. This avoids the problem of hierarchical logical conflicts caused by the independent prediction of each level in traditional methods, and ensures the structural self-consistency and credibility of the recognition results on the biological classification tree.

[0014] 2. This invention fundamentally eliminates cross-level semantic conflicts by dynamically generating a set of legal candidate categories for the next level based on the prediction results of the previous level and applying mask constraints to the classification logic values. Simultaneously, by designing an elastic backoff mechanism, it effectively suppresses the problem of the cumulative propagation of fine-grained classification errors. Furthermore, by constructing a multi-task joint loss function that includes object detection loss, hierarchical consistency loss, classification loss, and fine-grained loss, and introducing adaptive weights that are inversely correlated with the number of category samples, it significantly improves the recognition performance of categories with few samples under long-tail distributions, enhancing the model's generalization ability and engineering deployment value in real-world long-tail data scenarios.

[0015] 3. This invention transforms the hierarchical structure of order, family, subfamily, genus, and species into conditional dependencies between levels through multi-level conditional probability chain modeling. Combined with a dynamic routing decision function based on sample size, class imbalance, and network confidence, it achieves adaptive adjustment of classification depth under different sample conditions, further improving the model's recognition stability in cases of scarce data or extreme class imbalance. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the biometric model of the present invention; Figure 2 This is a flowchart illustrating the hierarchical structure constraints of biological taxonomy in this invention. Detailed Implementation

[0017] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0018] To address the technical problem in existing technologies where the lack of effective path constraint mechanisms between classification branches at different levels leads to potential logical conflicts in prediction results across different levels when predicting biological categories using a multi-level classification architecture, this embodiment provides a biological image recognition method based on dynamic hierarchical semantic routing. This method includes the following steps: Acquire images of the organism to be identified The biological images include images collected from natural scenes in the field, standardized images from the laboratory, or images from standard databases. To adapt to the inconsistent distribution of images from different sources under factors such as changes in illumination, scale differences, and complex background interference, a unified image input representation space can be constructed first.

[0019] Then, the biometric image to be identified is input into a pre-trained biometric model that includes at least an object detection network and a multi-level classification network to obtain the biometric result, such as... Figure 1 As shown, the specific steps include: First, the biological image to be identified is input into a pre-trained target detection network for forward propagation processing. The target detection network employs existing single-stage or two-stage target detection models, including the YOLO series or Faster R-CNN models. These models jointly model the input features through regression and classification branches. The regression branch predicts the target bounding box location parameters, while the classification branch predicts the target category probability, thus achieving collaborative computation of target localization and semantic recognition. Through the regression and classification branches, the spatial geometric topology information of the target biological individual is output, specifically represented by a set of target bounding boxes. And the corresponding location reliability score, the expression of which is: in, This represents the nonlinear mapping function of the object detection network.

[0020] This design transforms the whole-image classification problem into a target-level instance recognition problem, thereby reducing background interference and improving detection robustness and localization accuracy in complex scenes, outputting a set of bounding boxes. The location reliability score is used to describe the spatial location of the target and to characterize the detection reliability, providing an input basis for subsequent ROI extraction and hierarchical classification.

[0021] After obtaining the spatial geometric topological information of the target biological individual, the biological image to be identified is processed based on the spatial geometric topological information. Perform spatial matrix clipping operation This allows us to obtain images of the target region of interest (ROI). Its expression is: To adapt to situations involving varying target scales, partial occlusion, and high detection uncertainty in outdoor scenarios, this embodiment performs target bounding box assemblies before ROI clipping. The adaptive expansion compensation process includes the following steps: For each bounding box in the target bounding box set ,in, and Indicates the first bounding boxes The x and y coordinates of the center position, and They represent the first bounding boxes The width and height, based on their corresponding location reliability scores. Determine the expansion coefficient The expansion coefficient is negatively correlated with the location confidence score; that is, the lower the location confidence score, the larger the expansion coefficient, thus improving the coverage of potential target areas. Subsequently, the bounding box is scaled and its location updated based on the expansion coefficient to obtain the corrected bounding box. Its update process can be represented as: In the above formula, and Represents the corrected bounding box Width and height, and Represents the corrected bounding box The x and y coordinates of the center position are determined; finally, based on the corrected bounding box... The biological image to be identified is cropped to obtain the target region of interest with more complete structural information. In addition, to reduce the impact of imaging differences on subsequent feature extraction and classification processes, selective targeting of the target's region of interest can be performed. Standardized preprocessing is performed, including size normalization, brightness equalization, noise suppression, and data augmentation.

[0022] After obtaining the target region of interest Then, the target region of interest is input into a multi-level classification network. This multi-level classification network reasones step by step from the higher level to the lower level according to the hierarchical order of the pre-constructed biological classification hierarchical topology. During the step-by-step reasoning process, a dynamic hierarchical semantic routing mechanism is introduced to adaptively adjust the classification depth according to the data distribution characteristics and model uncertainty, and finally outputs the biometric identification result.

[0023] In this embodiment, the biological classification hierarchical topology is a pre-constructed, coarse-to-fine hierarchical topological structure tree based on prior rules of taxonomy. This structure describes the semantic hierarchical relationships between different biological categories. The biological classification hierarchical topology includes M semantic levels: Order, Family, Subfamily, Genus, and Species, thus extending the traditional planar classification labels into a hierarchical path structure with structural constraints. During training, statistical analysis is performed on the number N of samples in each category within the training set to form a category distribution vector. Based on this distribution, a unique hierarchical path label is assigned to each category node to construct a label system with hierarchical consistency constraints. Furthermore, to enhance the learnability of this hierarchical structure, each level is defined as a conditional dependency, thus forming a structured hierarchical classification space.

[0024] The dynamic hierarchical semantic routing mechanism introduced in this embodiment performs joint decision-making based at least on a dynamic routing decision function constructed from sample size, class imbalance, and network confidence. The class imbalance is determined by the ratio of the maximum number of samples in all classes to the number of samples in the current class, and its expression is: In the above formula, Indicates the first The degree of class imbalance in each category Indicates the first The number of samples in each category This represents the maximum number of samples across all categories. This is used to prevent extremely small positive numbers with a denominator of 0.

[0025] Network confidence Defined as: In the above formula, This represents the probability of the subsequent Softmax normalized output.

[0026] Based on this, a dynamic routing decision function is constructed. The expression is: in, This represents the sample size normalization function. , and These are the weighting coefficients.

[0027] Finally, the classification depth is determined by comparing the calculation results of the dynamic routing decision function with a preset threshold. For example, thresholds T1 and T2 (T1 < T2) are set, and the classification path is dynamically controlled as follows: in, , and These correspond to different semantic levels. Through the above mechanism, the model adaptively adjusts the classification depth based on data distribution characteristics and model uncertainty during the step-by-step inference process. When the decision function value is small (sufficient data, class balance, high confidence), the classification depth is deep, and fine-grained inference is performed; when the decision function value is large (scarce data, class imbalance, low confidence), the classification depth is shallow, and the model terminates at a higher level.

[0028] The multi-level classification network used in this embodiment employs a hierarchical conditional probability chain modeling approach to model the multi-semantic hierarchical classification process. It transforms the predictions at different semantic levels into conditional probability decompositions, enabling step-by-step reasoning from higher to lower levels. Its joint probability distribution is expressed as: Among them, when When the initial conditional probability is defined as: In the above formula, Indicates the first The prediction results of a semantic level, which includes at least M levels from high to low, such as Order, Family, Subfamily, Genus, and Species.

[0029] To implement the hierarchical conditional probability chain modeling process described above, the multi-level classification network in this embodiment includes a shared feature extraction backbone network (Backbone) and multiple hierarchical classification branches. The output of each hierarchical classification branch... Corresponding to the first in the conditional probability model Layer probability distribution This achieves a one-to-one mapping between the probabilistic model and the network structure. The mapping can be expressed as: in, Indicates the first The output of the first hierarchical classification branch is also the first... Prediction results at each semantic level Indicates in features and Conditional probability distribution under the conditions of the prediction results of the first level; In this implementation, the shared feature extraction backbone network is used to extract deep feature representations from the target region of interest. Its expression is as follows: In the above formula, Indicates a shared feature extraction network. This represents the depth feature representation extracted from the ROI image.

[0030] Based on this, each level of classification branch is used to implement the conditional probability of each level. The parameterized modeling transforms the probability decomposition process into a trainable neural network structure, which is implemented as follows: in, Indicates the first The classification function corresponding to each semantic level is used to approximate the conditional probability distribution, thereby completing the step-by-step reasoning process from high level to low level.

[0031] Furthermore, in this embodiment, when the multi-level classification network performs inference from high to low levels, before the Softmax calculation, to ensure that the multi-level classification results conform to the hierarchical structure constraints of biological taxonomy, the prediction results of the current level are used to constrain the candidate category space of the next level. A valid candidate set is generated based on the classification results of the previous level. Furthermore, illegal categories are masked to avoid cross-level error jumps and improve the structural consistency of hierarchical reasoning. The specific steps include the following: like Figure 2As shown, firstly, based on the prediction results of the current level and the biological classification hierarchical topology, the set of legal candidate categories for the next level is determined. The set of legal candidate categories consists of all subclass nodes corresponding to the prediction results of the current level in the biological classification hierarchical topology. Specifically, let the first... Layer prediction results are Then, based on the pre-constructed hierarchical mapping function of biological classification , obtained the The set of valid candidates for a layer is expressed as follows: in, This represents the set of all lower-level categories that have parent-child relationship constraints with the previous-level category in the biological classification tree structure, which is also the set of all subclass nodes corresponding to the current level prediction result in the biological classification hierarchical topology.

[0032] Based on this, in the next level of classification calculation, the classification logic value logits is subjected to mask constraint processing to obtain the corrected classification logic value. The classification logic values ​​corresponding to categories that do not belong to the set of legal candidate categories are assigned negative infinity, and the classification logic values ​​corresponding to categories that belong to the set of legal candidate categories are mapped according to preset rules. In this embodiment, the classification logic values ​​corresponding to categories that belong to the set of legal candidate categories remain unchanged after correction. For any category... The corrected expression for calculating the classification logical value can be: In the above formula, The corrected classification logic value. This represents the original classification logical value.

[0033] Subsequently, the corrected classification logistic values ​​are subjected to Softmax normalization to obtain the probability distribution of the next level: The final prediction result and confidence level are defined as follows: in, Used to indicate the current level The confidence level of the prediction is used as the input for subsequent hierarchical reasoning and backoff mechanisms.

[0034] In this embodiment, during the inference phase, the system performs predictions step by step according to the semantic hierarchy from high to low, and outputs the corresponding classification result for each level. and its confidence score Confidence score This is used to characterize the reliability and uncertainty of the prediction result at the current layer. When the prediction confidence of any layer during the step-by-step inference process is lower than a preset threshold, or when the number of training samples corresponding to the current category is lower than a set few-sample threshold, the prediction result at that layer is determined to be unreliable, triggering an elastic backoff mechanism. This terminates the inference to lower layers and backoffs step-by-step upwards according to the hierarchical order of the biological classification hierarchy, in order to avoid the downward propagation of erroneous fine-grained classification results, until the prediction confidence of the target layer meets the preset conditions or the backoff reaches the highest layer. The prediction result of the target layer that meets the preset conditions is then output as the final biometric result.

[0035] In this embodiment, "previous-level prediction result" refers to the prediction result of the parent node of the current-level node in the hierarchical classification tree structure, which is formally represented as: In the above formula, for The prediction result of the parent node of the corresponding node. This indicates the extraction of the prediction result from the parent node, used to describe the stable semantic state of the current semantic level in the next higher level of the classification tree. For example, in the hierarchical system of Order → Family → Subfamily → Genus → Species, when the system's prediction confidence at the Genus level is insufficient, it will automatically fall back to the prediction result at the Subfamily level. If the confidence requirement at the Subfamily level is still not met, it will continue to fall back level by level to the Family level and even the Order level, thereby gradually improving the stability of the prediction result. Through this elastic fallback mechanism, the system automatically terminates the low-level fine-grained inference process under conditions of low confidence or scarce data, and outputs prediction results with higher confidence and more stable semantics at higher levels. This effectively suppresses the problem of error accumulation and propagation, and improves the overall classification robustness and interpretability of the model under complex backgrounds, long-tail categories, and small sample conditions.

[0036] The biometric model in this embodiment uses a joint loss function for end-to-end training. At least for target detection loss Hierarchical consistency loss Classification loss and fine-grained loss The weighted sum, which simultaneously constrains target detection performance, hierarchical consistency, classification accuracy, and long-tail distribution adaptability, is expressed as follows: Among them, target detection loss To optimize target localization and classification accuracy, the bounding box regression loss is used. Target confidence loss and classification loss during the detection stage Composition, bounding box regression loss It can be either IoU loss or L1 loss, and its expression is: Hierarchical consistency loss This is used to constrain the consistency of predictions across different semantic levels, enabling higher-level predictions to guide lower-level learning. Its expression is: in, This represents the distribution distance function, which can be either KL divergence or L2 distance.

[0037] Classification loss The expression used to constrain the accuracy of classification predictions at each level is: in, For the first Layered real labels.

[0038] Fine-grained loss The expression used to enhance the classification ability at the lowest level is: in, For the first Layered real labels.

[0039] Weighting of fine-grained loss Based on the number of samples by category It is determined that the weighting weights are inversely correlated with the number of class samples, thereby enhancing the ability to learn classes with fewer samples. The expression for this weighting weights can be: During the training phase of the biometric model, biometric image samples with real labels are input into the model, which consists of an object detection network, a shared feature extraction backbone network, and a multi-level classification network. Forward propagation calculates the joint loss function, and backpropagation is used to synchronously update all trainable parameters in the biometric model, thereby achieving end-to-end collaborative optimization of the detection and classification tasks.

[0040] This embodiment also provides a biological image recognition system, including a processor and a memory, wherein the memory is used to store a computer program, and the computer program implements the biological image recognition method when executed by the processor.

[0041] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0042] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0043] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A biological image recognition method based on dynamic hierarchical semantic routing, characterized in that, The method includes the following steps: Acquire images of the organism to be identified; The biometric image to be identified is input into a pre-trained biometric model that includes at least an object detection network and a multi-level classification network to obtain the biometric result. The specific steps include: Spatial geometric topological information of target biological individuals in biological images to be identified is extracted using an object detection network; Extracting the region of interest (ROI) of the target in the biological image to be identified based on spatial geometric topological information; The target region of interest is input into a multi-level classification network. The multi-level classification network performs reasoning step by step from the higher level to the lower level according to the hierarchical order of the pre-constructed biological classification hierarchical topology. During the step-by-step reasoning process, a dynamic hierarchical semantic routing mechanism is introduced to adaptively adjust the classification depth according to the data distribution characteristics and model uncertainty, and the final biometric identification result is output.

2. The biometric image recognition method according to claim 1, characterized in that, The spatial geometric topology information includes at least the set of target bounding boxes and the corresponding location confidence scores; The step of extracting the target region of interest from the biological image to be identified based on spatial geometric topological information specifically includes: For each bounding box in the target bounding box set, an expansion coefficient is determined based on its corresponding location confidence score, wherein the expansion coefficient is negatively correlated with the location confidence score; The bounding box is scaled and its position updated according to the dilation coefficient to obtain the corrected bounding box. The target region of interest is extracted from the biological image to be identified based on the corrected bounding box.

3. The biometric image recognition method according to claim 1, characterized in that, The dynamic hierarchical semantic routing mechanism includes: Joint decision-making is performed at least based on a dynamic routing decision function constructed from sample size, class imbalance, and network confidence; the class imbalance is determined based on the ratio of the maximum number of samples in all classes to the number of samples in the current class. The classification depth is determined by comparing the calculation results of the dynamic routing decision function with a preset threshold.

4. The biometric image recognition method according to claim 1, characterized in that, The multi-level classification network includes a shared feature extraction backbone network and multiple hierarchical classification branches. The shared feature extraction backbone network is used to extract the deep feature representation from the target region of interest. Each of the hierarchical classification branches is used to implement conditional probability modeling at each level; The multi-level classification network adopts a hierarchical conditional probability chain modeling method to model the multi-semantic hierarchical classification process, transforming the predictions of different semantic levels into conditional probability decomposition forms, and realizing step-by-step reasoning from high-level to low-level. Its joint probability distribution is expressed as: Among them, when When the initial conditional probability is defined as: In the above formula, Indicates the first The prediction results for each semantic level, which, from high to low, includes at least order, family, subfamily, genus, and species.

5. The biological image recognition method according to claim 1, characterized in that, When the multi-level classification network reasones from high to low levels, the prediction result of the current level is used to constrain the candidate category space of the next level, specifically including: Based on the prediction results of the current level and the biological classification hierarchical topology, determine the set of legal candidate categories for the next level. The set of legal candidate categories consists of all subclass nodes corresponding to the prediction results of the current level in the biological classification hierarchical topology. In the next level of classification calculation, the classification logic value is processed by mask constraint to obtain the corrected classification logic value. The classification logic value corresponding to the category that does not belong to the set of legal candidate categories is assigned to negative infinity, and the classification logic value corresponding to the category that belongs to the set of legal candidate categories is mapped according to the preset rules. The corrected classification logistic values ​​are Softmax normalized to obtain the probability distribution of the next level and the final prediction result and confidence level are calculated.

6. The biological image recognition method according to claim 1, characterized in that, When the prediction confidence of any level in the step-by-step reasoning process is lower than a preset threshold, an elastic backoff mechanism is triggered to terminate the reasoning to a lower level and backoff step by step according to the hierarchical order of the biological classification hierarchical topology until the prediction confidence of the target level meets the preset condition or backoffs to the highest level. The prediction results of the target level that meet the preset conditions will be output as the final biometric result.

7. The biometric image recognition method according to claim 1, characterized in that, The biometric model is trained end-to-end using a joint loss function, which is at least a weighted sum of object detection loss, hierarchical consistency loss, classification loss, and fine-grained loss. The target detection loss is used to optimize the target localization and classification accuracy, and consists of bounding box regression loss, target confidence loss and classification loss in the detection stage. The hierarchical consistency loss is used to constrain the prediction consistency between different semantic levels, so that the prediction results of higher levels can guide the learning of lower levels. The fine-grained loss is used to enhance the classification ability at the lowest level; The classification loss is used to constrain the accuracy of classification predictions at each level; The weighting of the fine-grained loss is determined based on the number of class samples, and the weighting is inversely correlated with the number of class samples.

8. A system for implementing the biometric image recognition method according to any one of claims 1-7, characterized in that, It includes a processor and a memory, the memory being used to store a computer program that, when executed by the processor, implements the biometric image recognition method as described in any one of claims 1-7.

Images