Bird biometric feature recognition method based on end-side lightweight neural network

Abstract

This invention discloses a bird biometric recognition method using a lightweight edge-side neural network, relating to the field of fine-grained image recognition technology. The method includes constructing a backbone structure based on a lightweight convolutional network and embedding an attention-guided module that integrates prior knowledge of bird anatomy within the backbone structure, resulting in a student network capable of focusing on key biometric features. Based on a large-scale cloud-based bird recognition model, three-level soft labels (family, genus, and species) and intermediate semantic features are generated as supervisory signals for knowledge distillation. The supervisory signals are used to train the student network using hierarchical knowledge distillation, enabling the student network to inherit high-order classification logic and fine-grained discriminative ability under limited parameter conditions, resulting in a compressed edge-side recognition model. The compressed edge-side recognition model is deployed on an edge device, and a lightweight complexity evaluation sub-network is configured at the model's front end. The lightweight complexity evaluation sub-network performs scene complexity analysis on the input image.

Description

Technical Field

[0001] This invention relates to the field of fine-grained image recognition technology, and in particular to a method for bird biometric recognition using a lightweight end-side neural network. Background Technology

[0002] Fine-grained image recognition technology refers to a computer vision task that aims to distinguish subcategories within the same parent category in an image with high precision, given the small differences between categories and the large variations within a category. Unlike ordinary image classification, which focuses on broad categories, fine-grained recognition focuses on intra-species subdivision. Its core challenge lies in capturing subtle morphological, textural, or structural differences in key local regions and suppressing the influence of interference factors such as pose, lighting, and occlusion. It typically requires combining attention mechanisms, part localization, hierarchical modeling, or domain prior knowledge to improve discriminative ability.

[0003] Existing lightweight models lack biological prior guidance, making it difficult to effectively capture subtle morphological differences among bird species under conditions of limited parameters and computational resources. This results in low accuracy in fine-grained recognition. Furthermore, traditional knowledge distillation methods only focus on imitating the final output layer, ignoring the semantic logic inherent in the taxonomic hierarchy, which prevents student models from inheriting the systematic classification capabilities of teacher models. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a bird biometric identification method using a lightweight end-side neural network. This solves the problem that existing lightweight models lack biological prior guidance, making it difficult to effectively capture subtle morphological differences between bird species under conditions of limited parameters and computational resources, resulting in low accuracy in fine-grained identification. Furthermore, traditional knowledge distillation methods only focus on imitating the final output layer, ignoring the semantic logic inherent in the taxonomic hierarchical structure, which makes it impossible for student models to inherit the systematic classification capabilities of teacher models.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for bird biometric identification using a lightweight end-side neural network, comprising: A backbone structure based on a lightweight convolutional network was constructed, and an attention-guided module that integrates prior knowledge of bird anatomy was embedded in the backbone structure to obtain a student network with the ability to focus on key biological features. Based on a large-scale cloud-based bird identification model, three levels of soft labels (family, genus, and species) and intermediate semantic features are generated as supervisory signals for knowledge distillation. By using supervisory signals to train the student network through hierarchical perception knowledge distillation, the student network can inherit high-order classification logic and fine-grained discrimination ability under the condition of limited parameters, resulting in a compressed end-side recognition model. The compressed edge recognition model is deployed on edge devices, and a lightweight complexity evaluation subnetwork is configured at the front end of the model. A lightweight complexity evaluation subnetwork is used to perform scene complexity analysis on the input image and output a complexity score. The execution state of multiple convolutional blocks in the main structure is dynamically controlled based on the complexity score, skipping redundant computation paths, completing adaptive inference, and outputting bird species identification results.

[0007] As a preferred embodiment of the bird biometric identification method using a lightweight end-side neural network described in this invention, the attention guidance module that integrates prior knowledge of bird anatomy comprises the following steps: On a standard bird image dataset, professionals with a taxonomic background annotated three types of anatomical regions with species discrimination value: the beak region, the eye ring region, and the wing patch region, forming three sets of binary mask images; A two-dimensional Gaussian kernel is applied to each set of binary mask images for smoothing, generating a continuously distributed region importance map; Each region importance map is compressed into a single-channel weight vector using a global average pooling operation. The dimension of this vector is consistent with the number of channels in the output feature map of the current convolutional block of the backbone network; The weight vector is introduced into the channel weight calculation path of the standard SE attention module. As a bias term, the channel attention output is expressed as: ; in, The channel statistics are obtained after global average pooling of the input feature map. and The learnable weight matrices of the two fully connected layers form a bottleneck structure for dimensionality reduction and dimensionality increase. Represents the ReLU activation function; This is the original bias vector; Represents the Sigmoid function; These are the final channel attention weights after modulation; Using the channel attention weights Channel-by-channel scaling is performed on the input feature map to enhance local anatomical features that are strongly correlated with bird classification.

[0008] As a preferred embodiment of the bird biometric recognition method using the end-side lightweight neural network described in this invention, the Gaussian smoothing process employs a two-dimensional Gaussian kernel with a fixed standard deviation to perform a convolution operation on a binary mask, ensuring a smooth transition of the anatomical region boundaries and avoiding gradient discontinuities caused by hard boundaries. This ensures the smoothness of the subsequently generated weight vector. It should reasonably reflect the spatial influence range of key areas.

[0009] As a preferred embodiment of the bird biometric identification method using a lightweight end-side neural network described in this invention, the method involves: using supervised signals to perform hierarchical knowledge distillation training on the student network, enabling the student network to inherit high-order classification logic and fine-grained discriminative ability under limited parameter conditions, resulting in a compressed end-side identification model, including: Extract the three-level output probability distributions of species, genus, and family from a large-scale cloud-based bird identification model; The corresponding three-level prediction distribution is obtained synchronously from the student network; Calculate the seed-level cross-entropy loss to measure how well the student network fits the real labels; Calculate the Kullback-Leibler divergence loss at the genus and family levels to force the student network to mimic the output distribution of the teacher model at higher classification levels; Extract the intermediate feature maps from the convolutional layers of the teacher and student networks that embed the Bio-SE module, and calculate the feature alignment loss, expressed as: ; in, This refers to the feature map output by the teacher network at a specific convolutional layer. Feature maps of the corresponding layers in the student network; This represents a global average pooling operation, which compresses the spatial information of each channel into a single value. The various losses are summed according to preset weights to form the total loss function, which is used to optimize student network parameters end-to-end.

[0010] As a preferred embodiment of the bird biometric identification method using the end-side lightweight neural network described in this invention, the feature alignment loss is implemented only at the convolutional block location where the Bio-SE attention module is embedded in the backbone network.

[0011] As a preferred embodiment of the bird biometric recognition method using a lightweight edge-side neural network described in this invention, the following steps are taken: The compressed edge-side recognition model is deployed on an edge device, and a lightweight complexity evaluation sub-network is configured at the front end of the model. The lightweight complexity evaluation sub-network performs scene complexity analysis on the input image and outputs a complexity score. The input image is fed into a sub-network consisting of two 1x1 convolutional layers. The first layer outputs eight channels, and the second layer outputs a single channel. Perform global average pooling on the output of the second layer to obtain a scalar value; Applying the Sigmoid function to the scalar value generates a normalized complexity score. The complexity score reflects the number of targets in the image, the degree of background clutter, and the pose occlusion situation; The subnetwork is jointly optimized with the backbone student network during the training phase. Its supervision signal comes from manually labeled scene complexity levels, and the mapping relationship is learned through regression.

[0012] As a preferred embodiment of the bird biometric identification method using a lightweight end-side neural network described in this invention, the specific steps of dynamically controlling the execution state of multiple convolutional blocks in the backbone structure based on complexity scores, skipping redundant computational paths, completing adaptive inference, and outputting bird species identification results are as follows: Configure an independent learnable skip threshold for each MBConv convolutional block in the backbone network. ; Complexity score and If a comparison is made, < If the input features are not found in the first convolutional block, the convolutional block is skipped and the input features are directly passed to the next module; otherwise, the complete convolution operation is performed. all Joint optimization during knowledge distillation training enables a synergistic balance between skipping strategies and recognition accuracy, achieving on-demand allocation of computing resources.

[0013] As a preferred embodiment of the bird biometric recognition method using a lightweight edge-side neural network described in this invention, the learnable skip threshold is dynamically adjusted based on the real-time battery status of the edge device, and the adjustment rule is as follows: ; in, This is the normalized value of the remaining battery power of the device. This is the preset sensitivity coefficient; when the battery is low, Enlargement, leading to This increases the number of convolutional blocks, thus triggering more convolutional blocks to be skipped. When the battery is fully charged Close to the original value Preserve the complete calculation path.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the bird biometric identification method using an end-side lightweight neural network as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the bird biometric identification method using an end-side lightweight neural network as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by integrating prior knowledge of bird anatomy to guide the attention mechanism, and combining hierarchical perception knowledge distillation with dynamic reasoning strategies based on scene complexity and device power consumption, the lightweight neural network on the edge can still accurately focus on key biological features with extremely low computational resource consumption, effectively improving the accuracy, energy efficiency and environmental adaptability of fine-grained bird identification, and effectively solving the technical problem of balancing model performance and power consumption on edge devices. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the bird biometric identification method using a lightweight end-side neural network in Example 1. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1, referring to Figure 1 As one embodiment of the present invention, this embodiment provides a bird biometric recognition method using an edge-side lightweight neural network, comprising the following steps: S1. Construct a backbone structure based on a lightweight convolutional network, and embed an attention-guided module that integrates prior knowledge of bird anatomy into the backbone structure to obtain a student network with the ability to focus on key biological features.

[0023] Furthermore, on a standard bird image dataset, professionals with taxonomic backgrounds labeled three anatomical regions with species-discriminating value: the beak region, the eye ring region, and the wing patch region, forming three sets of binary mask images. A two-dimensional Gaussian kernel was applied to each set of binary mask images for smoothing, generating continuously distributed region importance maps. Each region importance map was then compressed into a single-channel weight vector using global average pooling. The dimension of this vector is consistent with the number of channels in the output feature map of the current convolutional block of the backbone network.

[0024] Introduce a weight vector into the channel weight calculation path of the standard SE attention module. As a bias term, the channel attention output is expressed as: ; in, The channel statistics are obtained after global average pooling of the input feature map. and The learnable weight matrices of the two fully connected layers form a bottleneck structure for dimensionality reduction and dimensionality increase. Represents the ReLU activation function; This is the original bias vector; Represents the Sigmoid function; The final modulated channel attention weights; using the channel attention weights Channel-wise scaling is performed on the input feature map to enhance local anatomical features strongly correlated with bird classification. Smoothing is achieved by convolving a binary mask with a two-dimensional Gaussian kernel of fixed standard deviation to ensure smooth transitions at anatomical region boundaries, avoiding gradient discontinuities caused by hard boundaries. This ensures the smooth generation of subsequent weight vectors. It should reasonably reflect the spatial influence range of key areas.

[0025] It should be noted that by transforming key anatomical regions marked by avian taxonomy experts into attention-guided signals and integrating them into the channel attention mechanism of lightweight networks, the model can purposefully focus on the most valuable visual parts for species identification in the early stages of training. This effectively compensates for the shortcomings of small networks that are difficult to autonomously discover fine-grained features due to limited capacity, and enhances its recognition accuracy and biological rationality in resource-constrained scenarios.

[0026] S2. Generate three-level soft labels (family, genus, and species) and intermediate semantic features based on a large-scale cloud-based bird identification model, as a supervisory signal for knowledge distillation.

[0027] Furthermore, the large-scale cloud-based bird recognition model is a deep convolutional neural network pre-trained on a massive bird image dataset. The same input image is fed into both the large-scale cloud-based model and the student network to be trained. The original logical outputs for all bird species are extracted from the classifier head of the large-scale cloud-based model and aggregated at the family, genus, and species levels to form three independent probability distributions, i.e., three-level soft labels. Feature maps are extracted from multiple intermediate convolutional layers in the backbone network of the large-scale cloud-based model corresponding to the Bio-SE module of the student network as high-order semantic features. These three-level soft labels and intermediate layer semantic features together constitute a complete supervision signal to guide the training of the student network.

[0028] It should be noted that the soft labels output by the large cloud model at the three classification levels of family, genus, and species, along with their deep intermediate features, together constitute the supervision signal. This not only conveys the final judgment result of the teacher model but also preserves its semantic understanding ability at different levels of abstraction. This enables the student network to learn macro-classification logic and micro-discrimination details simultaneously while compressing the parameter scale, thereby achieving high-fidelity knowledge transfer.

[0029] S3. Use the supervisory signal to train the student network with hierarchical perception knowledge distillation, so that the student network can inherit high-order classification logic and fine-grained discrimination ability under the condition of limited parameters, and obtain the compressed end-side recognition model.

[0030] Furthermore, the three-level output probability distributions (species, genus, and family) are extracted from the large-scale cloud-based bird recognition model; the corresponding three-level prediction distributions are simultaneously obtained from the student network; the species-level cross-entropy loss is calculated to measure the student network's fit to the true labels; the genus- and family-level Kullback-Leibler divergence losses are calculated to force the student network to mimic the teacher model's output distribution at higher classification levels; the intermediate feature maps of the teacher and student networks at the convolutional layer outputs of the embedded Bio-SE module are extracted, and the feature alignment loss is calculated, expressed as: ; in, This refers to the feature map output by the teacher network at a specific convolutional layer. Feature maps of the corresponding layers in the student network; The global average pooling operation compresses the spatial information of each channel into a single value; the various losses are weighted and summed according to preset weights to form the total loss function, which is used to optimize the student network parameters end-to-end; the feature alignment loss is only implemented at the convolutional block location of the Bio-SE attention module embedded in the backbone network.

[0031] It should be noted that the feature alignment loss only applies to specific layers in the backbone network that have embedded the anatomical prior attention module. This ensures that the distillation process focuses on strengthening key feature pathways that have been enhanced by biological knowledge, avoiding the imposition of invalid constraints on irrelevant or redundant intermediate representations. This improves training efficiency and prevents knowledge interference, enabling the student network's internal representation to more accurately inherit the semantic structures in the teacher model that are most relevant to bird recognition.

[0032] S4. Deploy the compressed edge recognition model on the edge device and configure a lightweight complexity evaluation sub-network at the front end of the model.

[0033] Furthermore, the input image is fed into a subnetwork consisting of two 1x1 convolutional layers. The first layer outputs eight channels, and the second layer outputs a single channel. Global average pooling is performed on the output of the second layer to obtain a scalar value. The sigmoid function is then applied to the scalar value to generate a normalized complexity score. The complexity score reflects the number of targets in the image, the clutter of the background, and the occlusion of the pose. The sub-network is jointly optimized with the backbone student network during the training phase. Its supervision signal comes from the manually labeled scene complexity level, and the mapping relationship is learned through regression.

[0034] It should be noted that the complexity evaluation subnetwork achieves a rapid prediction of the overall recognition difficulty of the input image with a minimal structure. Its output directly reflects key challenge factors in actual field scenarios, such as target clarity, background interference, and occlusion. This subnetwork is jointly trained with the backbone model, making its evaluation results highly task-relevant and providing a reliable and low-overhead decision basis for subsequent dynamic inference.

[0035] S5. Perform scene complexity analysis on the input image through a lightweight complexity evaluation subnetwork and output a complexity score.

[0036] Furthermore, the lightweight complexity evaluation subnetwork is a miniature convolutional neural network independent of the backbone recognition network. The original input image is directly fed into this subnetwork without any preprocessing. The first layer of the subnetwork uses a 1x1 convolutional kernel to compress the number of input channels to 8 to quickly extract preliminary global information. The second layer also uses a 1x1 convolutional kernel to further compress the number of channels to 1, generating a single-channel attention heatmap. Global average pooling is performed on this heatmap to obtain a single real value. This real value is normalized using the Sigmoid activation function and mapped to an open interval of 0 to 1 to obtain the final scene complexity score. Values ​​close to 0 represent simple image content and clear targets, while values ​​close to 1 represent cluttered backgrounds, severe target occlusion, or the presence of multiple interfering objects.

[0037] It should be noted that the scene complexity score describes the difficulty of image content in a continuous manner, which can effectively distinguish between simple and clear samples and complex and interfering samples, providing a basis for the system to process differently; the pre-analysis mechanism avoids the waste of using uniform computation intensity for all inputs, enabling the model to reduce energy consumption and latency when facing a large number of simple scenes, and improve overall operating efficiency.

[0038] S6. Dynamically control the execution state of multiple convolutional blocks in the main structure based on the complexity score, skip redundant calculation paths, complete adaptive inference, and output bird species identification results.

[0039] Furthermore, an independent learnable skip threshold is configured for each MBConv convolutional block in the backbone network. Complexity scoring and If a comparison is made, < If the input features are not found in the first convolutional block, the convolutional block is skipped, and the input features are directly passed to the next module; otherwise, the full convolution operation is performed. Joint optimization during knowledge distillation training achieves a synergistic balance between skipping strategies and recognition accuracy, enabling on-demand allocation of computing resources. The learnable skipping threshold is dynamically adjusted based on the real-time power status of edge devices, with the following adjustment rules: ; in, This is the normalized value of the remaining battery power of the device. This is the preset sensitivity coefficient; when the battery is low, Enlargement, leading to Increase the power level, thus triggering more convolutional blocks to be skipped; when the power is sufficient... Close to the original value Preserve the complete calculation path.

[0040] It should be noted that the skip threshold can be adjusted in real time according to the current power status of the device. When the power is sufficient, priority is given to ensuring recognition accuracy, and when the power is low, the computing intensity is actively reduced to extend the battery life. This mechanism of adaptively adjusting the inference depth according to the hardware environment makes the system more robust and capable of continuous operation in real field deployment, and takes into account the dynamic balance between performance and energy efficiency.

[0041] This embodiment also provides a computer device applicable to the bird biometric identification method using an edge-side lightweight neural network, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the bird biometric identification method using an edge-side lightweight neural network as proposed in the above embodiment.

[0042] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0043] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the bird biometric recognition method based on a lightweight neural network on the edge side as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0044] In summary, this invention integrates prior knowledge of bird anatomy to guide attention, combines hierarchical perception with knowledge distillation and dynamic reasoning strategies based on scene complexity and device power consumption, enabling the lightweight edge neural network to accurately focus on key biological features with extremely low computational resource consumption. This effectively improves the accuracy, energy efficiency, and environmental adaptability of fine-grained bird identification, and effectively solves the technical challenge of balancing model performance and power consumption on edge devices.

[0045] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A bird biometric feature recognition method of end-side lightweight neural network, characterized in that: include: A backbone structure based on a lightweight convolutional network was constructed, and an attention-guided module that integrates prior knowledge of bird anatomy was embedded in the backbone structure to obtain a student network with the ability to focus on key biological features. Based on a large-scale cloud-based bird identification model, three levels of soft labels (family, genus, and species) and intermediate semantic features are generated as supervisory signals for knowledge distillation. By using supervisory signals to train the student network through hierarchical perception knowledge distillation, the student network can inherit high-order classification logic and fine-grained discrimination ability under the condition of limited parameters, resulting in a compressed end-side recognition model. The compressed edge recognition model is deployed on edge devices, and a lightweight complexity evaluation subnetwork is configured at the front end of the model. A lightweight complexity evaluation subnetwork is used to perform scene complexity analysis on the input image and output a complexity score. The execution state of multiple convolutional blocks in the main structure is dynamically controlled based on the complexity score, skipping redundant computation paths, completing adaptive inference, and outputting bird species identification results. 2.The end-side lightweight neural network-based bird biometric feature recognition method of claim 1, wherein: The attention-guided module that integrates prior knowledge of bird anatomy includes the following steps: On a standard bird image dataset, professionals with a taxonomic background annotated three types of anatomical regions with species discrimination value: the beak region, the eye ring region, and the wing patch region, forming three sets of binary mask images; A two-dimensional Gaussian kernel is applied to each set of binary mask images for smoothing, generating a continuously distributed region importance map; each region importance map is compressed into a single-channel weight vector by a global average pooling operation the vector dimension is consistent with the channel number of the feature map output by the current convolution block of the backbone network introducing the weight vector in the channel weight computation path of a standard SE attention module as a bias term, the channel attention output is expressed as: ; wherein, is the channel statistics obtained after global average pooling of the input feature map; and are the learnable weight matrices of two fully connected layers, constituting a bottleneck structure of dimension reduction-dimension increase; denotes a ReLU activation function; is the original bias vector; denotes a Sigmoid function; is the final modulated channel attention weight; Utilizing the channel attention weights Performing per-channel scaling on the input feature map, completing the enhancement of the local anatomical features strongly related to bird classification.

3. The bird biometric recognition method using a lightweight end-side neural network as described in claim 2, characterized in that: The smoothing processing adopts a two-dimensional Gaussian kernel with a fixed standard deviation to perform convolution operation on the binary mask, so as to smoothly transition the boundary of the dissection region and avoid the gradient discontinuity problem caused by the hard boundary, and is used to ensure the weight vector generated subsequently Reasonably reflect the spatial influence range of the key region.

4. The end-side lightweight neural network bird biometric feature recognition method of claim 3, wherein: The method of using supervised signals to perform hierarchical perception knowledge distillation training on the student network enables the student network to inherit high-order classification logic and fine-grained discriminative ability under limited parameter conditions, resulting in a compressed edge recognition model, including: Extract the three-level output probability distributions of species, genus, and family from a large-scale cloud-based bird identification model; The corresponding three-level prediction distribution is obtained synchronously from the student network; Calculate the seed-level cross-entropy loss to measure how well the student network fits the real labels; Calculate the Kullback-Leibler divergence loss at the genus and family levels to force the student network to mimic the output distribution of the teacher model at higher classification levels; Extract the intermediate feature maps from the convolutional layers of the teacher and student networks that embed the Bio-SE module, and calculate the feature alignment loss, expressed as: ; wherein, is the feature map output by the teacher network at a certain convolutional layer, is the feature map of the corresponding layer of the student network; denotes a global average pooling operation that compresses the spatial information of each channel into a single numerical value; The various losses are summed according to preset weights to form the total loss function, which is used to optimize student network parameters end-to-end.

5. The bird biometric recognition method using a lightweight end-side neural network as described in claim 4, characterized in that: The feature alignment loss is implemented only at the convolutional block locations where the Bio-SE attention module is embedded in the backbone network.

6. The bird biometric recognition method using a lightweight end-side neural network as described in claim 5, characterized in that: The compressed edge recognition model is deployed on an edge device, and a lightweight complexity evaluation sub-network is configured at the front end of the model. The lightweight complexity evaluation sub-network performs scene complexity analysis on the input image and outputs a complexity score. The specific steps are as follows: The input image is fed into a sub-network consisting of two 1x1 convolutional layers. The first layer outputs eight channels, and the second layer outputs a single channel. Perform global average pooling on the output of the second layer to obtain a scalar value; Applying the Sigmoid function to the scalar value generates a normalized complexity score. The complexity score reflects the number of targets in the image, the degree of background clutter, and the pose occlusion situation; The subnetwork is jointly optimized with the backbone student network during the training phase. Its supervision signal comes from manually labeled scene complexity levels, and the mapping relationship is learned through regression.

7. The bird biometric recognition method using a lightweight end-side neural network as described in claim 6, characterized in that: The specific steps for dynamically controlling the execution state of multiple convolutional blocks in the backbone structure based on complexity scores, skipping redundant computation paths, completing adaptive inference, and outputting bird species identification results are as follows: Configure an independent learnable skip threshold for each MBConv convolutional block in the backbone network. ; Complexity score and If a comparison is made, < If the input features are not found in the first convolutional block, the convolutional block is skipped and the input features are directly passed to the next module; otherwise, the complete convolution operation is performed. all Joint optimization during knowledge distillation training enables a synergistic balance between skipping strategies and recognition accuracy, achieving on-demand allocation of computing resources.

8. The bird biometric recognition method using a lightweight end-side neural network as described in claim 7, characterized in that: The learnable skip threshold is dynamically adjusted based on the real-time power status of the edge device, and the adjustment rule is as follows: ; in, This is the normalized value of the remaining battery power of the device. This is the preset sensitivity coefficient; when the battery is low, Enlargement, leading to This increases the number of convolutional blocks, thus triggering more convolutional blocks to be skipped. When the battery is fully charged Close to the original value Preserve the complete calculation path.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the bird biometric identification method using an end-side lightweight neural network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the bird biometric identification method of the end-side lightweight neural network according to any one of claims 1 to 7.

Images