An image recognition and classification method based on deep learning

Abstract

The application relates to a deep learning-based image recognition and classification method, and relates to the technical field of biometric feature recognition and deep learning, and comprises the following steps: collecting biometric feature original image data through a multimodal sensor of an electronic device, and using a generative model to synthesize and enhance the biometric feature original image data to construct a training data set. The deep learning-based image recognition and classification method constructs a dynamic sparse biometric feature backbone network and integrates a self-adaptive gating mechanism and a channel pruning module, realizes on-demand allocation of computing resources, can dynamically adjust a network activation level and remove low-contribution neuron channels in real time according to the definition, the degree of occlusion and the background complexity of an input image, thereby significantly reduces the reasoning algorithm consumption and the memory occupation of a mobile terminal device, guarantees high recognition accuracy in a complex scene, and solves the problem of low running efficiency of a traditional static model on a resource-limited device.

Description

Technical Field

[0001] This invention relates to the fields of biometric recognition and deep learning technology, specifically to an image recognition and classification method based on deep learning. Background Technology

[0002] With the rapid development of biometric recognition technology, deep learning-based image recognition and classification methods have been widely applied in scenarios such as facial recognition access control, fingerprint payment, and iris verification. In existing technologies, convolutional neural networks (CNNs) or Transformers are typically used as the backbone network. These technologies collect multimodal biometric image data, such as visible light, infrared, or depth images, from electronic devices (e.g., cameras, sensors). Supervised learning is then performed using large-scale labeled datasets to train the feature extraction model. During the recognition process, the system maps the input image into a high-dimensional feature vector and makes classification decisions by calculating the distance or similarity between feature vectors. To improve the generalization ability of the model, some existing technologies have also introduced data augmentation strategies and transfer learning techniques to try to solve the problem of insufficient samples in specific scenarios. In addition, to improve security, some solutions attempt to add a liveness detection module after feature extraction, which distinguishes between real-person attacks by analyzing the texture features or simple motion information of the image.

[0003] However, existing deep learning-based biometric recognition methods still have significant technical shortcomings in practical applications. First, existing models are highly dependent on labeled data and struggle to effectively handle long-tailed distribution scenarios, leading to decreased recognition accuracy when registering small samples. Second, traditional static network architectures cannot dynamically adjust computational resources based on the complexity of the input image, resulting in wasted computing power or excessive latency on mobile devices, and lacking channel-level pruning mechanisms to optimize inference speed. Third, existing liveness detection methods are often independent of the main recognition network, making them vulnerable to sophisticated adversarial attacks, high-resolution photo replay attacks, or 3D mask attacks, and lacking quantitative evaluation mechanisms for the uncertainty of recognition results. Furthermore, the inherent "black box" nature of deep learning models makes recognition decisions lack interpretability and fails to provide specific geometric or texture matching criteria to meet compliance requirements in high-security scenarios. Finally, existing systems are prone to catastrophic forgetting when continuously learning new biometrics, and the lack of effective privacy protection mechanisms for centralized data storage or transmission poses a risk of biometric data leakage. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an image recognition and classification method based on deep learning, which has advantages such as dynamic sparse network resource adaptive allocation and solves the problem of serious computational waste in static models in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an image recognition and classification method based on deep learning, comprising the following steps:

[0006] Step S1: Collect raw biometric image data through the multimodal sensor of the electronic device, use a generative model to synthesize and enhance the raw biometric image data to construct a training dataset, and use a contrastive self-supervised learning algorithm to pre-train a feature extraction network on unlabeled biometric images;

[0007] Step S2: Construct a dynamic sparse biometric backbone network. The dynamic sparse biometric backbone network includes an adaptive gating module and a channel pruning module. Based on the clarity, occlusion degree and background complexity of the input biometric image, the activation state of the network layers is dynamically adjusted through the adaptive gating module, and the channel pruning module removes neuron channels in the current calculation path whose contribution is lower than a preset threshold in real time.

[0008] Step S3: Embed a liveness detection branch in the middle layer of the dynamic sparse biometric backbone network, use remote photoplethysmography pulse wave technology or micro-motion texture analysis algorithm to detect blood flow changes or micro-vibrations in the biometric region, and at the same time perform perturbation sample testing generated by adversarial training, and output the predicted classification result and its corresponding uncertainty quantification score.

[0009] Step S4: Map the biometric key point coordinates and texture feature vectors output by the feature extraction network into atomic facts, input them into the neural symbolic logic reasoning engine, perform logical verification on the atomic facts according to the preset biometric knowledge graph rule base, correct the preliminary classification results, and generate a classification report containing feature matching basis.

[0010] Step S5: Initialize the classifier parameters using the model-independent meta-learning algorithm. When a new user registers, receive a very small number of biometric samples and use the pre-trained general feature space for rapid fine-tuning to generate the user's unique feature vector. This allows for updating the unique feature vector based on incremental samples in subsequent use.

[0011] Step S6: Establish a non-volatile memory buffer in the local storage medium to store compressed representation data of historical biometric samples. When performing a new category learning task, use the elastic weight consolidation algorithm to calculate the penalty term of the key parameters of the old task and limit the parameter update range to preserve the recognition performance of the old task.

[0012] Step S7: Perform the biometric recognition inference process locally on the electronic device. Do not upload the original biometric image data. Only upload the model parameter gradients after differential privacy processing to the server for federated aggregation. Apply homomorphic encryption to the biometric template during storage and transmission.

[0013] Furthermore, in step S1, the multimodal sensor includes a visible light camera, an infrared camera, and a depth structured light sensor; the generative model is a diffusion model, and the synthetic augmented data it generates includes image samples covering different light intensities, different angles and poses, scenes with occlusions, and scenes with impaired biometric features.

[0014] Furthermore, in step S1, the contrastive self-supervised learning algorithm adopts a positive and negative sample contrast loss function, where a positive sample pair is defined as an image pair of the same biological individual at different times or in different modalities, and a negative sample pair is defined as an image pair of different biological individuals, in order to optimize the intra-class compactness and inter-class separability of the feature extraction network in the feature space.

[0015] Furthermore, in step S2, the dynamic sparse biometric backbone network adopts a hybrid architecture of convolutional neural network and Transformer; the specific logic of the adaptive gating module is as follows: when the signal-to-noise ratio of the input image is detected to be higher than the first threshold and the proportion of the unoccluded area is higher than the second threshold, the deep attention module is skipped and only the shallow feature extraction module is activated; when the input image is detected to be blurry, occluded or has a cluttered background, the deep attention module is activated to enhance the global feature extraction capability.

[0016] Furthermore, in step S3, the liveness detection branch extracts the color component temporal signal of the facial skin region, calculates the pulse amplitude change rate, and distinguishes the physiological signals generated by blood flow in a real face from those that cannot be generated by photos, video playback, or 3D mask models.

[0017] Furthermore, in step S3, the uncertainty quantification score is obtained through the Monte Carlo Dropout mechanism, specifically by randomly discarding some neurons multiple times during the inference phase and repeatedly calculating, and statistically analyzing the variance of the multiple output results as the uncertainty quantification score; when the uncertainty quantification score exceeds a preset safety threshold, a secondary verification process is triggered, requiring the electronic device to control the user to perform blinking, mouth opening, or head turning actions.

[0018] Furthermore, in step S4, the biometric knowledge graph rule base stores biological anatomical geometric constraint rules, including pupil distance range constraints, nasal bridge height ratio constraints, and skin texture continuity constraints; the logical verification specifically involves: if the geometric attributes in the atomic facts violate the anatomical constraint rules, it is determined to be a forgery attack and the classification result is corrected to "invalid"; if the atomic facts conform to the anatomical constraint rules, the confidence of the original classification result is maintained or increased.

[0019] Furthermore, in step S4, the classification report includes the following structured data fields: the unique identifier ID of the identified object, the classification matching degree value, the deviation of key biometric points, the texture hash value comparison result, and the description of the logic path for excluding liveness attacks.

[0020] Furthermore, in step S5, the rapid fine-tuning process includes: inputting a very small number of biometric samples into the meta-learner, calculating the gradient of the loss function with respect to the model parameters, and updating the model parameters to minimize the classification error on the small number of samples, thereby completing the construction of the exclusive feature vector without retraining the entire network weights.

[0021] Furthermore, in step S6, the compressed representation data is the feature vector after dimensionality reduction by principal component analysis; in step S7, the differential privacy processing is achieved by adding Laplacian noise to the gradient of the model parameters, and the homomorphic encryption uses the Paillier encryption algorithm or the ElGamal encryption algorithm to encrypt, store, and transmit the biometric template.

[0022] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0023] 1. This deep learning-based image recognition and classification method constructs a dynamic sparse biometric backbone network and integrates an adaptive gating mechanism and a channel pruning module to achieve on-demand allocation of computing resources. This method can dynamically adjust the network activation level and remove low-contribution neuron channels in real time according to the clarity, occlusion degree and background complexity of the input image. This significantly reduces the inference computing power consumption and memory occupation of mobile devices while ensuring high recognition accuracy in complex scenes, and solves the problem of low running efficiency of traditional static models on resource-constrained devices.

[0024] 2. This deep learning-based image recognition and classification method effectively improves the security, robustness, and interpretability of biometric recognition systems by integrating a neural symbolic logic reasoning engine with a multimodal liveness detection branch. Utilizing remote photoplethysmography (LPP) technology combined with anatomical geometric constraints, it can accurately distinguish between real human physiological signals and forgeries such as photos, videos, or 3D masks. Furthermore, it outputs uncertainty quantification scores through a Monte Carlo Dropout mechanism to trigger secondary verification, significantly reducing the false recognition rate. Simultaneously, it transforms deep network features into atomic facts and performs logical verification, generating classification reports containing specific matching criteria, meeting the transparency requirements of high-security scenarios. In addition, it employs meta-learning-driven small-sample rapid registration, a lifelong memory mechanism with elastic weight consolidation, and a privacy-preserving architecture based on federated learning and homomorphic encryption. This achieves rapid new user data entry, anti-forgetting updates of old user features, and full lifecycle security of biometric data, overcoming the shortcomings of existing technologies such as strong data dependence, poor continuous learning ability, and high privacy leakage risks. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the identification and classification method of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] A deep learning-based image recognition and classification method includes the following steps:

[0028] Step S1: Collect raw biometric image data through the multimodal sensor of the electronic device, use a generative model to synthesize and enhance the raw biometric image data to construct a training dataset, and use a contrastive self-supervised learning algorithm to pre-train a feature extraction network on unlabeled biometric images;

[0029] Step S2: Construct a dynamic sparse biometric backbone network. The dynamic sparse biometric backbone network includes an adaptive gating module and a channel pruning module. Based on the clarity, occlusion degree and background complexity of the input biometric image, the activation state of the network layers is dynamically adjusted through the adaptive gating module, and the channel pruning module removes neuron channels in the current calculation path whose contribution is lower than a preset threshold in real time.

[0030] Step S3: Embed a liveness detection branch in the middle layer of the dynamic sparse biometric backbone network, use remote photoplethysmography pulse wave technology or micro-motion texture analysis algorithm to detect blood flow changes or micro-vibrations in the biometric region, and at the same time perform perturbation sample testing generated by adversarial training, and output the predicted classification result and its corresponding uncertainty quantification score.

[0031] Step S4: Map the coordinates of biometric key points and texture feature vectors output by the feature extraction network into atomic facts, input them into the neural symbolic logic reasoning engine, perform logical verification on the atomic facts according to the preset biometric knowledge graph rule base, correct the preliminary classification results, and generate a classification report containing feature matching basis.

[0032] Step S5: Initialize the classifier parameters using the model-independent meta-learning algorithm. When a new user registers, receive a very small number of biometric samples and use the pre-trained general feature space for rapid fine-tuning to generate the user's unique feature vector. This allows for updating the unique feature vector based on incremental samples in subsequent use.

[0033] Step S6: Establish a non-volatile memory buffer in the local storage medium to store compressed representation data of historical biometric samples. When performing a new category learning task, use the elastic weight consolidation algorithm to calculate the penalty term of the key parameters of the old task and limit the parameter update range to preserve the recognition performance of the old task.

[0034] Step S7: Perform the biometric recognition inference process locally on the electronic device. Do not upload the original biometric image data. Only upload the model parameter gradients that have undergone differential privacy processing to the server for federated aggregation. Apply homomorphic encryption to the biometric template during storage and transmission.

[0035] It should be noted that by constructing a full-process approach that includes multimodal acquisition, generative data augmentation, contrastive self-supervised pre-training, dynamic sparse backbone network, liveness detection branch, neural symbolic logic reasoning, meta-learning fast registration, elastic weight consolidation, and federated privacy protection, this approach solves the comprehensive technical problems of traditional biometric recognition systems, such as high dependence on labeled data, static architecture inability to adapt to complex scenarios, lack of effective liveness attack defense mechanisms, unexplainable decisions, susceptibility to catastrophic forgetting, and the risk of data leakage.

[0036] In step S1, the multimodal sensor includes a visible light camera, an infrared camera, and a depth structured light sensor; the generative model is a diffusion model, and the generated synthetic augmented data includes image samples covering different light intensities, different angles and poses, scenes with occlusions, and scenes with damaged biometric features.

[0037] It should be noted that by limiting the specific combination of multimodal sensors (visible light, infrared, structured light) and diffusion model generative enhancement, it is possible to synthesize diverse samples covering extreme lighting, multiple poses, occlusion and damage scenarios, thereby significantly enhancing the generalization ability of the feature extraction network in long-tailed distribution and harsh environments, and solving the problems of difficulty in acquiring single-modal data and insufficient diversity of synthesized data.

[0038] In step S1, the contrastive self-supervised learning algorithm uses a positive and negative sample contrast loss function, where a positive sample pair is defined as an image pair of the same biological individual at different times or in different modalities, and a negative sample pair is defined as an image pair of different biological individuals, in order to optimize the intra-class compactness and inter-class separability of the feature extraction network in the feature space.

[0039] It should be noted that the positive and negative sample contrast loss function is used for self-supervised pre-training. Images of the same biological individual at different times or modalities are used as positive sample pairs, and images of different individuals are used as negative sample pairs. This optimizes the intra-class compactness and inter-class separation in the feature space and solves the problems of low utilization of unlabeled data and low feature discrimination.

[0040] In step S2, the dynamic sparse biometric backbone network adopts a hybrid architecture of convolutional neural network and Transformer. The specific logic of the adaptive gating module is as follows: when the signal-to-noise ratio of the input image is detected to be higher than the first threshold and the proportion of the unoccluded area is higher than the second threshold, the deep attention module is skipped and only the shallow feature extraction module is activated; when the input image is detected to be blurry, occluded or cluttered, the deep attention module is activated to enhance the global feature extraction capability.

[0041] It should be noted that by dynamically configuring the adaptive gating module and the channel pruning module, the activation state of shallow or deep networks is intelligently adjusted and low-contribution channels are removed based on the signal-to-noise ratio, occlusion, and background complexity of the input image. This achieves on-demand allocation of computing resources, resolves the contradiction between the limited computing power of mobile devices and the requirement for high recognition accuracy, and reduces inference latency and energy consumption.

[0042] In step S3, the liveness detection branch extracts the color component temporal signal of the facial skin region, calculates the pulse amplitude change rate, and distinguishes the physiological signals generated by blood flow in a real face from those that cannot be generated by photos, video playback, or 3D mask models.

[0043] It should be noted that by using rPPG technology to extract the temporal signal of facial skin color components and calculate the rate of change of pulse wave amplitude, the physiological blood flow characteristics of real people can be accurately captured to distinguish between photos, video playback, or 3D masks. This solves the technical defect of traditional liveness detection methods based on texture or simple motion analysis, which are easily deceived by high-resolution forgery attacks.

[0044] In step S3, the uncertainty quantification score is obtained through the Monte Carlo Dropout mechanism. Specifically, during the inference phase, some neurons are randomly dropped multiple times and the calculation is repeated. The variance of the multiple output results is used as the uncertainty quantification score. When the uncertainty quantification score exceeds the preset safety threshold, a secondary verification process is triggered, requiring the electronic device to control the user to perform blinking, mouth opening, or head turning actions.

[0045] It should be noted that the Monte Carlo Dropout mechanism is introduced to quantify the uncertainty score of the inference result, and active interaction verification such as blinking and opening the mouth is triggered when the threshold is exceeded. This constructs a closed loop of secondary verification based on confidence, which effectively reduces the false recognition rate caused by environmental interference or adversarial examples and improves the security of the system in high-risk scenarios.

[0046] In step S4, the biometric knowledge graph rule base stores biological anatomical geometric constraint rules, including pupil distance range constraints, nasal bridge height ratio constraints, and skin texture continuity constraints. The logical verification is as follows: if the geometric attributes in the atomic facts violate the anatomical constraint rules, it is determined to be a forgery attack and the classification result is corrected to "invalid"; if the atomic facts conform to the anatomical constraint rules, the confidence of the original classification result is maintained or increased.

[0047] It should be noted that by embedding anatomical geometric constraints (such as pupil distance, nasal bridge ratio, and texture continuity) into the logical verification process, the classification results are corrected by verifying whether atomic facts conform to prior knowledge of biological anatomy. This solves the problem that deep learning "black box" models lack interpretability and are difficult to identify sophisticated forgery attacks from a geometric logic level.

[0048] In step S4, the classification report includes the following structured data fields: the unique identifier ID of the identified object, the classification matching degree value, the deviation of key biometric points, the texture hash value comparison result, and the description of the logic path for excluding liveness attacks.

[0049] It should be noted that a structured classification report format is defined, which includes identity identifiers, matching degree, deviation, hash comparison, and logical path. This provides a clear chain of evidence and traceability basis for identification decisions, meeting the stringent requirements of transparency and audit interpretability for algorithmic decisions in highly compliant scenarios such as finance and security.

[0050] In step S5, the rapid fine-tuning process includes: inputting a very small number of biometric samples into the meta-learner, calculating the gradient of the loss function with respect to the model parameters, and updating the model parameters to minimize the classification error on the small number of samples, thereby completing the construction of the exclusive feature vector without retraining the entire network weights.

[0051] It should be noted that the Model-Independent Meta-Learning (MAML) algorithm is used to quickly fine-tune the model within a general feature space using a very small number of new samples to construct a unique feature vector. This allows for new user registration without retraining the entire network, thus solving the problems of high training cost, slow convergence, and difficulty in cold start for traditional deep learning models in small sample scenarios.

[0052] In step S6, the compressed representation data is the feature vector after dimensionality reduction by principal component analysis; in step S7, differential privacy processing is achieved by adding Laplacian noise to the gradient of the model parameters, and homomorphic encryption uses the Paillier encryption algorithm or the ElGamal encryption algorithm to encrypt the storage and transmission of biometric templates.

[0053] This invention constructs a highly generalizable training set through multimodal biometric feature acquisition and generative data augmentation, utilizes contrastive self-supervised learning to pre-train a feature extraction network, and then dynamically adjusts the sparse structure and computational resources of the backbone network according to the complexity of the input image. In the recognition stage, it integrates rPPG physiological signal detection, Monte Carlo Dropout uncertainty quantification, and a neural symbolic logic reasoning engine based on anatomical geometric constraints to perform liveness attack defense and interpretability verification on features and output a structured decision report. At the same time, it combines meta-learning to achieve rapid registration of small samples, elastic weight consolidation to prevent catastrophic forgetting, and uses federated learning and homomorphic encryption technology to ensure data privacy and security throughout the entire lifecycle, thus forming an intelligent biometric recognition closed-loop system that combines high accuracy, high security, strong real-time performance, and interpretability.

[0054] It should be noted that by combining principal component analysis for dimensionality reduction and compressed representation, differential privacy gradient uploading, and homomorphic encrypted template protection, federated aggregation and local secure storage are achieved without leaking the original biometric data. This solves the privacy leakage risk caused by centralized storage of biometric data and the problem of forgetting the performance of old tasks during continuous learning.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based image recognition and classification method, characterized in that, Includes the following steps: Step S1: Collect raw biometric image data through the multimodal sensor of the electronic device, use a generative model to synthesize and enhance the raw biometric image data to construct a training dataset, and use a contrastive self-supervised learning algorithm to pre-train a feature extraction network on unlabeled biometric images; Step S2: Construct a dynamic sparse biometric backbone network. The dynamic sparse biometric backbone network includes an adaptive gating module and a channel pruning module. Based on the clarity, occlusion degree and background complexity of the input biometric image, the activation state of the network layers is dynamically adjusted through the adaptive gating module, and the channel pruning module removes neuron channels in the current calculation path whose contribution is lower than a preset threshold in real time. Step S3: Embed a liveness detection branch in the middle layer of the dynamic sparse biometric backbone network, use remote photoplethysmography pulse wave technology or micro-motion texture analysis algorithm to detect blood flow changes or micro-vibrations in the biometric region, and at the same time perform perturbation sample testing generated by adversarial training, and output the predicted classification result and its corresponding uncertainty quantification score. Step S4: Map the biometric key point coordinates and texture feature vectors output by the feature extraction network into atomic facts, input them into the neural symbolic logic reasoning engine, perform logical verification on the atomic facts according to the preset biometric knowledge graph rule base, correct the preliminary classification results, and generate a classification report containing feature matching basis. Step S5: Initialize the classifier parameters using the model-independent meta-learning algorithm. When a new user registers, receive a very small number of biometric samples and use the pre-trained general feature space for rapid fine-tuning to generate the user's unique feature vector. This allows for updating the unique feature vector based on incremental samples in subsequent use. Step S6: Establish a non-volatile memory buffer in the local storage medium to store compressed representation data of historical biometric samples. When performing a new category learning task, use the elastic weight consolidation algorithm to calculate the penalty term of the key parameters of the old task and limit the parameter update range to preserve the recognition performance of the old task. Step S7: Perform the biometric recognition inference process locally on the electronic device. Do not upload the original biometric image data. Only upload the model parameter gradients after differential privacy processing to the server for federated aggregation. Apply homomorphic encryption to the biometric template during storage and transmission.

2. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S1, the multimodal sensor includes a visible light camera, an infrared camera, and a depth structured light sensor; the generative model is a diffusion model, and the synthetic augmented data it generates includes image samples covering different light intensities, different angles and poses, scenes with occlusions, and scenes with impaired biometric features.

3. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S1, the contrastive self-supervised learning algorithm uses a positive and negative sample contrast loss function, where a positive sample pair is defined as an image pair of the same biological individual at different times or in different modalities, and a negative sample pair is defined as an image pair of different biological individuals, in order to optimize the intra-class compactness and inter-class separability of the feature extraction network in the feature space.

4. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S2, the dynamic sparse biometric backbone network adopts a hybrid architecture of convolutional neural network and Transformer; the specific logic of the adaptive gating module is as follows: when the signal-to-noise ratio of the input image is detected to be higher than the first threshold and the proportion of the unoccluded area is higher than the second threshold, the deep attention module is skipped and only the shallow feature extraction module is activated; when the input image is detected to be blurry, occluded or has a cluttered background, the deep attention module is activated to enhance the global feature extraction capability.

5. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S3, the liveness detection branch extracts the color component temporal signal of the facial skin region, calculates the pulse amplitude change rate, and distinguishes the physiological signals generated by blood flow in a real face from those that cannot be generated by photos, video playback, or 3D mask models.

6. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S3, the uncertainty quantification score is obtained through the Monte Carlo Dropout mechanism, specifically by randomly discarding some neurons multiple times during the inference phase and repeatedly calculating, and statistically analyzing the variance of the multiple output results as the uncertainty quantification score; when the uncertainty quantification score exceeds a preset safety threshold, a secondary verification process is triggered, requiring the electronic device to control the user to perform blinking, opening mouth, or turning head actions.

7. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S4, the biometric knowledge graph rule base stores biological anatomical geometric constraint rules, including pupil distance range constraints, nasal bridge height ratio constraints, and skin texture continuity constraints; the logical verification is specifically as follows: if the geometric attributes in the atomic facts violate the anatomical constraint rules, it is determined to be a forgery attack and the classification result is corrected to "invalid"; if the atomic facts conform to the anatomical constraint rules, the confidence of the original classification result is maintained or increased.

8. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S4, the classification report includes the following structured data fields: the unique identifier ID of the identified object, the classification matching degree value, the deviation of key biometric points, the texture hash value comparison result, and the description of the logic path for excluding liveness attacks.

9. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S5, the rapid fine-tuning process includes: inputting a very small number of biometric samples into the meta-learner, calculating the gradient of the loss function with respect to the model parameters, and updating the model parameters to minimize the classification error on the small number of samples, thereby completing the construction of the exclusive feature vector without retraining the entire network weights.

10. The image recognition and classification method based on deep learning according to claim 1, characterized in that: In step S6, the compressed representation data is the feature vector after dimensionality reduction by principal component analysis; in step S7, the differential privacy processing is achieved by adding Laplacian noise to the gradient of the model parameters, and the homomorphic encryption uses the Paillier encryption algorithm or the ElGamal encryption algorithm to encrypt, store and transmit the biometric template.

Images