Image Multi-Task Recognition Method and System Based on Progressive Learning

Abstract

This invention proposes an image multi-task recognition method and system based on advanced learning, constructing an end-to-end advanced multi-task deep learning network model for simultaneous facial attribute recognition. This advanced multi-task learning network consists of three main modules: a primary network, an advanced network, and an expert-level network. The primary network is constructed using a lifelong learning strategy, training on facial attribute samples to acquire basic facial attribute recognition capabilities. The advanced network is constructed using a multi-task learning strategy, training on specially collected facial multi-attribute samples for multi-task collaborative learning, enabling advanced recognition of multiple facial attributes simultaneously. The expert-level network is constructed using a soft-label learning strategy, training on specially collected, confidence-based facial multi-attribute training samples for discriminative learning, enabling expert-level identification capabilities that see beyond appearances.

Description

Technical Field

[0001] This invention belongs to the field of machine learning, specifically relating to image processing and pattern recognition. Background Technology

[0002] In multi-task image recognition, the traditional approach is to train a deep neural network for each task and tackle them one by one. This approach has several drawbacks: ① It isolates each task, preventing cross-task learning (co-learning and transfer learning). ② It suffers from few-shot learning problems. ③ It leads to duplication and waste of computational and storage resources. To address these issues, researchers have proposed multi-task deep learning models that utilize a single model for simultaneous multi-task co-learning. However, because different tasks have varying levels of difficulty, trying to accommodate the majority often results in difficult-to-classify tasks dragging down easier-to-classify tasks. Furthermore, current multi-task learning only involves co-learning through network parameter sharing and does not perform task-specific knowledge transfer learning. Finally, the ease of collecting single-task training samples compared to multi-task training samples hinders the application of multi-task deep learning. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes an advanced multi-task learning method, which aims to construct an end-to-end advanced multi-task deep learning network model for simultaneous multi-task image recognition.

[0004] This invention proposes a multi-task facial attribute recognition method based on progressive learning, constructing an end-to-end progressive multi-task deep learning network model for simultaneous recognition of multiple facial attributes. This progressive multi-task learning network consists of three main modules: a primary network, a high-level network, and an expert-level network. The primary network is constructed using a lifelong learning strategy, training on facial attribute samples to acquire basic recognition capabilities for each attribute. The high-level network is constructed using a multi-task learning strategy, training on specially collected multi-attribute facial samples for multi-task collaborative learning, enabling advanced recognition of multiple facial attributes simultaneously. The expert-level network is constructed using a soft-label learning strategy, training on specially collected confidence-based multi-attribute facial training samples for discriminative learning, enabling expert-level identification capabilities that see beyond appearances.

[0005] To achieve the above objectives, the technical solution provided by this invention is: an image multi-task recognition method based on progressive learning, comprising the following steps:

[0006] Step 1: Construct multi-level training samples;

[0007] Step 2: Construct a multi-level image recognition network model with progressive learning;

[0008] The advanced learning multi-level image recognition network model consists of three stages. The first stage utilizes lifelong learning to build a primary network, which is trained using single-attribute samples to enable the model to initially recognize the performance of various tasks. The second stage, based on the transfer of knowledge from the primary network, utilizes multi-attribute samples for collaborative training and learning, and uses multi-task collaborative learning to build an advanced network, enabling the model to fully utilize the correlation and heterogeneity between different tasks and complete the function of simultaneous recognition of multiple tasks. The third stage, based on the transfer of knowledge from the primary and advanced networks, utilizes soft-labeled samples for training and learning to obtain an expert network, and uses soft-label learning to build an expert network, enabling the model to have the essential image recognition ability to distinguish between true and false images, thereby improving the robustness and generalization ability of the model.

[0009] Step 3: Train the image recognition network model using training samples of the corresponding level;

[0010] Step 4: Use the trained multi-level image recognition network model to achieve multi-task image recognition.

[0011] Furthermore, the multi-level training samples constructed in step 1 include single-label primary network training samples, multi-label advanced network training samples, and multi-label expert network training samples with confidence.

[0012] Furthermore, the advanced learning multi-level image recognition network model performs knowledge transfer in two ways: first, input transfer: to utilize the experience of previous hierarchical networks, the current hierarchical network inputs the current task's input into all previous hierarchical networks, and inputs the outputs of each layer of the previous hierarchical networks and the outputs of each layer of the current hierarchical network into the next layer; second, output transfer: to utilize the knowledge of previous hierarchical networks, the loss function of each hierarchical network is designed as follows:

[0013]

[0014] Where i and j represent class, y i This is the actual result of the current class network, q j is the output of the previous class network, pred is the output of the current class network, and loss represents the loss function.

[0015] Furthermore, in order to perform knowledge transfer and avoid catastrophic forgetting, the primary network, after training for each attribute task, calculates the importance Ω of each parameter in the network to that task. ij That is, the proportion of the parameter value in the i-th row and j-th column to the total parameter value, which is used in subsequent training tasks; Ω ij It is added to the loss function as a regularization term whenever a new task is trained: for Ω ijFor larger parameters, minimize their changes during gradient descent, as these parameters are important for past tasks and their values ​​need to be preserved to avoid catastrophic forgetting; while for Ω... ij Smaller parameters allow for gradient updates with larger magnitudes, resulting in better performance on new tasks. Therefore, the loss function for the nth task is:

[0016]

[0017] Among them, L n (θ) is the loss function for the current task, λ is the balance factor, and θ ij These are the parameters in the i-th row and j-th column of the current model parameter matrix. Ω consists of model parameters obtained after training the first n-1 tasks. After training on each task, Ω... ij They will all be updated, especially when performing the first task. Take 0.

[0018] Furthermore, the primary network takes single-attribute training samples as input and outputs single-attribute labels, with its backbone network employing RestNet or EfficientNet.

[0019] Furthermore, a high-level network is constructed using multi-task collaborative learning. This network learns collaboratively using multi-attribute samples based on the transfer of knowledge from the primary network. The backbone of the high-level network is the same as that of the primary network.

[0020] Suppose we use N training samples to learn T tasks, using Let represent the i-th sample vector in the t-th task ∈ {1,…,T} and its corresponding task label. Then the objective of multi-task collaborative learning is:

[0021]

[0022] Where, x i This represents the i-th training sample. Let f(x) represent the label of the i-th sample in the t-th attribute task. i Φ(θ) represents the deep network function, i.e., the backbone network of the high-level network, where θ is the network parameter to be solved, and Φ(θ) is the regularization penalty term, using the L2 penalty term. t (·) represents the loss function for different tasks, λ t This represents the weight allocation for different loss functions.

[0023] Furthermore, the backbone of the expert network is the same as that of the primary network. In the expert network, hard labels y are used... k Softening yields a new loss function:

[0024]

[0025] in, This indicates that the label has been softened. The value is determined based on the specific circumstances of the sample, c k p represents the confidence level of the sample label. k For the network prediction results, y k The result is either 1 or 0, where K is the number of label categories.

[0026] Furthermore, in the process of multi-attribute recognition of facial images, a cross-validation strategy is adopted to find the difficulty of each task and the learning ranking problem. The final recognition order is determined as gender, illumination, resolution, pose, age, and expression.

[0027] Furthermore, a pre-trained multi-level image recognition network model is used to identify the gender, age, expression, pose, lighting, and resolution attributes of facial images.

[0028] On the other hand, the present invention provides an image multi-task recognition system based on progressive learning, comprising the following modules:

[0029] The training sample construction module is used to construct training samples at multiple levels;

[0030] The model building module is used to build multi-level image recognition network models that enable progressive learning.

[0031] The advanced learning multi-level image recognition network model consists of three stages. The first stage utilizes lifelong learning to build a primary network, which is trained using single-attribute samples to enable the model to initially recognize the performance of various tasks. The second stage, based on the transfer of knowledge from the primary network, utilizes multi-attribute samples for collaborative training and learning, and uses multi-task collaborative learning to build an advanced network, enabling the model to fully utilize the correlation and heterogeneity between different tasks and complete the function of simultaneous recognition of multiple tasks. The third stage, based on the transfer of knowledge from the primary and advanced networks, utilizes soft-labeled samples for training and learning to obtain an expert network, and uses soft-label learning to build an expert network, enabling the model to have the essential image recognition ability to distinguish between true and false images, thereby improving the robustness and generalization ability of the model.

[0032] The model training module is used to train the image recognition network model using training samples of the corresponding level.

[0033] The recognition module is used to achieve multi-task image recognition using a trained multi-level image recognition network model.

[0034] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0035] This invention mimics the human learning process, fully utilizing lifelong learning, multi-task collaborative learning, and soft-label learning to construct an advanced multi-task learning network for simultaneous multi-task image recognition. Its significant features are:

[0036] 1. A unique multi-task learning network structure. Compared to existing progressive networks: ① Each level inputs the same type of image, thus better sharing of information from lower-level networks. ② Each level has different image labels (single-label, multi-label, and multi-label with confidence). ③ The output labels of each level are not independent; there are correlations and inclusion relationships between the output labels of different levels, thus facilitating knowledge transfer through loss function design. ④ Progressive multi-task learning networks do not need to focus too much on the "catastrophic forgetting" problem encountered in lifelong learning; they only need to consider the knowledge transfer problem between different levels, thus reducing the complexity of the problem.

[0037] 2. Wide range of applications. This advanced learning network can be widely used in multi-task recognition of face images, vehicle images, character images, and clothing images, etc. Attached Figure Description

[0038] Figure 1 A schematic diagram of the structure of a multi-task image recognition model based on progressive learning;

[0039] Figure 2 A technical roadmap for multi-task face image recognition.

[0040] Figure 3 A schematic diagram of an advanced multi-task learning network structure used for multi-attribute face recognition. Detailed Implementation

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0043] This embodiment takes the multi-task recognition of six facial attributes, including gender, age, expression, pose, lighting, and resolution, as an example to further explain the image multi-task recognition method based on progressive learning.

[0044] A multi-task facial attribute recognition method based on progressive learning, the overall technical approach is as follows: Figure 2As shown, the input is a face image with arbitrary attributes, and the output is a multi-attribute face category. It mainly includes two parts: construction of multi-attribute face training samples (completed offline) and face attribute analysis.

[0045] 1. Construction of multi-attribute training samples for faces

[0046] Multi-attribute facial training samples are the foundation and prerequisite for advanced learning-based multi-task facial attribute recognition. Since the construction of facial samples is done offline, it can fully utilize 3D scanning equipment, publicly available facial databases, and human interaction. This invention constructs multi-attribute facial training samples based on facial attributes such as gender, resolution, pose, lighting, expression, and age to serve different network models.

[0047] 1.1 Construction of Initial Network Training Samples

[0048] Since primary networks based on lifelong learning are trained task-by-task and perform task knowledge transfer, their training samples are relatively easy to construct. A suitable training library can be easily built using publicly available face databases (with attributes) and simple annotations of faces collected online.

[0049] 1.2 Construction of Advanced Network Training Samples

[0050] Since high-level networks based on multi-task collaborative learning employ a multi-task co-training approach, each training sample must contain labels for all tasks. Specifically, each face training sample must be labeled with its gender, age, expression, pose, illumination, resolution, and identity. This invention first uses a primary network to run through the collected unlabeled training samples to generate initial multi-task label information; then, it uses manual correction to generate the final high-level network training samples.

[0051] 1.3 Construction of Expert Network Training Samples

[0052] The training samples for expert networks not only need to contain multi-task labels, but also need to provide their confidence levels. This invention first uses an advanced network to run through the collected training samples to generate initial multi-task labels and their corresponding confidence levels; then, it uses manual correction to generate the final expert network training samples.

[0053] 2. Multi-attribute facial analysis and recognition

[0054] This invention proposes an advanced multi-task learning network model and applies it to multi-task facial attribute recognition, such as... Figure 3As shown, it has the following characteristics: ① The input for each level is a face image, thus allowing the sharing of information from lower-level networks. ② The labels of the face images are different for each level (single-label, multi-label, and multi-label with confidence). ③ The output labels of each level are not independent; there are correlations and inclusion relationships between the output labels of different levels, thus facilitating knowledge transfer through the design of loss functions. ④ The advanced multi-task learning network does not need to focus on the "catastrophic forgetting" problem encountered in lifelong learning; it only needs to consider the knowledge transfer problem between different levels, thus reducing the complexity of the problem.

[0055] The advanced multi-task learning network model consists of three levels, with each level leveraging the experience of the previous level to improve itself. Level 1 utilizes lifelong learning to build a basic network, enabling the model to initially identify the performance of each task. Level 2 utilizes multi-task collaborative learning to build a high-level network, allowing the model to fully utilize the correlation and heterogeneity between different tasks to achieve simultaneous multi-task recognition. Level 3 utilizes soft-label learning to build an expert network, enabling the model to identify the essence of facial attribute recognition by filtering out false positives, thereby improving the model's robustness and generalization ability. The advanced multi-task learning network performs knowledge transfer in two ways. First, input transfer: to utilize the experience of previous levels, the current level network inputs the current task's input into all previous levels, and inputs the outputs of each layer of previous levels and the current level network's outputs together into the next layer. Second, output transfer: to utilize the knowledge of previous levels (considered as an input-to-output mapping), the loss function of each level network is designed as follows:

[0056]

[0057] Where i and j represent class, y i This is the actual result of the current class network, q j The output of the previous class network is given by , pred is given by the current class network, and loss represents the loss function, which can be the existing cross-entropy loss or other loss functions.

[0058] 2.1, Construction of a Primary Network Based on Lifelong Learning

[0059] Traditionally, multi-task deep learning involves training a separate deep neural network for each task to recognize different tasks. This approach has drawbacks: it isolates each task, preventing cross-task learning (co-learning); furthermore, it leads to computational and storage resource constraints. To address these issues, this invention introduces lifelong learning, utilizing the same network for continuous training across tasks to obtain a multi-task primary network. This network employs a lifelong learning strategy to learn each task, minimizing forgetting knowledge from previous tasks while learning new ones, and transferring knowledge from old tasks to new tasks to improve performance. It represents a forward knowledge transfer strategy—that is, how to leverage knowledge from old tasks to enhance learning of new tasks.

[0060] To facilitate knowledge transfer and avoid catastrophic forgetting, the primary network calculates the importance Ω of each parameter in the network to the task after training for that specific attribute task. ij This represents the proportion of the parameter value in the i-th row and j-th column to the total parameter values, and is used in subsequent training tasks. Ω ij It is added to the loss function as a regularization term whenever a new task is trained: for Ω ij For larger parameters, minimize their changes during gradient descent, as these parameters are important for past tasks and their values ​​need to be preserved to avoid catastrophic forgetting; while for Ω... ij Smaller parameters allow for larger gradient updates, resulting in better performance on new tasks. Therefore, the loss function for the nth task is:

[0061]

[0062] Among them, L n (θ) is the loss function for the current task, λ is the balance factor, and θ ij These are the parameters in the i-th row and j-th column of the current model parameter matrix. Ω consists of model parameters obtained after training the first n-1 tasks. After training on each task, Ω... ij They will all be updated, especially when performing the first task. Take 0.

[0063] The primary network takes single-attribute face training samples as input and outputs single-attribute labels. Its backbone network can be RestNet or EfficientNet. When using a lifelong learning network for multi-task learning of face attributes, the learning order of different attribute tasks has a significant impact on the final performance of the primary network. This invention employs a cross-validation strategy to determine the difficulty of each task and the learning order. The final determined recognition order is gender, illumination, resolution, pose, age, and expression.

[0064] 2.2, Construction of Advanced Networks Based on Multi-Task Collaborative Learning

[0065] Primary networks based on lifelong learning do not consider collaborative learning between tasks, i.e., they do not consider backward transfer of knowledge (using knowledge from new tasks to improve the learning performance of old tasks). Therefore, this invention utilizes multi-task collaborative learning to construct a high-level network, which, based on the transfer of knowledge from the primary network, uses multi-attribute samples for collaborative training and learning. The backbone of the high-level network is the same as the backbone of the primary network; its input is multi-attribute face samples, and its output is multi-attribute labels. Assuming N training samples are used to learn T tasks, ... Let represent the i-th sample vector in the t-th task ∈ {1,…,T} and its corresponding task label. Then the objective of multi-task collaborative learning is:

[0066]

[0067] Where, x i This represents the i-th training sample. Let f(x) represent the label of the i-th sample in the t-th attribute task. i Φ(θ) represents the deep network function, i.e., the backbone network in a high-level network, where θ is the network parameter to be solved, and Φ(θ) is the regularization penalty term, using the L2 penalty term. t (·) represents the loss function for different tasks. The choice of loss function can vary randomly depending on the situation, such as cross-entropy loss, center loss, triplet loss, arcface loss, and A-softmax loss. t represents six attributes: gender, pose, expression, illumination, resolution, and age. Since the difficulty of recognition tasks varies depending on the attribute, λ is used. t This represents the weight allocation for different loss functions.

[0068] 2.3 Research on Expert Networks Based on Soft Label Learning

[0069] In image classification tasks, image labels are often one-hot encoded (hard labels), meaning the label has only two possibilities: 0 or 1. In practice, due to various noises, interference, or special circumstances, we cannot assign a 100% probability to any image. For example, if a male has delicate features but the image is blurry, it's difficult to accurately determine his gender; in this case, we can only say it's 60% likely to resemble a male. However, conventional deep learning approaches simply assign a rigid category label regardless of image clarity. This approach has several drawbacks: ① It hinders the network model from finding essential classification features because it treats easily classified and difficult-to-classify samples the same, preventing it from exploring the reasons for sample classification difficulty. ② It hinders the design of the loss function. During training, the network should give more attention (higher weights) to difficult-to-classify samples, but hard labels cannot provide information about the difficulty of the samples.

[0070] To address the above issues, this invention proposes a soft-label training method with attached confidence scores to construct an expert network. This method utilizes soft-labeled samples for training and learning, based on the transfer of knowledge from the primary and advanced networks. First, training samples with confidence scores are collected; each sample not only records a classification label but also a label confidence value (given based on human subjective judgment). Then, the expert network is trained using these soft-labeled samples, based on the transfer of knowledge from the primary and advanced networks. The backbone of the expert network is identical to that of the primary network.

[0071] For classification problems, the cross-entropy loss function is commonly used:

[0072]

[0073] Where, p k For the network prediction results, y k The labels are either 1 or 0 (hard labels). In multi-class classification problems, labels are often converted into one-hot vectors, which leads to two problems: overfitting and the model becoming overly reliant on its own judgments, resulting in a very high probability that a sample belongs to a particular class. Therefore, some researchers have proposed Label Smoothing Regularization (LSR), which adds noise to the output y to constrain the model and reduce overfitting. LSR affects y... k Perform regularization:

[0074]

[0075] Where α is a very small constant smoothing factor, and K is the number of label categories. This invention also applies to hard labels y. kSoftening yields a new loss function, ArcFace loss:

[0076]

[0077] in, c k This represents the confidence level of the sample label. Unlike LSR, The value is determined based on the sample itself (by the confidence level given at the time of sample collection); it is a dynamic value, not a fixed one. Furthermore, to enhance p... k Inspired by Fisher's criterion (keeping intra-class members as close as possible and inter-class members as far apart as possible), the classification ability of the expert network is enhanced by using the ArcFace loss method, that is, using the ArcFace loss function instead of the Cross-Entropy loss function to achieve enhancement.

[0078] This invention also provides an image multi-task recognition system based on progressive learning, comprising the following modules:

[0079] The training sample construction module is used to construct training samples at multiple levels;

[0080] The model building module is used to build multi-level image recognition network models that enable progressive learning.

[0081] The advanced learning multi-level image recognition network model consists of three stages. The first stage utilizes lifelong learning to build a primary network, which is trained using single-attribute samples to enable the model to initially recognize the performance of various tasks. The second stage, based on the transfer of knowledge from the primary network, utilizes multi-attribute samples for collaborative training and learning, and uses multi-task collaborative learning to build an advanced network, enabling the model to fully utilize the correlation and heterogeneity between different tasks and complete the function of simultaneous recognition of multiple tasks. The third stage, based on the transfer of knowledge from the primary and advanced networks, utilizes soft-labeled samples for training and learning to obtain an expert network, and uses soft-label learning to build an expert network, enabling the model to have the essential image recognition ability to distinguish between true and false images, thereby improving the robustness and generalization ability of the model.

[0082] The model training module is used to train the image recognition network model using training samples of the corresponding level.

[0083] The recognition module is used to achieve multi-task image recognition using a trained multi-level image recognition network model.

[0084] The specific implementation methods of each module and the corresponding steps are not described in this invention.

[0085] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. An image multi-task recognition method based on progressive learning, characterized in that, Includes the following steps: Step 1: Construct multi-level training samples; Step 2: Construct a multi-level image recognition network model with progressive learning; The advanced learning multi-level image recognition network model consists of three stages. The first stage utilizes lifelong learning to build a primary network, which is trained using single-attribute samples to enable the model to initially recognize the performance of various tasks. The second stage, based on the transfer of knowledge from the primary network, utilizes multi-attribute samples for collaborative training and learning, and uses multi-task collaborative learning to build an advanced network, enabling the model to fully utilize the correlation and heterogeneity between different tasks and complete the function of simultaneous recognition of multiple tasks. The third stage, based on the transfer of knowledge from the primary and advanced networks, utilizes soft-labeled samples for training and learning to obtain an expert network, and uses soft-label learning to build an expert network, enabling the model to have the essential image recognition ability to distinguish between true and false images, thereby improving the robustness and generalization ability of the model. Advanced networks are constructed using multi-task collaborative learning. Based on the transfer of knowledge from primary networks, they are collaboratively trained using multi-attribute samples. The backbone of the advanced network is the same as that of the primary network. Suppose we use N training samples to learn T tasks, using Let represent the i-th sample vector in the t-th task ∈ {1,…,T} and its corresponding task label. Then the objective of multi-task collaborative learning is: Where, x i This represents the i-th training sample. Let f(x) represent the label of the i-th sample in the t-th attribute task. i Φ(θ) represents the deep network function, i.e., the backbone network of the high-level network, where θ is the network parameter to be solved, and Φ(θ) is the regularization penalty term, using the L2 penalty term. t (·) represents the loss function for different tasks, λ t This represents the weight allocation for different loss functions; The backbone of an expert network is the same as that of a primary network. In an expert network, hard labels y are used... k Softening yields a new loss function: in, This indicates that the label has been softened. The value is determined based on the specific circumstances of the sample, c k p represents the confidence level of the sample label. k For the network prediction results, y k The true result is either 1 or 0, where K is the number of label categories; Step 3: Train the image recognition network model using training samples of the corresponding level; Step 4: Use the trained multi-level image recognition network model to achieve multi-task image recognition.

2. The image multi-task recognition method based on progressive learning as described in claim 1, characterized in that: The multi-level training samples constructed in step 1 include single-label primary network training samples, multi-label advanced network training samples, and multi-label expert network training samples with confidence.

3. The image multi-task recognition method based on progressive learning as described in claim 1, characterized in that: The multi-level image recognition network model with progressive learning performs knowledge transfer in two ways: first, input transfer: to utilize the experience of previous level networks, the current level network inputs the current task's input into all previous level networks, and inputs the outputs of each layer of the previous level networks and the outputs of each layer of the current level network into the next layer; second, output transfer: to utilize the knowledge of previous level networks, the loss function of each level network is designed as follows: Where i and j represent class, y i This is the actual result of the current class network, q j is the output of the previous class network, pred is the output of the current class network, and loss represents the loss function.

4. The image multi-task recognition method based on progressive learning as described in claim 1, characterized in that: To facilitate knowledge transfer and avoid catastrophic forgetting, the primary network, after training for each attribute task, calculates the importance Ω of each parameter in the network to that task. ij That is, the proportion of the parameter value in the i-th row and j-th column to the total parameter value, which is used in subsequent training tasks; Ω ij It is added to the loss function as a regularization term. The loss function for the nth task is: Among them, L n (θ) is the loss function for the current task, λ is the balance factor, and θ ij These are the parameters in the i-th row and j-th column of the current model parameter matrix. Ω consists of model parameters obtained after training the first n-1 tasks. After training on each task, Ω... ij They will all be updated, especially when performing the first task. Take 0.

5. The image multi-task recognition method based on progressive learning as described in claim 1, characterized in that: The primary network takes single-attribute training samples as input and outputs single-attribute labels. Its backbone network uses RestNet or EfficientNet.

6. The image multi-task recognition method based on progressive learning as described in claim 4, characterized in that: In the process of multi-attribute facial image recognition, a cross-validation strategy is adopted to find the difficulty of each task and the learning ranking problem. The final recognition order is determined as gender, illumination, resolution, pose, age, and expression.

7. The image multi-task recognition method based on progressive learning as described in claim 1, characterized in that: A pre-trained multi-level image recognition network model is used to identify the gender, age, expression, pose, lighting, and resolution attributes of facial images.

8. An image multi-task recognition system based on progressive learning, characterized in that, Includes the following modules: The training sample construction module is used to construct training samples at multiple levels; The model building module is used to build multi-level image recognition network models that enable progressive learning. The advanced learning multi-level image recognition network model consists of three stages. The first stage utilizes lifelong learning to build a primary network, which is trained using single-attribute samples to enable the model to initially recognize the performance of various tasks. The second stage, based on the transfer of knowledge from the primary network, utilizes multi-attribute samples for collaborative training and learning, and uses multi-task collaborative learning to build an advanced network, enabling the model to fully utilize the correlation and heterogeneity between different tasks and complete the function of simultaneous recognition of multiple tasks. The third stage, based on the transfer of knowledge from the primary and advanced networks, utilizes soft-labeled samples for training and learning to obtain an expert network, and uses soft-label learning to build an expert network, enabling the model to have the essential image recognition ability to distinguish between true and false images, thereby improving the robustness and generalization ability of the model. Advanced networks are constructed using multi-task collaborative learning. Based on the transfer of knowledge from primary networks, they are collaboratively trained using multi-attribute samples. The backbone of the advanced network is the same as that of the primary network. Suppose we use N training samples to learn T tasks, using Let represent the i-th sample vector in the t-th task ∈ {1,…,T} and its corresponding task label. Then the objective of multi-task collaborative learning is: Where, x i This represents the i-th training sample. Let f(x) represent the label of the i-th sample in the t-th attribute task. i Φ(θ) represents the deep network function, i.e., the backbone network of the high-level network, where θ is the network parameter to be solved, and Φ(θ) is the regularization penalty term, using the L2 penalty term. t (·) represents the loss function for different tasks, λ t This represents the weight allocation for different loss functions; The backbone of an expert network is the same as that of a primary network. In an expert network, hard labels y are used... k Softening yields a new loss function: in, This indicates that the label has been softened. The value is determined based on the specific circumstances of the sample, c k p represents the confidence level of the sample label. k For the network prediction results, y k The true result is either 1 or 0, where K is the number of label categories; The model training module is used to train the image recognition network model using training samples of the corresponding level. The recognition module is used to achieve multi-task image recognition using a trained multi-level image recognition network model.

Images