Lightweight image recognition system

Abstract

The invention provides a lightweight image recognition system, and relates to the field of image processing.The lightweight image recognition system comprises a preparation module, a data preprocessing module, a model establishment module, a model training module, a model evaluation module and a model deployment module.According to the lightweight image recognition system, regularization parameters are set, and the image recognition accuracy is improved; comprising an L2 regularization parameter, group sparse regularization parameters lambda1, lambda2 and lambda3 and a loss rate, the overfitting phenomenon of the BMtLS-RG model in the training process is effectively inhibited, an objective function is optimized in combination with an alternating direction multiplier method, the generalization performance and stability of the BMtLS-RG model are further improved, and the problem that the BMtLS-RG model has high robustness is solved. The method aims at solving the overfitting problem of an existing BMtLS-RG model when a data set is small or correlation between tasks is not strong.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular, to a lightweight image recognition system. Background Art

[0002] With the continuous development of image recognition technology, its applications in various fields are becoming increasingly widespread, such as security monitoring, autonomous driving, medical image analysis, etc.

[0003] Existing journals have disclosed a generalized multi-task learning system with group sparse regularization. Journal authors: Huang Jintao, Chen Chuangquan. It is recorded that the Broad Learning System (BLS) has achieved success in applications due to its lightweight, incremental expansion, and strong generalization ability. Despite these advantages, BLS still faces challenges in multi-task learning (MTL) scenarios. Its limited capabilities cannot solve multiple complex tasks simultaneously. Existing BLS models cannot fully capture and utilize the key information across tasks, thus reducing their effectiveness and efficiency in MTL scenarios. To address these issues, an innovative MTL framework specifically designed for BLS, called the Group Sparse Regularized Width Multi-Task Learning System Based on Related Tasks (BMtLS-RG), was proposed. This framework combines a task-related BLS learning mechanism and a group sparse optimization strategy, significantly enhancing the generalization ability of BLS in the MTL environment. The task-related learning component utilizes task relevance to achieve shared learning and efficiently optimize parameters. At the same time, the group sparse optimization method helps to minimize the impact of irrelevant or noisy data, thereby enhancing the robustness and stability of BLS in complex learning scenarios. To meet the diverse requirements of MTL challenges, two additional variants of BMtLS-RG were proposed: BMtLS-RG with shared feature mapping node parameters (BMtLS-RGf), which integrates a shared feature mapping layer; and BMtLS-RGf with enhanced nodes (BMtLSRGfe), which further adds an enhanced node layer on top of the shared feature mapping structure. These adaptations provide customized solutions for the diversity of MTL problems. Through comprehensive experimental evaluations on multiple real MTL and UCI datasets, BMtLS-RG was compared with state-of-the-art (SOTA) MTL and BLS algorithms. BMtLS-RG achieved the best performance in 97.81% of the classification tasks.

[0004] It has been found that the above BMtLS-RG model still has deficiencies in use. Although group sparse regularization helps to reduce overfitting, in some cases, such as when the dataset is small or the correlation between tasks is weak, the BMtLS-RG model may still face the risk of overfitting. Specifically, the BMtLS-RG model performs well on the training set but its performance drops on the test set.

[0005] Therefore, it is necessary to provide a new lightweight image recognition system to solve the above technical problems. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a lightweight image recognition system.

[0007] The lightweight image recognition system provided by the present invention includes: a preparation module, configured to select original image data from a database, a network, or a local storage device, and extract an image recognition task, where the image recognition task includes face recognition, object recognition, and scene recognition;

[0008] a data preprocessing module, which preprocesses the original image and extracts image features using a pre-trained convolutional neural network to establish an image dataset;

[0009] a model establishment module, which establishes a BMtLS-RG model according to the preprocessed image dataset and groups it according to the type of image features;

[0010] a model training module, after grouping, divides the preprocessed image dataset into a training set, a validation set, and a test set, uses the divided training set to train the BMtLS-RG model, incorporates the set regularization parameters and dropout rate into the BMtLS-RG model for updating, combines L2 regularization and group sparse regularization, optimizes the objective function by the alternating direction method of multipliers, and monitors the loss function value and performance metrics on the validation set during the training process, and adjusts the model parameters and hyperparameters according to the training effect; where the dropout rate is also denoted as the Dropout rate;

[0011] a model evaluation module, which evaluates the trained model using the test set and calculates the performance metrics of the model on the test set;

[0012] a model deployment module, which packages the trained BMtLS-RG model into an executable program or an API interface.

[0013] Preferably, the model establishment module further includes the following operating steps:

[0014] Step 1: First, initialize the basic structure of the BMtLS-RG model, including an input layer, a feature mapping layer, an enhanced node layer, and an output layer;

[0015] Step 2: After initialization, set the number of feature mapping nodes and the number of enhanced nodes, and initialize the regularization parameters λ1, λ2, λ3, and the dropout rate;

[0016] Step 3: Group according to the type of image features or their positions in the image to form multiple feature groups;

[0017] Step 4: After grouping is completed, initialize the basic objective function, including the loss function and the L2 regularization term. According to the result of feature grouping, calculate the L2 norm of each feature group, sum up the L2 norms of all feature groups, and multiply by the group sparse regularization parameter λ2 to form the group sparse regularization term. At this time, add the group sparse regularization term to the basic objective function.

[0018] Preferably, in the data preprocessing module, the preprocessing includes image enhancement, normalization, and image cropping. Among them, the input image size for image cropping is 112x112 pixels.

[0019] Preferably, in the preparation module, after selecting the original image data, extract the image recognition tasks, where the image recognition tasks include face recognition, object recognition, and scene recognition.

[0020] Preferably, in Step 2, initialize the regularization parameters λ1, λ2, λ3 and the dropout rate. Among them, λ1 is the L1 norm regularization parameter, which is used to perform sparse constraints on the weights of the BMtLS-RG model;

[0021] λ2 is the task sparse regularization parameter, which is used to control the sparsity of the weights between different image recognition tasks;

[0022] λ3 is the group sparse regularization parameter, which is used to perform sparse constraints on the feature groups.

[0023] Preferably, in the model training module, the L2 regularization parameter is used to control the intensity of L2 norm regularization. By imposing a penalty on the sum of squares of the weights of the BMtLS-RG model, the weights of the BMtLS-RG model are made as small as possible to avoid the model being too complex and overfitting.

[0024] Preferably, the input layer is: the preprocessed image data;

[0025] The feature mapping layer is: multi-type features extracted from the preprocessed image data, including edge features, texture features, and color features;

[0026] The enhanced node layer is: designing specific enhanced nodes for each image recognition task through non-linear transformation, further processing the output of the feature mapping layer, including edge feature enhancement, texture feature enhancement, and color feature enhancement, and extracting task-specific information; among them, the edge feature is the contour feature of the object in the image, the texture feature is the directionality and roughness of the object in the image, and the color feature is the statistical feature of the color distribution of the object in the image;

[0027] The output layer is: defining output nodes for each task and outputting the prediction results corresponding to the tasks.

[0028] Preferably, the edge feature is the contour feature of the object in the image, the texture feature is the directionality and roughness of the object in the texture feature image, and the color feature is the statistical feature of the color distribution of the object in the image.

[0029] Preferably, the model evaluation module further includes the following operating steps:

[0030] S1. Model evaluation: Use the test set to evaluate the trained BMtLS-RG model, and calculate the accuracy rate, recall rate, and F1 score of the BMtLS-RG model on the test set;

[0031] S2. Data analysis: According to the evaluation results, analyze the performance of the BMtLS-RG model on different tasks;

[0032] S3. Define task relevance: According to the image recognition task, define the relevance between image recognition tasks, introduce a weight fusion criterion, realize shared learning between image recognition tasks, and improve the generalization ability of the BMtLS-RG model.

[0033] Compared with the related technology, the lightweight image recognition system provided by the present invention has the following beneficial effects:

[0034] 1. By setting regularization parameters, including L2 regularization parameters and group sparse regularization parameters λ1, λ2, λ3 and dropout rate, the present invention effectively suppresses the overfitting phenomenon of the BMtLS-RG model during the training process. Combining the alternating direction method of multipliers to optimize the objective function further improves the generalization performance and stability of the BMtLS-RG model, and solves the overfitting problem faced by the existing BMtLS-RG model when the data set is small or the correlation between tasks is not strong;

[0035] 2. The BMtLS-RG in the present invention combines a task-related learning mechanism and a group sparse optimization strategy, can make full use of the relevance between tasks to achieve shared learning, and significantly improves the generalization ability of the BMtLS-RG model in multi-task scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a schematic flow chart of the lightweight image recognition system provided by the present invention;

[0037] Figure 2 is a schematic flow chart of model establishment. DETAILED DESCRIPTION OF THE INVENTION

[0038] The present invention will be further described below with reference to the drawings and embodiments.

[0039] Please refer to Figures 1 to 2 , wherein, Figure 1 is a schematic flow chart of the lightweight image recognition system provided by the present invention;Figure 2 Flow chart for model establishment

[0040] In the specific implementation process, as Figures 1 to 2 shown, it includes a preparation module for selecting original image data from a database, network, or local storage device;

[0041] A data preprocessing module preprocesses the original images and extracts image features using a pre-trained convolutional neural network to establish an image dataset; among them, the preprocessing includes image enhancement, normalization, and image cropping, and the input image size for image cropping is 112x112 pixels;

[0042] Preprocessing steps:

[0043] Image enhancement: Improve image quality and feature saliency.

[0044] Normalization: Convert image data to a unified scale for subsequent processing.

[0045] Image cropping: Crop the input image size to 112x112 pixels to ensure the consistency of the input for the subsequent BMtLS-RG model.

[0046] Features: Through preprocessing, the quality and applicability of image data are improved, providing a high-quality image dataset for the subsequent training of the BMtLS-RG model;

[0047] A model establishment module establishes a BMtLS-RG model based on the preprocessed image dataset and groups them according to the type of image features;

[0048] A model training module, after grouping, divides the preprocessed image dataset into a training set, a validation set, and a test set, uses the divided training set to train the BMtLS-RG model, incorporates the set regularization parameters and dropout rate into the BMtLS-RG model for updating, combines L2 regularization and group sparse regularization, optimizes the objective function through the alternating direction method of multipliers (ADMM), and monitors the loss function value during training and the performance metrics on the validation set, and adjusts the model parameters and hyperparameters according to the training effect;

[0049] A model evaluation module evaluates the trained model using the test set and calculates the performance metrics of the model on the test set;

[0050] A model deployment module packages the trained BMtLS-RG model into an executable program or API interface, facilitating the deployment and use of the BMtLS-RG model in practical applications, and improving the practicality and convenience of the system.

[0051] The model evaluation module further includes the following operating steps:

[0052] S1, Model Evaluation: Use the test set to evaluate the trained BMtLS-RG model, and calculate the accuracy, recall rate, and F1 score of the BMtLS-RG model on the test set;

[0053] S2, Data Analysis: Analyze the performance of the BMtLS-RG model on different tasks according to the evaluation results;

[0054] S3, Define Task Relevance: According to the image recognition task, define the relevance between image recognition tasks, introduce the weight fusion criterion, achieve shared learning between image recognition tasks, and improve the generalization ability of the BMtLS-RG model.

[0055] Among them, for example: The objective function is:

[0056]

[0057] Among them, L(W) is the loss function, W is the weight parameter of the model, λ1, λ2, λ3 are regularization parameters, is the L2 regularization term, ∑ g ||W g ||2 is the group sparse regularization term, and Wg represents the g-th group of weight parameters.

[0058] First, set the initial weight parameter W 0 , the auxiliary variable Z 0 (with the same dimension as W), and the Lagrange multiplier U 0 (with the same dimension as W and initialized to zero).

[0059] Set the regularization parameters λ1, λ2, λ3 and the penalty parameter ρ of ADMM

[0060] Perform iterative updates:

[0061] Step 1: Update W:

[0062]

[0063] Step 2: Update Z:

[0064] Specifically, for each group Zg, perform the following update:

[0065]

[0066] Among them, shrinkage is the soft threshold function, which is defined for the vector x and the threshold τ as:

[0067] shrinkage(x,τ) = sign(x) max(|x| - τ, 0);

[0068] For group soft thresholding, the same threshold is applied to the entire group;

[0069] Step 3: Update U: U K+1 = U K + W K+1 — Z K+1 ;

[0070] This step is the update of the Lagrange multiplier, which is used to ensure that the constraint W = Z is gradually satisfied during the iteration;

[0071] Check whether the difference between W K+1 and Z K+1 is less than a preset tolerance, or check whether the change in the objective function value is less than a certain threshold.

[0072] If the convergence condition is met, stop the iteration; otherwise, return to Step 2 and continue the iteration

[0073] After convergence, the final W or Z can be used as the trained model parameters. Since the values of W and Z are close, the BMtLS-RG model can be further evaluated, tuned, and deployed.

[0074] Specifically, ADMM decomposes the original problem into multiple simpler sub-problems, enabling each sub-problem to be solved independently, thereby reducing the optimization difficulty. By introducing group sparse regularization, ADMM helps to learn more sparse and robust model parameters, thus improving the generalization ability of the BMtLS-RG model.

[0075] Refer to Figures 1 to 2 As shown, the model building module further includes the following operating steps:

[0076] Step 1: First, initialize the basic structure of the BMtLS-RG model, including the input layer, feature mapping layer, enhanced node layer, and output layer;

[0077] The input layer is: the preprocessed image data;

[0078] The feature mapping layer is: multiple types of features extracted from the preprocessed image data, including edge features, texture features, and color features;

[0079] The enhanced node layer is: designing specific enhanced nodes for each image recognition task through non-linear transformation, further processing the output of the feature mapping layer, including edge feature enhancement, texture feature enhancement, and color feature enhancement, and extracting task-specific information;

[0080] The output layer is: Define output nodes for each task and output the prediction results of the corresponding tasks.

[0081] Step 2: After initialization, set the number of feature mapping nodes and the number of enhancement nodes, and initialize the regularization parameters λ1, λ2, λ3 and the dropout rate;

[0082] Among them, λ1 is the L1-norm regularization parameter, which is used for sparse constraint on the weights of the BMtLS-RG model;

[0083] λ2 is the task sparse regularization parameter, which is used to control the sparsity of the weights between different image recognition tasks;

[0084] λ3 is the group sparse regularization parameter, which is used for sparse constraint on the feature groups;

[0085] Step 3: Group according to the type of image features or their positions in the image to form multiple feature groups;

[0086] Step 4: After grouping, initialize the basic objective function, including the loss function and the L2 regularization term. According to the result of feature grouping, calculate the L2 norm of each feature group, sum the L2 norms of all feature groups, and multiply by the group sparse regularization parameter λ2 to form the group sparse regularization term. At this time, add the group sparse regularization term to the basic objective function;

[0087] In the model training module, the L2 regularization parameter is used to control the intensity of L2-norm regularization. By imposing a penalty on the sum of squares of the weights of the BMtLS-RG model, the weights of the BMtLS-RG model are made as small as possible to avoid the BMtLS-RG model from being too complex and overfitting.

[0088] Example 1, as shown in Figures 1 to 2 For enhancing the generalization ability of the BMtLS-RG model and reducing overfitting, there is also the following analysis:

[0089] First, introduce the L2 regularization term into the original loss function of the BMtLS-RG model. Suppose the original loss function is:

[0090]

[0091] Among them, W is the weight parameter of the model, Y t represents the target output or label at time t, W G represents a specific grouping of the model weight parameters, Wt represents the weight parameters at time t or related to a specific task / layer / module, and λ1, λ2, λ3 are the group sparse regularization parameters;

[0092] After introducing the L2 regularization, the loss function becomes:

[0093]

[0094] Among them, λ4 is a parameter that controls the strength of L2 regularization;

[0095] At this time, initial values are set for all regularization parameters (λ1, λ2, λ3, λ4). For example: λ1 = λ2 = λ3 = 0.01, λ4 = 0.001;

[0096] At this time, cross-validation is performed: divide the dataset: divide the image dataset into a training set, a validation set, and a test set. Training set: 70%; Validation set: 15%; Test set: 15%.

[0097] Set the search grid for the regularization parameters. For example:

[0098] λ1, λ2, λ3 ∈ {0.001, 0.01, 0.1, 0.5, 1};

[0099] λ4 ∈ {0.0001, 0.001, 0.01, 0.1};

[0100] For each set of parameter values in the grid, use the training set to train the BMtLS-RG model;

[0101] Evaluate the performance of the model on the validation set, and record metrics such as the validation set loss or accuracy corresponding to each set of parameters;

[0102] Select the best parameters: Select the set of regularization parameter values that optimizes the performance of the validation set. For example:

[0103] λ1 = 0.01; λ2 = 0.05; λ3 = 0.01; λ4 = 0.001

[0104] Initial parameter setting: Use the initial values selected by the above cross-validation:

[0105] λ1 = 0.01; λ2 = 0.05; λ3 = 0.01

[0106] Starting from the initial values, gradually increase the value of the group sparse regularization parameter by 0.01 each time:

[0107] 1. Increase λ2, λ2 = 0.06, retrain the model and evaluate the performance on the validation set;

[0108] 2. Increase λ3, λ3 = 0.02, retrain the model and evaluate the performance on the validation set;

[0109] At this time, according to the performance of the validation set, select the set of group sparse regularization parameter values that optimizes the generalization ability of the model. For example: λ1 = 0.01, λ2 = 0.06, λ3 = 0.02.

[0110] Apply dropout rate:

[0111] During the training process, randomly discard the outputs of a part of the feature mapping nodes or enhancement nodes, and set the dropout rate to 0.5.

[0112] Cross-validate the dropout rate: First, set the search range for the dropout rate: 0.1, 0.2, 0.3, 0.4, 0.5.

[0113] For each dropout rate value, combine it with the previously selected regularization parameter values (λ1 = 0.01, λ2 = 0.06, λ3 = 0.02, λ4 = 0.001), retrain the model and evaluate the performance on the validation set.

[0114] At this time, select the dropout rate that optimizes the performance of the validation set, that is, the dropout rate: For example, select the dropout rate to be 0.3, and combine L2 regularization, group sparse regularization and Dropout, and the regularization parameter combination is:

[0115] λ1 = 0.01; λ2 = 0.06; λ3 = 0.02; λ4 = 0.001; dropout rate = 0.3

[0116] Use the regularization parameter combination that performs best on the validation set to retrain the model on the complete training set, including the previous validation set;

[0117] Evaluate the performance of the final model on the test set, and the training results are:

[0118] Accuracy: 93%;

[0119] Recall rate: 92%;

[0120] F1 score: 92.5%.

[0121] Through the above embodiments, the generalization ability of the BMtLS-RG model is successfully enhanced and overfitting is reduced, specifically reflected in the significant improvement of the performance metrics on the test set;

[0122] Specifically, by introducing L2 regularization, that is, weight decay and group sparse regularization, the complexity of the weights of the BMtLS-RG model is effectively restricted, preventing the BMtLS-RG model from overfitting on the training data;

[0123] Specifically, the application of the Dropout technique further enhances the robustness of the BMtLS-RG model. By randomly discarding the outputs of a part of the nodes, the BMtLS-RG model learns to rely more on multiple independent features rather than on specific noises or patterns in the training data;

[0124] Furthermore, the optimized BMtLS-RG model can capture the contour features of objects in images more accurately. Due to the introduction of the regularization method, the BMtLS-RG model pays more attention to the overall structure and edge information of the image during the training process, thereby improving the recognition accuracy of object contours. That is, the BMtLS-RG model also shows a stronger ability to extract texture features in the image. By tuning the regularization parameters, the BMtLS-RG model can learn a more robust texture feature representation, enabling it to maintain a high recognition accuracy when facing images with different texture complexities.

[0125] Furthermore, the optimized BMtLS-RG model performs better in identifying the directionality of objects in images. By introducing L2 regularization and group sparse regularization, the BMtLS-RG model can learn a more directional feature representation, thereby improving the recognition ability of object directionality. And the BMtLS-RG model also shows a better recognition effect on the roughness of objects in the image. The introduction of the regularization method makes the BMtLS-RG model pay more attention to the detailed information in the image during the training process, thus improving the perception ability of object roughness.

[0126] Even further, the optimized BMtLS-RG model can extract the statistical features of the color distribution of objects in images more accurately. By tuning the regularization parameters and applying the Dropout technique, the BMtLS-RG model can learn a more robust color feature representation during the training process, enabling it to maintain a high recognition accuracy when facing images with different color distributions and lighting conditions.

[0127] The circuits and controls involved in the present invention are all prior arts and will not be elaborated herein.

[0128] The above are only embodiments of the present invention and do not limit the patent scope of the present invention accordingly. Any equivalent structural or equivalent process transformation made using the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A lightweight image recognition system, characterized in that, It includes a preparation module for selecting original image data from a database, network, or local storage device; a data preprocessing module for preprocessing the original image and using a pre-trained convolutional neural network to extract image features to establish an image dataset; a model establishment module for establishing a BMtLS-RG model based on the preprocessed image dataset and grouping according to the type of image features; a model training module. After grouping, the preprocessed image dataset is divided into a training set, a validation set, and a test set. The BMtLS-RG model is trained using the divided training set. The set regularization parameters and dropout rate are incorporated into the BMtLS-RG model for updating. Combining L2 regularization and group sparsity regularization, the objective function is optimized by the alternating direction method of multipliers, and the loss function value and performance metrics on the validation set during the training process are monitored. The model parameters and hyperparameters are adjusted according to the training effect; a model evaluation module for evaluating the trained model using the test set and calculating the performance metrics of the model on the test set; a model deployment module for encapsulating the trained BMtLS-RG model into an executable program or API interface.

2. The lightweight image recognition system according to claim 1, wherein The model establishment module further includes the following operation steps: Step 1: First, initialize the basic structure of the BMtLS-RG model, including an input layer, a feature mapping layer, an enhancement node layer, and an output layer; Step 2: After initialization, set the number of feature mapping nodes and the number of enhancement nodes, and initialize the regularization parameters λ1, λ2, λ3, and the dropout rate; Step 3: Group according to the type of image features or their positions in the image to form multiple feature groups.

3. The lightweight image recognition system according to claim 2, wherein In the data preprocessing module, the preprocessing includes image enhancement, normalization, and image cropping.

4. The lightweight image recognition system according to claim 3, wherein In the preparation module, after selecting the original image data, the image recognition task is extracted.

5. The lightweight image recognition system according to claim 4, wherein In Step 2, the regularization parameters λ1, λ2, λ3, and the dropout rate are initialized, where λ1 is the L1 norm regularization parameter for sparse constraint of the BMtLS-RG model weights; λ2 is the task sparsity regularization parameter for controlling the sparsity of weights between different image recognition tasks; λ3 is the group sparsity regularization parameter for sparse constraint of the feature groups.

6. The lightweight image recognition system according to claim 5, wherein In the model training module, the L2 regularization parameter is used to control the intensity of L2 norm regularization. By imposing a penalty on the sum of squares of the BMtLS-RG model weights, the BMtLS-RG model weights are made as small as possible to avoid the BMtLS-RG model from being too complex and overfitting.

7. The lightweight image recognition system according to claim 6, wherein The input layer is: the preprocessed image data; The feature mapping layer is: multi-type features extracted from the preprocessed image data, including edge features, texture features, and color features; The enhancement node layer is: designing specific enhancement nodes for each image recognition task through nonlinear transformation to further process the output of the feature mapping layer, including edge feature enhancement, texture feature enhancement, and color feature enhancement, and extracting task-specific information; The output layer is: defining output nodes for each task and outputting the prediction results of the corresponding tasks.

8. The lightweight image recognition system according to claim 7, characterized in that The edge feature is the contour feature of the object in the image, the texture feature is the directionality and roughness of the object in the texture feature image, and the color feature is the statistical feature of the color distribution of the object in the image.

9. The lightweight image recognition system according to claim 8, wherein, The model evaluation module further includes the following operation steps: S1. Model evaluation: Use the test set to evaluate the trained BMtLS-RG model, and calculate the accuracy rate, recall rate, and F1 score of the BMtLS-RG model on the test set; S2. Data analysis: Analyze the performance of the BMtLS-RG model on different tasks according to the evaluation results; S3. Define task relevance: According to the image recognition task, define the relevance between image recognition tasks, introduce the weight fusion criterion, realize shared learning between image recognition tasks, and improve the generalization ability of the BMtLS-RG model.

10. The lightweight image recognition system according to claim 9, characterized in that, In the preprocessing, the input image size for image cropping is 112x112 pixels.

Images