An Incremental Target Detection Method Based on Gradient Decomposition and Alignment

Abstract

This invention belongs to the field of computer vision technology, specifically disclosing an incremental object detection method based on gradient decomposition and alignment, including the following steps: constructing an incremental object detection model based on gradient decomposition and alignment; and using the incremental object detection model based on gradient decomposition and alignment to complete incremental object detection. This invention effectively solves the problem of unlabeled old category samples in incremental object detection through a pseudo-label generation and filtering mechanism, and proposes a gradient decomposition and alignment strategy to achieve dynamic adjustment of the gradients of the old and new categories during incremental training. This ensures that the learning of the new category does not impair the preservation of knowledge of the old category, improving the balance between the stability and plasticity of the incremental object detection model, thereby solving the problems of catastrophic forgetting, category overlap, and the difficulty in optimizing the balance between learning the old and new categories in incremental object detection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, specifically relating to an incremental target detection method based on gradient decomposition and alignment. Background Technology

[0002] Incremental object detection is an important research direction in computer vision, aiming to enable object detection models to maintain their ability to detect older categories while continuously learning new ones. Incremental learning is particularly important in dynamically changing application scenarios, such as intelligent surveillance and autonomous driving, where systems need to adapt to new object categories in real time without losing existing knowledge.

[0003] Traditional object detection methods are typically trained on static datasets, with all category labels provided at the start of training. However, in real-world scenarios, the number of object categories often increases gradually, leading to a catastrophic forgetting problem where the model loses its ability to detect older categories when learning new ones. Furthermore, incremental learning suffers from category overlap; in new training data, targets of older categories are often unlabeled and easily misclassified as background, further exacerbating the knowledge forgetting phenomenon.

[0004] Most existing incremental object detection methods rely on knowledge distillation or data replay, but these methods perform poorly in scenarios with overlapping categories or are limited by issues such as data privacy and storage resources. Furthermore, these methods fail to effectively balance the model's stability (preserving old knowledge) and plasticity (learning new knowledge), making it difficult to simultaneously optimize the ability to learn new categories and detect old categories.

[0005] To address the aforementioned issues, there is an urgent need for a new incremental object detection method that can efficiently learn new categories in dynamic environments while mitigating catastrophic forgetting and improving the system's detection performance and adaptability. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of catastrophic forgetting, class overlap, and difficulty in optimizing the learning balance between old and new classes in incremental target detection when performing target monitoring in dynamic environments. A gradient decomposition and alignment-based incremental target detection method is proposed.

[0007] The technical solution of this invention is: an incremental target detection method based on gradient decomposition and alignment, comprising the following steps:

[0008] S1. Construct an incremental target detection model based on gradient decomposition and alignment;

[0009] S2. Input the image to be detected into the incremental target detection model based on gradient decomposition and alignment to complete the incremental target detection.

[0010] Preferably, step S1 specifically includes the following steps:

[0011] Construct an incremental target detection model and perform initial training on the incremental target detection model;

[0012] Incremental learning is performed on the incremental target detection model using a pseudo-label generation and filtering module based on Gaussian mixture model and a gradient decomposition and alignment module, resulting in an incremental target detection model based on gradient decomposition and alignment.

[0013] The pseudo-label generation and filtering module based on Gaussian mixture model is used to receive the prediction results of the incremental target detection model after initial training, and generate and filter pseudo-labels based on the prediction results of the incremental target detection model after initial training.

[0014] The gradient decomposition and alignment module is used to receive the pseudo-labels, dynamically decompose the new label gradient and pseudo-label gradient of the incremental target detection model, perform gradient alignment when gradient directions conflict, and then update the parameters of the incremental target detection model according to the updated gradients to obtain an incremental target detection model based on gradient decomposition and alignment.

[0015] Preferably, the incremental target detection model includes a feature extraction unit, a candidate region extraction network, and a ROIHead unit;

[0016] The feature extraction unit is used to extract visual feature information from the input image and output the original image visual feature map.

[0017] The candidate region extraction network is used to receive the visual feature map of the original image and generate candidate regions;

[0018] The ROI Head unit is used to extract feature vectors from the candidate regions and to classify and fine-tune the bounding boxes of the feature vectors.

[0019] Preferably, the feature extraction unit includes multiple serial convolutional residual blocks;

[0020] The formula for extracting the visual feature map of the original image is:

[0021]

[0022] in, Represents the visual feature map of the original image, with , Represents the set of real numbers. , and These represent the height, width, and number of channels of the original image's visual feature map, respectively. Represents a feature extraction unit. This represents the original input image.

[0023] Preferably, the candidate region extraction network receives the visual feature map of the original image and generates candidate regions, specifically including the following steps:

[0024] The original image visual feature map is received, and an intermediate feature map is obtained by sliding a convolution kernel across the original image visual feature map. The formula for the intermediate feature map is as follows:

[0025]

[0026] in, This represents the intermediate feature map. Represents the convolution kernel. Represents the visual feature map of the original image;

[0027] Multiple predefined anchor boxes are generated at each location in the intermediate feature map, each anchor box... The coordinates are ,in, Indicates anchor frame center point coordinate, Indicates anchor frame center point coordinate, Indicates anchor frame width, Indicates anchor frame Height;

[0028] The candidate region extraction network uses two parallel classification and regression branches to perform binary classification and bounding box regression on the intermediate feature map to obtain the anchor box. The offset, the formula for the binary classification of the classification branch is:

[0029]

[0030] in, Indicates the first The classification probability of each anchor box. This represents a fully connected layer network used for classification tasks. This represents the normalized exponential function;

[0031] The formula for the bounding box regression of the regression branch is as follows:

[0032]

[0033] in, Indicates anchor frame The offset, Includes 4 offsets, namely: anchor frame center point coordinate offset Anchor frame center point coordinate offset Anchor frame Width offset and anchor frame height offset , This represents a fully connected layer network used for regression tasks;

[0034] According to the anchor frame coordinates and anchor frame The offset is used to calculate the candidate region, and the calculation formula is as follows:

[0035]

[0036] in, Indicates the center point of the candidate region coordinate, Indicates the center point of the candidate region coordinate, Indicates the width of the candidate region. Indicates the height of the candidate region. Represents the natural base.

[0037] Preferably, the ROI Head unit extracts feature vectors from the candidate regions and performs classification and bounding box fine-tuning on the feature vectors, specifically including the following steps:

[0038] The ROI pooling layer in the ROI Head unit extracts a fixed-size feature vector for each candidate region from the visual feature map of the original image. The formula for extracting the feature vector is as follows:

[0039]

[0040] in, Indicates the first The feature vectors corresponding to each candidate region Represents the visual feature map of the original image. Indicates the first Candidate regions, Indicates the ROI pooling unit;

[0041] The feature vector is input into a fully connected layer for classification and bounding box fine-tuning, and the formula is as follows:

[0042]

[0043] in, Indicates the first The classification probability of each candidate region Represents the normalized exponential function, This represents a fully connected layer network used for classification tasks. This represents the finely adjusted bounding box. This represents a fully connected layer network used for regression tasks.

[0044] As a preferred option, the incremental target detection model is initially trained using candidate region extraction network loss and ROI-Head loss;

[0045] The candidate region extraction network loss The formula for expressing this is:

[0046]

[0047] in, This indicates the category prediction for the candidate region. Indicates the true label of the candidate region. This indicates the predicted bounding box of the candidate region. This represents the true bounding box of the candidate region. The normalization factor represents the normalization factor for the classification task. Represents cross-entropy loss, Indicates the first Category prediction for each candidate region Indicates the first The real labels of each candidate region have 1 represents the target, and 0 represents the background. This represents the normalization factor for the regression task. Indicates smoothness loss, Indicates the first Boundary box prediction for each candidate region Indicates the first The true bounding boxes of each candidate region;

[0048] The ROI-Head loss The formula for expressing this is:

[0049]

[0050] in, This indicates the predicted category of the feature vector. This represents the true class label of the feature vector. Indicates the first The predicted category of each feature vector. Indicates the first The true class labels of each feature vector are: , Indicates the background category.

[0051] Preferably, the pseudo-label generation and filtering module based on the Gaussian mixture model generates and filters pseudo-labels based on the prediction results of the incremental target detection model after initial training, specifically including the following steps:

[0052] Using an incremental target detection model, inference is performed on the data of the current stage to generate a set of prediction results, which is:

[0053]

[0054] in, This represents the set of prediction results from the incremental object detection model. The first element of the incremental target detection model is represented by the second element. The bounding box positions of the candidate regions Indicates the first The predicted category of each feature vector. The first element of the incremental target detection model is represented by the second element. The confidence scores of each prediction result are: ;

[0055] A Gaussian mixture model is used to model the set of confidence scores in the prediction result set, resulting in a Gaussian mixture model of the confidence score set, which is expressed as follows:

[0056]

[0057] in, A Gaussian mixture model representing the set of confidence scores. The mean is variance is Gaussian distribution, The mean is variance is Gaussian distribution, This represents the weight of regions where the confidence score is below a threshold. This represents the weight of regions where the confidence score is higher than the threshold. , , , , and It is estimated using an unsupervised expectation-maximization algorithm;

[0058] Based on the Gaussian mixture model of the confidence score set, the prediction results in the prediction result set are divided into three confidence regions, namely:

[0059] Regions with low confidence are discarded directly; among them... Indicates hyperparameters, with ;

[0060] This represents the region of medium confidence.

[0061] The region with high confidence is extracted as a pseudo-label.

[0062] Preferably, the gradient decomposition and alignment module dynamically decomposes the new label gradient and the pseudo-label gradient, specifically including the following steps:

[0063] In the incremental learning phase of the incremental object detection model, the total loss function of the incremental object detection model is... It includes the loss for the new category and the loss for the old category, and its formula is as follows:

[0064]

[0065] in, This represents the loss for the new category at the current stage, calculated by combining the prediction results of the incremental object detection model with the true labels in the current dataset. The loss representing the pseudo-labels, i.e. the loss of the old category, is calculated from the prediction results of the incremental object detection model and the pseudo-labels generated by the pseudo-label generation and screening module based on the Gaussian mixture model.

[0066] Based on the total loss function of the incremental target detection model The gradient of the incremental target detection model It can be decomposed into two parts: the gradient of the new label and the gradient of the pseudo-label. The formula is as follows:

[0067]

[0068] in, Representing the gradient of the new label, we have The gradient of the new label is the gradient of the new category. These represent the parameters of the object detection model. Representing the pseudo-label gradient, we have The pseudo-label gradient is the old category gradient.

[0069] Preferably, the gradient decomposition and alignment module performs gradient alignment when gradient directions conflict, specifically as follows:

[0070] Gradient of incremental target detection model With pseudo-label gradient When the directions are inconsistent, that is At that time, the pseudo-label gradient Mapped to the gradient of the incremental object detection model The vertical direction is represented as:

[0071]

[0072] in, Represents pseudo-label gradient transpose, The norm of the pseudo-label gradient is represented. This indicates the gradient direction of the adjusted incremental target detection model;

[0073] Gradient direction based on the adjusted incremental target detection model Update the parameters of the incremental target detection model to obtain the incremental target detection model based on gradient decomposition and alignment.

[0074] The beneficial effects of this invention are:

[0075] 1. Dynamic pseudo-tag generation and filtering mechanism

[0076] The pseudo-label generation and filtering mechanism proposed in this invention can effectively solve the problem of unlabeled old category samples in incremental object detection during image object detection in dynamic scenes. By combining the inference results of the previous stage of incremental object detection with a Gaussian mixture model (GMM), this invention achieves accurate filtering of high-confidence pseudo-labels, effectively preserves the knowledge representation of old categories, and significantly alleviates the catastrophic forgetting problem.

[0077] 2. Gradient Decomposition and Alignment Strategies

[0078] Through an innovative gradient decomposition and alignment strategy, this invention achieves dynamic adjustment of gradients between old and new categories during incremental training for image object detection in dynamic scenes, ensuring that the learning of new categories does not compromise the preservation of knowledge from old categories. This mechanism improves the balance between model stability and plasticity, making the model more robust to detection performance at different stages. Attached Figure Description

[0079] Figure 1 The diagram shows a flowchart of an incremental target detection method based on gradient decomposition and alignment provided in an embodiment of the present invention.

[0080] Figure 2 The diagram shows the structure of the incremental target detection model based on gradient decomposition and alignment provided in an embodiment of the present invention. Detailed Implementation

[0081] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the invention, and are not intended to limit the scope of the invention.

[0082] like Figure 1 As shown, an incremental target detection method based on gradient decomposition and alignment includes the following steps:

[0083] S1. Construct an incremental target detection model based on gradient decomposition and alignment;

[0084] S2. Input the image to be detected into the incremental target detection model based on gradient decomposition and alignment to complete the incremental target detection.

[0085] In this embodiment, step S1 specifically includes the following steps:

[0086] Construct an incremental target detection model and perform initial training on the incremental target detection model;

[0087] Incremental learning is performed on the incremental target detection model using a pseudo-label generation and filtering module based on Gaussian mixture model and a gradient decomposition and alignment module, resulting in an incremental target detection model based on gradient decomposition and alignment.

[0088] The pseudo-label generation and filtering module based on Gaussian mixture model is used to receive the prediction results of the incremental target detection model after initial training, and generate and filter pseudo-labels based on the prediction results of the incremental target detection model after initial training.

[0089] The gradient decomposition and alignment module is used to receive the pseudo-labels, dynamically decompose the new label gradient and pseudo-label gradient of the incremental target detection model, perform gradient alignment when gradient directions conflict, and then update the parameters of the incremental target detection model according to the updated gradients to obtain an incremental target detection model based on gradient decomposition and alignment.

[0090] In this embodiment, as Figure 2 As shown, the incremental target detection model includes a feature extraction unit (FeatureExtractor), a candidate region extraction network (RPN), and a ROI head unit;

[0091] The feature extraction unit is used to extract visual feature information from the input image and output the original image visual feature map.

[0092] The candidate region extraction network is used to receive the visual feature map of the original image and generate candidate regions;

[0093] The ROI Head unit is used to extract feature vectors from the candidate regions and to classify and fine-tune the bounding boxes of the feature vectors. In this embodiment, the feature extraction module includes multiple layers of serial convolutional residual blocks;

[0094] The formula for extracting the visual feature map of the original image is:

[0095]

[0096] in, Represents the visual feature map of the original image, with , Represents the set of real numbers. , and These represent the height, width, and number of channels of the original image's visual feature map, respectively. This indicates the feature extraction module. This represents the original input image.

[0097] In this embodiment, the candidate region extraction network receives the visual feature map of the original image and generates candidate regions, specifically including the following steps:

[0098] Receive the visual feature map of the original image, and generate multiple predefined anchor boxes at each location in the visual feature map of the original image, each anchor box... The coordinates are ,in, Indicates anchor frame center point coordinate, Indicates anchor frame center point coordinate, Indicates anchor frame width, Indicates anchor frame The height; In this embodiment, the candidate region extraction network first uses a convolutional kernel with a scale of 3×3, a stride of 1, and padding of 1 to slide on the visual feature map of the original image, so as to keep the spatial size of the output intermediate feature map unchanged, as shown below:

[0099]

[0100] in, This represents the intermediate feature map. Represents the convolution kernel. Represents the visual feature map of the original image;

[0101] The candidate region extraction network processes intermediate feature maps through two parallel branches: a classification branch and a regression branch. The classification branch uses a 1×1 convolutional kernel to perform binary classification (target or background) on each anchor point. This branch outputs 2k channels, where k is the number of anchor points generated by each sliding window. The regression branch uses a 1×1 convolutional kernel to perform bounding box regression on each anchor point, predicting four offsets to adjust the anchor box for a more accurate match with the ground truth bounding box. This branch outputs 4k channels; in this embodiment, k=9.

[0102] The formula for describing the binary classification of the classification branch is as follows:

[0103]

[0104] in, Indicates the first The classification probability of each anchor box. This represents a fully connected layer network used for classification tasks. This represents the normalized exponential function;

[0105] The formula for the bounding box regression of the regression branch is as follows:

[0106]

[0107] in, Indicates anchor frame The offset, Includes 4 offsets, namely: anchor frame center point coordinate offset Anchor frame center point coordinate offset Anchor frame Width offset and anchor frame height offset , This represents a fully connected layer network used for regression tasks;

[0108] According to the anchor frame coordinates and anchor frame The offset is used to calculate the candidate region, and the calculation formula is as follows:

[0109]

[0110] in, Indicates the center point of the candidate region coordinate, Indicates the center point of the candidate region coordinate, Indicates the width of the candidate region. Indicates the height of the candidate region. Represents the natural base.

[0111] In this embodiment, the ROI Head module extracts feature vectors from the candidate regions and performs classification and bounding box fine-tuning on the feature vectors, specifically including the following steps:

[0112] The ROI pooling unit in the ROI Head module extracts a fixed-size feature vector for each candidate region from the visual feature map of the original image. The formula for extracting the feature vector is as follows:

[0113]

[0114] in, Indicates the first The feature vectors corresponding to each candidate region Represents the visual feature map of the original image. Indicates the first Candidate regions, Indicates the ROI pooling unit;

[0115] The feature vector is input into a fully connected layer for classification and bounding box fine-tuning, and the formula is as follows:

[0116]

[0117] in, Indicates the first The classification probability of each candidate region Represents the normalized exponential function, This represents a fully connected layer network used for classification tasks. This represents the finely adjusted bounding box. This represents a fully connected layer network used for regression tasks.

[0118] In this embodiment, the incremental target detection model is initially trained using candidate region extraction network loss and ROI-Head loss;

[0119] The candidate region extraction network loss The formula for expressing this is:

[0120]

[0121] in, This indicates the category prediction for the candidate region. Indicates the true label of the candidate region. This indicates the predicted bounding box of the candidate region. This represents the true bounding box of the candidate region. The normalization factor represents the normalization factor for the classification task. Represents cross-entropy loss, Indicates the first Category prediction for each candidate region Indicates the first The real labels of each candidate region have 1 represents the target, and 0 represents the background. This represents the normalization factor for the regression task. Indicates smoothness loss, Indicates the first Boundary box prediction for each candidate region Indicates the first The true bounding boxes of each candidate region;

[0122] The ROI-Head loss The formula for expressing this is:

[0123]

[0124] in, This indicates the predicted category of the feature vector. This represents the true class label of the feature vector. Indicates the first The predicted category of each feature vector. Indicates the first The true class labels of each feature vector are: , Indicates the background category.

[0125] In this embodiment, the pseudo-label generation and filtering module based on the Gaussian mixture model generates and filters pseudo-labels according to the prediction results of the incremental target detection model after initial training, specifically including the following steps:

[0126] Using an incremental target detection model, inference is performed on the data of the current stage to generate a set of prediction results, which is:

[0127]

[0128] in, This represents the set of prediction results from the incremental object detection model. The first element of the incremental target detection model is represented by the second element. The bounding box positions of the candidate regions Indicates the first The predicted category of each feature vector. The first element of the incremental target detection model is represented by the second element. The confidence scores of each prediction result are: ;

[0129] A Gaussian Mixture Model (GMM) is used to model the set of confidence scores in the prediction result set, resulting in a Gaussian Mixture Model for the confidence score set. The formula for this model is as follows:

[0130]

[0131] in, A Gaussian mixture model representing the set of confidence scores. The mean is variance is Gaussian distribution, The mean is variance is Gaussian distribution, This represents the weight of regions with confidence scores below a threshold, i.e., the weight of low-confidence regions. This represents the weight of regions with confidence scores above a threshold, i.e., the weight of high-confidence regions. , , , , and It is estimated using an unsupervised expectation-maximization algorithm;

[0132] Based on the Gaussian mixture model of the confidence score set, the prediction results in the prediction result set are divided into three confidence regions, namely:

[0133] Regions with low confidence are discarded directly; among them... Indicates hyperparameters, with ;

[0134] This represents the region of medium confidence.

[0135] The region with high confidence is extracted as a pseudo-label.

[0136] In this embodiment, the gradient decomposition and alignment module dynamically decomposes the new label gradient and the pseudo-label gradient, specifically including the following steps:

[0137] In the incremental learning phase of the incremental object detection model, the total loss function of the incremental object detection model is... It includes the loss for the new category and the loss for the old category, and its formula is as follows:

[0138]

[0139] in, The loss for the new category at the current stage is calculated by combining the output of the incremental object detection model based on gradient decomposition and alignment with the ground truth labels in the current dataset. The ground truth labels in the current dataset only contain the new category objects. The loss representing the pseudo-labels, i.e. the loss of the old category, is calculated from the output of the incremental target detection model based on gradient decomposition and alignment and the pseudo-labels generated by the pseudo-label generation and filtering module based on Gaussian mixture model. The pseudo-labels only contain old category targets.

[0140] The total loss function of the incremental target detection model based on gradient decomposition and alignment The gradient of the incremental target detection model based on gradient decomposition and alignment will be used to... It can be decomposed into two parts: the gradient of the new label and the gradient of the pseudo-label. The formula is as follows:

[0141]

[0142] in, Representing the gradient of the new label, we have The gradient of the new label is the gradient of the new category. These represent the parameters of the object detection model. Representing the pseudo-label gradient, we have The pseudo-label gradient is the old category gradient.

[0143] In this embodiment, the gradient decomposition and alignment module performs gradient alignment when gradient directions conflict, specifically as follows:

[0144] Gradient of an incremental target detection model based on gradient decomposition and alignment With pseudo-label gradient When the directions are inconsistent, that is At that time, the pseudo-label gradient Mapped to gradients of an incremental object detection model based on gradient decomposition and alignment The vertical direction is represented as:

[0145]

[0146] in, Represents pseudo-label gradient transpose, The norm of the pseudo-label gradient is represented. This indicates the gradient direction of the adjusted incremental target detection model;

[0147] Gradient direction based on the adjusted incremental target detection model Update the parameters of the incremental object detection model to obtain the incremental object detection model based on gradient decomposition and alignment. .

[0148] Example 2:

[0149] Based on Example 1, this embodiment of the invention provides a training method for an incremental target detection model based on gradient decomposition and alignment, comprising the following steps:

[0150] A1. Data Preprocessing

[0151] Input image size: All input images are uniformly adjusted to a size of 800×600×3, which balances computational efficiency and preservation of image details;

[0152] Data augmentation: During training, data augmentation techniques, including random cropping, horizontal flipping, and random erasing, are used to improve the generalization ability of the model. Data augmentation can effectively increase the diversity of data and prevent overfitting of incremental object detection models based on gradient decomposition and alignment.

[0153] A2. Model Initialization

[0154] The feature extraction module uses the ResNet50 architecture, which consists of multiple residual blocks, each containing skip connections.

[0155] A3. Training Hyperparameter Settings

[0156] Batch size: Set to 16, each batch contains 16 training set images;

[0157] Optimizer: The SGD optimizer is used, which has good convergence and stability.

[0158] Learning rate: The initial learning rate is set to 1×10^2, and it decays to 0.1 and 0.01 at the 8th and 10th epochs, respectively.

[0159] A4. Loss Function

[0160] Using combined candidate region extraction network loss and ROI-Head loss The total loss function. This is used to optimize the candidate region extraction network, enabling it to classify proposed regions and further refine the position of the bounding boxes. The ROI-Head module is used to optimize incremental object detection models based on gradient decomposition and alignment. The ROI-Head module focuses on more accurate classification and bounding box adjustment of these candidate regions to ultimately output object detection results.

[0161] A5. Training Process

[0162] Epochs: The model is usually trained for 12 epochs, and the timing of early stopping is determined based on the performance on the validation set;

[0163] Pseudo-label generation: First, the incremental object detection model based on gradient decomposition and alignment trained in the previous stage is used to infer the training data in the current stage, generating a set of prediction results. Then, a pseudo-label generation and filtering module based on Gaussian mixture model is used to filter high-quality prediction results from the set of prediction results as pseudo-labels.

[0164] Forward propagation: For each input image, visual feature information is extracted from the image to obtain the original image visual feature map. Candidate regions are generated by the candidate region extraction network. The ROI Head module then performs more accurate classification and bounding box adjustment on the candidate regions generated by the candidate region extraction network, and outputs the target detection results.

[0165] Backpropagation: The loss is calculated based on the detection results output by the incremental object detection model based on gradient decomposition and alignment, along with the dataset labels and pseudo-labels generated by the pseudo-label generation and filtering module based on Gaussian mixture model. During backpropagation, the gradient decomposition and alignment module corrects the gradients of some parameters, and finally, the parameters of the incremental object detection model based on gradient decomposition and alignment are updated based on the corrected gradients.

[0166] Early stopping strategy: If the performance on the validation set no longer improves within a certain number of epochs, training is stopped early to prevent overfitting.

[0167] Example 3:

[0168] Based on Examples 1 and 2, the incremental target detection model based on gradient decomposition and alignment, which uses the above training steps and strategies, is experimentally compared with other methods. The experimental performance comparison results are shown in Table 1.

[0169] Table 1. Performance comparison of different methods under different experimental settings on the POSCAL VOC2007 dataset.

[0170]

[0171] The experimental results for some key datasets are shown below:

[0172] Experimental setup for PASCAL VOC2007 dataset 19-1 (the dataset contains 20 categories, 19-1 indicates that task 1 contains 19 categories and task 2 contains 1 category): The incremental object detection model based on gradient decomposition and alignment proposed in this invention achieved an average accuracy (mAP) of 76.1% on the old task and 61.0% on the new task on this dataset.

[0173] Experimental setup for PASCAL VOC2007 dataset 15-5 (the dataset contains 20 categories, 15-5 indicates that task 1 contains 15 categories and task 2 contains 5 categories): The model of this invention achieved 76.7% mAP on the old task and 66.0% mAP on the new task on this dataset.

[0174] Experimental setup for PASCAL VOC2007 dataset 10-10 (the dataset contains 20 categories, 10-10 indicates that task 1 contains 10 categories and task 2 contains 10 categories): The model of this invention achieved 75.7% mAP on the old task and 74.0% mAP on the new task on this dataset.

[0175] The experimental results above demonstrate that the incremental target detection model based on gradient decomposition and alignment proposed in this invention achieves excellent performance under various task settings on the benchmark dataset.

[0176] The joint loss function designed in this invention comprehensively considers the optimization objectives of the true labels of the new category and the pseudo labels of the old category, balancing the learning ability of the new category with the detection performance of the old category. By integrating the Faster R-CNN framework with the incremental learning mechanism, this invention significantly improves the adaptability and performance of the model in incremental object detection tasks.

[0177] This invention reduces computational complexity while maintaining high detection performance through modular design. The pseudo-label generation module and gradient alignment module enable the model to achieve high-quality object detection with lower resource consumption, offering significant advantages in hardware requirements and energy consumption compared to traditional incremental detection methods.

[0178] Experiments on multiple target detection benchmark datasets have validated the superiority of this invention, demonstrating significant improvements in both average accuracy and detection precision compared to state-of-the-art methods. Furthermore, this invention possesses strong applicability and can be widely applied to practical scenarios such as intelligent surveillance, autonomous driving, and dynamic target monitoring, providing an efficient solution for target detection tasks in dynamic environments.

[0179] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. An incremental target detection method based on gradient decomposition and alignment, characterized in that, Includes the following steps: S1. Construct an incremental target detection model based on gradient decomposition and alignment; S2. Input the image to be detected into the incremental target detection model based on gradient decomposition and alignment to complete the incremental target detection; Step S1 specifically includes the following steps: Construct an incremental target detection model and perform initial training on the incremental target detection model; Incremental learning is performed on the incremental target detection model using a pseudo-label generation and filtering module based on Gaussian mixture model and a gradient decomposition and alignment module, resulting in an incremental target detection model based on gradient decomposition and alignment. The pseudo-label generation and filtering module based on the Gaussian mixture model is used to receive the prediction results of the incremental object detection model after initial training, and generate and filter pseudo-labels based on the prediction results of the incremental object detection model after initial training, specifically as follows: Using an incremental target detection model, inference is performed on the data of the current stage to generate a set of prediction results, which is: in, This represents the set of prediction results from the incremental object detection model. The incremental target detection model represents the first... The bounding box positions of the candidate regions Indicates the first The predicted category of each feature vector. The incremental target detection model represents the first... The confidence scores of each prediction result are: ; A Gaussian mixture model is used to model the set of confidence scores in the prediction result set, resulting in a Gaussian mixture model of the confidence score set, which is expressed as follows: in, A Gaussian mixture model representing the set of confidence scores. The mean is variance is Gaussian distribution, The mean is variance is Gaussian distribution, This represents the weight of regions where the confidence score is below a threshold. This represents the weight of regions where the confidence score is higher than the threshold. , , , , and It is estimated using an unsupervised expectation-maximization algorithm; Based on the Gaussian mixture model of the confidence score set, the prediction results in the prediction result set are divided into three confidence regions, namely: Regions with low confidence are discarded directly; among them... Indicates hyperparameters, with ; This represents the region of medium confidence. High-confidence regions are extracted as pseudo-labels, preserving the knowledge representation of the old categories and mitigating the problem of catastrophic forgetting; The gradient decomposition and alignment module is used to receive the pseudo-labels and, based on the total loss function of the incremental target detection model, perform the decomposition and alignment. The gradient of the incremental object detection model is dynamically decomposed into the gradient of the new label and the gradient of the pseudo-label, and then... With pseudo-label gradient Gradient when directions conflict Mapping to pseudo-label gradient The vertical direction ensures that the learning of new categories does not compromise the preservation of knowledge of old categories. Then, the parameters of the incremental object detection model are updated according to the updated gradient, resulting in an incremental object detection model based on gradient decomposition and alignment.

2. The incremental target detection method based on gradient decomposition and alignment according to claim 1, characterized in that, The incremental target detection model includes a feature extraction unit, a candidate region extraction network, and a ROI Head unit. The feature extraction unit is used to extract visual feature information from the input image and output the original image visual feature map. The candidate region extraction network is used to receive the visual feature map of the original image and generate candidate regions; The ROI Head unit is used to extract feature vectors from the candidate regions and to classify and fine-tune the bounding boxes of the feature vectors.

3. The incremental target detection method based on gradient decomposition and alignment according to claim 2, characterized in that, The feature extraction unit includes multiple serial convolutional residual blocks; The formula for extracting the visual feature map of the original image is: in, Represents the visual feature map of the original image, with , Represents the set of real numbers. , and These represent the height, width, and number of channels of the original image's visual feature map, respectively. Represents a feature extraction unit. This represents the original input image.

4. The incremental target detection method based on gradient decomposition and alignment according to claim 2, characterized in that, The candidate region extraction network receives the visual feature map of the original image and generates candidate regions, specifically including the following steps: The original image visual feature map is received, and an intermediate feature map is obtained by sliding a convolution kernel across the original image visual feature map. The formula for the intermediate feature map is as follows: in, This represents the intermediate feature map. Represents the convolution kernel. Represents the visual feature map of the original image; Multiple predefined anchor boxes are generated at each location in the intermediate feature map, each anchor box... The coordinates are ,in, Indicates anchor frame center point coordinate, Indicates anchor frame center point coordinate, Indicates anchor frame width, Indicates anchor frame Height; The candidate region extraction network uses two parallel classification and regression branches to perform binary classification and bounding box regression on the intermediate feature map to obtain the anchor box. The offset, the formula for the binary classification of the classification branch is: in, Indicates the first The classification probability of each anchor box. This represents a fully connected layer network used for classification tasks. This represents the normalized exponential function; The formula for the bounding box regression of the regression branch is as follows: in, Indicates anchor frame The offset, Includes 4 offsets, namely: anchor frame center point coordinate offset Anchor frame center point coordinate offset Anchor frame Width offset and anchor frame height offset , This represents a fully connected layer network used for regression tasks; According to the anchor frame coordinates and anchor frame The offset is used to calculate the candidate region, and the calculation formula is as follows: in, Indicates the center point of the candidate region coordinate, Indicates the center point of the candidate region coordinate, Indicates the width of the candidate region. Indicates the height of the candidate region. Represents the natural base.

5. The incremental target detection method based on gradient decomposition and alignment according to claim 2, characterized in that, The ROI Head unit extracts feature vectors from the candidate regions and performs classification and bounding box fine-tuning on the feature vectors, specifically including the following steps: The ROI pooling layer in the ROI Head unit extracts a fixed-size feature vector for each candidate region from the visual feature map of the original image. The formula for extracting the feature vector is as follows: in, Indicates the first The feature vectors corresponding to each candidate region Represents the visual feature map of the original image. Indicates the first Candidate regions, Indicates the ROI pooling unit; The feature vector is input into a fully connected layer for classification and bounding box fine-tuning, and the formula is as follows: in, Indicates the first The classification probability of each candidate region Represents the normalized exponential function, This represents a fully connected layer network used for classification tasks. This represents the finely adjusted bounding box. This represents a fully connected layer network used for regression tasks.

6. The incremental target detection method based on gradient decomposition and alignment according to claim 2, characterized in that, The incremental target detection model is initially trained using candidate region extraction network loss and ROI head loss. The candidate region extraction network loss The formula for expressing this is: in, This indicates the category prediction for the candidate region. Indicates the true label of the candidate region. This indicates the predicted bounding box of the candidate region. This represents the true bounding box of the candidate region. The normalization factor represents the normalization factor for the classification task. Represents cross-entropy loss, Indicates the first Category prediction for each candidate region Indicates the first The real labels of each candidate region have 1 represents the target, and 0 represents the background. This represents the normalization factor for the regression task. Indicates smoothness loss, Indicates the first Boundary box prediction for each candidate region Indicates the first The true bounding boxes of each candidate region; The ROI Head loss The formula for expressing this is: in, This indicates the predicted category of the feature vector. This represents the true class label of the feature vector. Indicates the first The predicted category of each feature vector. Indicates the first The true class labels of each feature vector are: , Indicates the background category.

7. The incremental target detection method based on gradient decomposition and alignment according to claim 1, characterized in that, The gradient decomposition and alignment module dynamically decomposes the gradient of the new label and the gradient of the pseudo-label, specifically including the following steps: In the incremental learning phase of the incremental object detection model, the total loss function of the incremental object detection model is... It includes the loss for the new category and the loss for the old category, and its formula is as follows: in, This represents the loss for the new category at the current stage, calculated by combining the prediction results of the incremental object detection model with the true labels in the current dataset. The loss representing the pseudo-labels, i.e. the loss of the old category, is calculated from the prediction results of the incremental object detection model and the pseudo-labels generated by the pseudo-label generation and screening module based on the Gaussian mixture model. Based on the total loss function of the incremental target detection model The gradient of the incremental target detection model It can be decomposed into two parts: the gradient of the new label and the gradient of the pseudo-label. The formula is as follows: in, Representing the gradient of the new label, we have The gradient of the new label is the gradient of the new category. These represent the parameters of the object detection model. Representing the pseudo-label gradient, we have The pseudo-label gradient is the old category gradient.

8. The incremental target detection method based on gradient decomposition and alignment according to claim 7, characterized in that, The gradient decomposition and alignment module is used in the gradient of the incremental target detection model. With pseudo-label gradient Gradient when directions conflict Mapping to pseudo-label gradient In the vertical direction, this ensures that learning new categories does not compromise the preservation of knowledge from old categories. Then, the parameters of the incremental object detection model are updated based on the updated gradient, resulting in an incremental object detection model based on gradient decomposition and alignment, specifically: Gradient of incremental target detection model With pseudo-label gradient When the directions are inconsistent, that is At that time, the gradient of the incremental target detection model will be... Mapping to pseudo-label gradient The vertical direction is represented as: in, Represents pseudo-label gradient transpose, The norm of the pseudo-label gradient is represented. This indicates the gradient direction of the adjusted incremental target detection model; Gradient direction based on the adjusted incremental target detection model Update the parameters of the incremental target detection model to obtain the incremental target detection model based on gradient decomposition and alignment.

Images