Incremental target detection method based on gradient decomposition and alignment

Abstract

The invention belongs to the technical field of computer vision, and particularly discloses an incremental target detection method based on gradient decomposition and alignment, which comprises the following steps: constructing an incremental target detection model based on gradient decomposition and alignment; and finishing incremental target detection by using an incremental target detection model based on gradient decomposition and alignment. According to the method, the problem that old class samples are not labeled in incremental target detection can be effectively solved through a pseudo label generation and screening mechanism, a gradient decomposition and alignment strategy is proposed, dynamic adjustment of new and old class gradients is realized in incremental training, new class learning is ensured, meanwhile, retention of old class knowledge is not damaged, and the method is suitable for large-scale popularization and application. The stability and plasticity balance of the incremental target detection model is improved, and then the problems of disastrous forgetting, category overlapping and difficulty in optimizing new and old category learning balance in incremental target detection are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and in particular relates to an incremental target detection method based on gradient decomposition and alignment. Background Art

[0002] The incremental object detection task is an important research direction in the field of computer vision. Its goal is to enable the object detection model to continuously learn new categories while maintaining the ability to detect old categories. Incremental learning is particularly important in dynamically changing application scenarios, such as intelligent monitoring and unmanned driving, where the system needs to adapt to new target categories in real time without losing existing knowledge.

[0003] Traditional object detection methods are usually trained on static datasets, and all category annotations are fully provided at the beginning of training. However, in real-world scenarios, the number of target categories tends to increase gradually, causing the model to face the problem of catastrophic forgetting, that is, losing the ability to detect old categories when learning new categories. In addition, incremental learning also has the problem of category overlap. In the training data of the new stage, the targets of the old categories are often not labeled and are easily misclassified as background, further exacerbating the phenomenon of knowledge forgetting.

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

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

[0006] The purpose of the present invention is to solve the problems of catastrophic forgetting, category overlap and difficulty in optimizing the learning balance of new and old categories in incremental target detection when monitoring targets in dynamic environments. An incremental target detection method based on gradient decomposition and alignment is proposed.

[0007] The technical solution of the present invention is: an incremental target detection method based on gradient decomposition and alignment, comprising the following steps: S1. Construct an incremental object 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 incremental target detection.

[0008] Preferably, the step S1 specifically includes the following steps: Build an incremental target detection model and perform initial training on the incremental target detection model; The incremental object detection model is incrementally learned using the pseudo-label generation and screening module based on the Gaussian mixture model and the gradient decomposition and alignment module to obtain an incremental object detection model based on gradient decomposition and alignment. The Gaussian mixture model-based pseudo-label generation and screening module is used to receive the prediction results of the incremental target detection model that has completed initial training, and generate and screen pseudo-labels according to the prediction results of the incremental target detection model that has completed initial training; The gradient decomposition and alignment module is used to receive the pseudo-label, dynamically decompose the new label gradient and pseudo-label gradient of the incremental target detection model, and perform gradient alignment when the gradient directions conflict, and then update the parameters of the incremental target detection model according to the updated gradient to obtain the incremental target detection model based on gradient decomposition and alignment.

[0009] Preferably, the incremental target detection model includes a feature extraction unit, a candidate region extraction network and a ROIHead unit; The feature extraction unit is used to extract visual feature information from the input image and output a visual feature map of the original image; The candidate region extraction network is used to receive the original image visual feature map and generate a candidate region; The ROI Head unit is used to extract feature vectors from the candidate region, and classify and fine-tune the boundary box of the feature vectors.

[0010] Preferably, the feature extraction unit comprises a multi-layer serial convolution residual block; The extraction formula of the original image visual feature map is:

[0011] in, Represents the visual feature map of the original image, , represents the set of real numbers, , and Respectively represent the height, width and number of channels of the original image visual feature map, represents the feature extraction unit, Represents the original input image.

[0012] Preferably, the candidate region extraction network receives the original image visual feature map and generates a candidate region, specifically comprising the following steps: The original image visual feature map is received, and the convolution kernel is used to slide on the original image visual feature map to obtain an intermediate feature map. The expression formula of the intermediate feature map is:

[0013] in, represents the intermediate feature map, represents the convolution kernel, Represents the visual feature map of the original image; A plurality of predefined anchor boxes are generated at each position of the intermediate feature map, each anchor box The coordinates of ,in, Represents the anchor box Center point coordinate, Represents the anchor box Center point coordinate, Represents the anchor box The width of Represents the anchor box Height; The candidate region extraction network uses two parallel classification branches and regression branches to perform binary classification and bounding box regression on the intermediate feature map to obtain the anchor box. The offset of the classification branch is expressed as follows:

[0014] in, Indicates The classification probability of anchor boxes, represents a fully connected layer network for classification tasks, represents the normalized exponential function; The regression branch performs bounding box regression as follows:

[0015] in, Represents the anchor box The offset of Includes 4 offsets, namely: anchor box Center point Coordinate offset , Anchor Frame Center point Coordinate offset , Anchor Frame Width offset and anchor box Height offset , represents a fully connected layer network for regression tasks; According to the anchor box The coordinates and anchor box of The candidate area is calculated by the offset, and the calculation formula is:

[0016] in, Represents the center point of the candidate region coordinate, Represents the center point of the candidate region coordinate, represents the width of the candidate region, represents the height of the candidate region, Represents the natural base.

[0017] Preferably, the ROI Head unit extracts a feature vector from the candidate region, and classifies the feature vector and fine-tunes the bounding box, specifically comprising the following steps: The ROI pooling layer in the ROI Head unit extracts a fixed-size feature vector for each candidate region from the original image visual feature map. The extraction formula of the feature vector is:

[0018] in, Indicates The feature vector corresponding to the candidate region is Represents the visual feature map of the original image, Indicates candidate regions, Represents ROI pooling unit; The feature vector is input into the fully connected layer for classification and bounding box fine-tuning, and the formula is:

[0019] in, Indicates The classification probability of the candidate regions is represents the normalized exponential function, represents a fully connected layer network for classification tasks, represents the fine-tuned bounding box, Represents a fully connected layer network for regression tasks.

[0020] Preferably, the incremental target detection model is initially trained using a candidate region extraction network loss and a ROI-Head loss; The candidate region extraction network loss The expression formula is:

[0021] in, represents the category prediction of the candidate region, represents the true label of the candidate region, represents the bounding box prediction of the candidate region, represents the ground-truth bounding box of the candidate region, represents the normalization factor for the classification task, represents the cross entropy loss, Indicates Category prediction of candidate regions, Indicates The true labels of the candidate regions are , 1 represents the target, 0 represents the background, represents the normalization factor for the regression task, Indicates smoothness loss, Indicates Bounding box predictions for candidate regions, Indicates The ground-truth bounding boxes of the candidate regions; The ROI-Head loss The expression formula is:

[0022] in, represents the predicted category of the feature vector, represents the true category label of the feature vector, Indicates The predicted category of the feature vector, Indicates The true category labels of the feature vectors are , Represents the background category.

[0023] Preferably, the pseudo-label generation and screening module based on the Gaussian mixture model generates and screens pseudo-labels according to the prediction results of the incremental target detection model that has completed initial training, and specifically includes the following steps: The incremental target detection model is used to infer the data in the current stage to generate a set of prediction results, which are:

[0024] in, represents the set of prediction results of the incremental target detection model, represents the incremental object detection model The bounding box locations of the candidate regions, Indicates The predicted category of the feature vector, represents the incremental object detection model The confidence score of the prediction result is ; Based on the Gaussian mixture model, the set of confidence scores in the prediction result set is modeled to obtain the Gaussian mixture model of the confidence score set, which is expressed as follows:

[0025] in, A Gaussian mixture model representing a set of confidence scores, Indicates the mean , the variance is Gaussian distribution, Indicates the mean , the variance is Gaussian distribution, represents the weight of the region whose confidence score is below the threshold, represents the weight of the region with confidence score above the threshold, , , , , and It is estimated by an unsupervised expectation maximization algorithm; According to the Gaussian mixture model of the confidence score set, the prediction results in the prediction result set are divided into three confidence regions, namely: , is a low confidence area and is directly discarded; among them, represents a super parameter, ; , is the medium confidence region; , which is a high confidence region, is extracted as a pseudo label.

[0026] Preferably, the gradient decomposition and alignment module dynamically decomposes the new label gradient and the pseudo label gradient, specifically comprising the following steps: In the incremental learning phase of the incremental target detection model, the total loss function of the incremental target detection model is Including the loss of new categories and the loss of old categories, the expression formula is:

[0027] in, Represents the loss of the new category in the current stage, which is calculated by the prediction results of the incremental target detection model and the true label annotated by the current dataset. represents the loss of pseudo labels, that is, the loss of old categories, which is calculated by 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; The total loss function of the incremental object detection model , the gradient of the incremental target detection model It is decomposed into two parts: new label gradient and pseudo label gradient, and its expression formula is:

[0028] in, represents the new label gradient, , the new label gradient is the new category gradient, represents the parameters of the target detection model, represents the pseudo-label gradient, , the pseudo-label gradient is the old category gradient.

[0029] Preferably, the gradient decomposition and alignment module performs gradient alignment when the gradient directions conflict, specifically: Incremental object detection model gradients With pseudo-labeled gradient When the directions are inconsistent, When the pseudo-label gradient Mapping to the gradient of the incremental object detection model The vertical direction is expressed as:

[0030] in, Represents pseudo-label gradient The transpose of represents the norm of the pseudo-label gradient, Represents the gradient direction of the adjusted incremental target detection model; Gradient directions based on the adjusted incremental object detection model The parameters of the incremental target detection model are updated, that is, an incremental target detection model based on gradient decomposition and alignment is obtained.

[0031] The beneficial effects of the present invention are: 1. Dynamic pseudo-label generation and screening mechanism The pseudo-label generation and screening mechanism proposed in the present invention can effectively solve the problem of unlabeled samples of old categories in incremental target detection when performing image target detection in dynamic scenes. By combining the reasoning results of incremental target detection in the previous stage with the Gaussian mixture model (GMM), the present invention achieves accurate screening of high-confidence pseudo-labels, effectively retains the knowledge representation of old categories, and significantly alleviates the problem of catastrophic forgetting.

[0032] 2. Gradient decomposition and alignment strategy Through innovative gradient decomposition and alignment strategies, the present invention achieves dynamic adjustment of new and old category gradients in incremental training for image target detection in dynamic scenes, ensuring that new categories are learned without compromising the retention of old category knowledge. This mechanism improves the balance between model stability and plasticity, making the model's detection performance more robust at different stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Shown is a flow chart of an incremental target detection method based on gradient decomposition and alignment provided in an embodiment of the present invention.

[0034] Figure 2 Shown is a structural diagram of an incremental target detection model based on gradient decomposition and alignment provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0036] like Figure 1 As shown, an incremental target detection method based on gradient decomposition and alignment includes the following steps: S1. Construct an incremental object 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 incremental target detection.

[0037] In this embodiment, step S1 specifically includes the following steps: Build an incremental target detection model and perform initial training on the incremental target detection model; The incremental object detection model is incrementally learned using the pseudo-label generation and screening module based on the Gaussian mixture model and the gradient decomposition and alignment module to obtain an incremental object detection model based on gradient decomposition and alignment. The Gaussian mixture model-based pseudo-label generation and screening module is used to receive the prediction results of the incremental target detection model that has completed initial training, and generate and screen pseudo-labels according to the prediction results of the incremental target detection model that has completed initial training; The gradient decomposition and alignment module is used to receive the pseudo-label, dynamically decompose the new label gradient and pseudo-label gradient of the incremental target detection model, and perform gradient alignment when the gradient directions conflict, and then update the parameters of the incremental target detection model according to the updated gradient to obtain the incremental target detection model based on gradient decomposition and alignment.

[0038] In this embodiment, if Figure 2As shown, the incremental target detection model includes a feature extraction unit (FeatureExtractor), a candidate region extraction network (RPN) and a ROI Head unit; The feature extraction unit is used to extract visual feature information from the input image and output a visual feature map of the original image; The candidate region extraction network is used to receive the original image visual feature map and generate a candidate region; The ROI Head unit is used to extract feature vectors from the candidate region, and classify and fine-tune the feature vectors. In this embodiment, the feature extraction module includes multiple layers of serial convolution residual blocks; The extraction formula of the original image visual feature map is:

[0039] in, Represents the visual feature map of the original image, , represents the set of real numbers, , and Respectively represent the height, width and number of channels of the original image visual feature map, represents the feature extraction module, Represents the original input image.

[0040] In this embodiment, the candidate region extraction network receives the original image visual feature map and generates a candidate region, specifically including the following steps: Receive the original image visual feature map, and generate a plurality of predefined anchor boxes at each position of the original image visual feature map, each anchor box The coordinates of ,in, Represents the anchor box Center point coordinate, Represents the anchor box Center point coordinate, Represents the anchor box The width of Represents the anchor box In this embodiment, the candidate region extraction network first uses a convolution kernel with a scale of 3×3, a step size of 1, and a padding of 1 to slide on the original image visual feature map to keep the spatial size of the output intermediate feature map unchanged, which is expressed as:

[0041] in, represents the intermediate feature map, represents the convolution kernel, Represents the visual feature map of the original image; The candidate region extraction network processes the intermediate feature map through two parallel classification branches and regression branches. The classification branch is a 1×1 convolution kernel, which is used to classify the anchor of each anchor box into two categories (target or background). The number of output channels of this branch is 2k, where k is the number of anchors generated by each sliding window; the regression branch is a 1×1 convolution kernel, which is used to regress the bounding box for each anchor point and predict 4 offsets to adjust the anchor box to more accurately match the real bounding box. The number of output channels of this branch is 4k. In this embodiment, k=9; The expression formula for the classification branch to perform binary classification is:

[0042] in, Indicates The classification probability of anchor boxes, represents a fully connected layer network for classification tasks, represents the normalized exponential function; The regression branch performs bounding box regression as follows:

[0043] in, Represents the anchor box The offset of Includes 4 offsets, namely: anchor box Center point Coordinate offset , Anchor Frame Center point Coordinate offset , Anchor Frame Width offset and anchor box Height offset , represents a fully connected layer network for regression tasks; According to the anchor box The coordinates and anchor box of The candidate area is calculated by the offset, and the calculation formula is:

[0044] in, Represents the center point of the candidate region coordinate, Represents the center point of the candidate region coordinate, represents the width of the candidate region, represents the height of the candidate region, Represents the natural base.

[0045] In this embodiment, the ROI Head module extracts feature vectors from the candidate regions, and classifies and fine-tunes the feature vectors, specifically including the following steps: The ROI pooling unit in the ROI Head module extracts a fixed-size feature vector for each candidate region from the original image visual feature map. The feature vector extraction formula is:

[0046] in, Indicates The feature vector corresponding to the candidate region is Represents the visual feature map of the original image, Indicates candidate regions, Represents ROI pooling unit; The feature vector is input into the fully connected layer for classification and bounding box fine-tuning, and the formula is:

[0047] in, Indicates The classification probability of the candidate regions is represents the normalized exponential function, represents a fully connected layer network for classification tasks, represents the fine-tuned bounding box, Represents a fully connected layer network for regression tasks.

[0048] In this embodiment, the incremental object detection model is initially trained using the candidate region extraction network loss and the ROI-Head loss; The candidate region extraction network loss The expression formula is:

[0049] in, represents the category prediction of the candidate region, represents the true label of the candidate region, represents the bounding box prediction of the candidate region, represents the ground-truth bounding box of the candidate region, represents the normalization factor for the classification task, represents the cross entropy loss, Indicates Category prediction of candidate regions, Indicates The true labels of the candidate regions are , 1 represents the target, 0 represents the background, represents the normalization factor for the regression task, Indicates smoothness loss, Indicates Bounding box predictions for candidate regions, Indicates The ground-truth bounding boxes of the candidate regions; The ROI-Head loss The expression formula is:

[0050] in, represents the predicted category of the feature vector, represents the true category label of the feature vector, Indicates The predicted category of the feature vector, Indicates The true category labels of the feature vectors are , Represents the background category.

[0051] In this embodiment, the pseudo-label generation and screening module based on the Gaussian mixture model generates and screens pseudo-labels according to the prediction results of the incremental target detection model that has completed the initial training, and specifically includes the following steps: The incremental target detection model is used to infer the data in the current stage to generate a set of prediction results, which are:

[0052] in, represents the set of prediction results of the incremental target detection model, represents the incremental object detection model The bounding box locations of the candidate regions, Indicates The predicted category of the feature vector, represents the incremental object detection model The confidence score of the prediction result is ; Based on the Gaussian mixture model (GMM), the set of confidence scores in the prediction result set is modeled to obtain the Gaussian mixture model of the confidence score set, which is expressed as follows:

[0053] in, A Gaussian mixture model representing a set of confidence scores, Indicates the mean , the variance is Gaussian distribution, Indicates the mean , the variance is Gaussian distribution, represents the weight of the area whose confidence score is lower than the threshold, that is, the weight of the low confidence area, represents the weight of the area whose confidence score is higher than the threshold, that is, the weight of the high confidence area, , , , , and It is estimated by an unsupervised expectation maximization algorithm; According to the Gaussian mixture model of the confidence score set, the prediction results in the prediction result set are divided into three confidence regions, namely: , is a low confidence area and is directly discarded; among them, represents a super parameter, ; , is the medium confidence region; , which is a high confidence region, is extracted as a pseudo label.

[0054] 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: In the incremental learning phase of the incremental target detection model, the total loss function of the incremental target detection model is Including the loss of new categories and the loss of old categories, the expression formula is:

[0055] in, represents the loss of the new category at the current stage, which is calculated by the output of the incremental target detection model based on gradient decomposition and alignment and the true labels annotated by the current dataset. The true labels annotated by the current dataset only contain new category targets. The pseudo-label loss, i.e., the loss of the old category, is calculated by the output of the incremental object detection model based on gradient decomposition and alignment and the pseudo-label generated by the pseudo-label generation and filtering module based on the Gaussian mixture model. The pseudo-label only contains the old category objects. The total loss function of the incremental object detection model based on gradient decomposition and alignment , the gradient of the incremental target detection model based on gradient decomposition and alignment It is decomposed into two parts: new label gradient and pseudo label gradient, and its expression formula is:

[0056] in, represents the new label gradient, , the new label gradient is the new category gradient, represents the parameters of the target detection model, represents the pseudo-label gradient, , the pseudo-label gradient is the old category gradient.

[0057] In this embodiment, the gradient decomposition and alignment module performs gradient alignment when the gradient directions conflict, specifically: Gradients of incremental object detection models based on gradient decomposition and alignment With pseudo-labeled gradient When the directions are inconsistent, When the pseudo-label gradient Gradients mapped to incremental object detection models based on gradient decomposition and alignment The vertical direction is expressed as:

[0058] in, Represents pseudo-label gradient The transpose of represents the norm of the pseudo-label gradient, Represents the gradient direction of the adjusted incremental target detection model; Gradient directions based on the adjusted incremental object detection model Update the parameters of the incremental target detection model, that is, obtain the incremental target detection model based on gradient decomposition and alignment .

[0059] Embodiment 2: Based on Example 1, the embodiment of the present invention provides a training method for an incremental target detection model based on gradient decomposition and alignment, comprising the following steps: A1. Data Preprocessing Input image size: All input images are uniformly resized to 800×600×3, which can balance computational efficiency and image detail preservation; Data enhancement: During the training process, data enhancement techniques including random cropping, horizontal flipping, and random erasing are used to improve the generalization ability of the model. Data enhancement can effectively increase the diversity of data and prevent overfitting of the incremental target detection model based on gradient decomposition and alignment.

[0060] A2. Model Initialization The feature extraction module uses the ResNet50 architecture, which consists of multiple residual blocks, each of which contains a skip connection.

[0061] A3. Training hyperparameter settings Batch size: set to 16, each batch contains 16 training set images; Optimizer: Use the SGD optimizer, which has good convergence and stability; Learning rate: The initial learning rate is set to 1×10^2, and decays to 0.1 and 0.01 at the 8th and 10th epochs.

[0062] A4. Loss Function Using the combined proposal region extraction network loss and ROI-Head loss The total loss function. Used to optimize the candidate region extraction network so that the candidate region extraction network can classify the proposed region and further refine the position of the bounding box. The ROI-Head module used to optimize the incremental target detection model based on gradient decomposition and alignment focuses on more accurate classification and bounding box adjustment of these candidate regions to ultimately output target detection results.

[0063] A5. Training process Epochs: The model is usually trained for 12 epochs, and the timing of early stopping is determined based on the performance of the validation set; Pseudo-label generation: First, use the incremental target detection model based on gradient decomposition and alignment trained in the previous stage to infer the training data in the current stage and generate a set of prediction results. Then use the pseudo-label generation and screening module based on the Gaussian mixture model to screen high-quality prediction results from the prediction result set as pseudo-labels; Forward propagation: For each input image, the visual feature information is extracted from the image to obtain the visual feature map of the original image, and the candidate region is generated by the candidate region extraction network. The ROI Head module then performs more accurate classification and bounding box adjustment on the candidate region generated by the candidate region extraction network, and outputs the target detection result.

[0064] Directional propagation: The loss is calculated based on the detection results output by the incremental object detection model based on gradient decomposition and alignment, the dataset labels, and the pseudo labels generated by the pseudo label generation and screening module based on the Gaussian mixture model. During the back propagation process, the gradient decomposition and alignment module corrects some parameter gradients, and finally updates the parameters of the incremental object detection model based on gradient decomposition and alignment based on the corrected gradients.

[0065] Early stopping strategy: If the performance of the validation set does not improve within a few epochs, stop training early to prevent overfitting.

[0066] Embodiment 3: Based on Example 1 and Example 2, an incremental object detection model based on gradient decomposition and alignment using the above training steps and strategies is experimentally compared with other methods. The experimental performance comparison results are shown in Table 1.

[0067] Table 1 Performance comparison of different methods under different experimental settings on the POSCAL VOC2007 dataset

[0068] The experimental results of some key data sets are as follows: PASCAL VOC2007 dataset 19-1 experimental setting (the dataset contains 20 categories in total, 19-1 means 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 precision (mAP) of 76.1% on the old task and 61.0% on the new task on this dataset.

[0069] PASCAL VOC2007 dataset 15-5 experimental setting (the dataset contains 20 categories in total, 15-5 means task 1 contains 15 categories, and task 2 contains 1 category): the proposed model achieved 76.7% mAP on the old task and 66.0% mAP on the new task on this dataset.

[0070] PASCAL VOC2007 dataset 10-10 experimental setting (the dataset contains 20 categories in total, 10-10 means task 1 contains 10 categories, and task 2 contains 10 categories): the proposed model achieved 75.7% mAP on the old task and 74.0% mAP on the new task on this dataset.

[0071] The above experimental results show that under various task settings of the benchmark dataset, the incremental target detection model based on gradient decomposition and alignment proposed in this invention achieves excellent performance.

[0072] The joint loss function designed by the present invention comprehensively considers the optimization goals of the real labeling of new categories and the pseudo labeling of old categories, taking into account the learning ability of new categories and the detection performance of old categories. By integrating the Faster R-CNN framework with the incremental learning mechanism, the present invention greatly improves the adaptability and performance of the model in incremental target detection tasks.

[0073] While retaining high detection performance, the present invention reduces computational complexity through modular design. The pseudo-label generation module and gradient alignment module enable the model to complete high-quality target detection with lower resource consumption, which has significant advantages in hardware requirements and energy consumption compared with traditional incremental detection methods.

[0074] The experiments on multiple target detection benchmark datasets have verified the superiority of the present invention. Compared with the most advanced methods, the average precision and detection accuracy have been significantly improved. In addition, the present invention has strong applicability and can be widely used in practical scenarios such as intelligent monitoring, autonomous driving and dynamic target monitoring, providing an efficient solution for target detection tasks in dynamic environments.

[0075] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. An incremental target detection method based on gradient decomposition and alignment, characterized in that: The following steps are involved: S1. Construct an incremental object 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 incremental target detection.

2. The incremental target detection method based on gradient decomposition and alignment according to claim 1, characterized in that: The step S1 specifically includes the following steps: Build an incremental target detection model and perform initial training on the incremental target detection model; The incremental object detection model is incrementally learned using the pseudo-label generation and screening module based on the Gaussian mixture model and the gradient decomposition and alignment module to obtain an incremental object detection model based on gradient decomposition and alignment. The Gaussian mixture model-based pseudo-label generation and screening module is used to receive the prediction results of the incremental target detection model that has completed initial training, and generate and screen pseudo-labels according to the prediction results of the incremental target detection model that has completed initial training; The gradient decomposition and alignment module is used to receive the pseudo-label, dynamically decompose the new label gradient and pseudo-label gradient of the incremental target detection model, and perform gradient alignment when the gradient directions conflict, and then update the parameters of the incremental target detection model according to the updated gradient to obtain the incremental target detection model based on gradient decomposition and alignment.

3. The incremental target detection method based on gradient decomposition and alignment according to claim 2, 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 a visual feature map of the original image; The candidate region extraction network is used to receive the original image visual feature map and generate a candidate region; The ROI Head unit is used to extract feature vectors from the candidate region, and classify and fine-tune the bounding box of the feature vectors.

4. The incremental target detection method based on gradient decomposition and alignment according to claim 3 is characterized in that: The feature extraction unit includes multiple layers of serial convolution residual blocks; The extraction formula of the original image visual feature map is: in, Represents the visual feature map of the original image, , represents the set of real numbers, , and Respectively represent the height, width and number of channels of the original image visual feature map, represents the feature extraction unit, Represents the original input image.

5. The incremental target detection method based on gradient decomposition and alignment according to claim 3, characterized in that: The candidate region extraction network receives the original image visual feature map and generates a candidate region, specifically comprising the following steps: The original image visual feature map is received, and the convolution kernel is used to slide on the original image visual feature map to obtain an intermediate feature map. The expression formula of the intermediate feature map is: in, represents the intermediate feature map, represents the convolution kernel, Represents the visual feature map of the original image; A plurality of predefined anchor boxes are generated at each position of the intermediate feature map, each anchor box The coordinates of ,in, Represents the anchor box Center point coordinate, Represents the anchor box Center point coordinate, Represents the anchor box The width of Represents the anchor box Height; The candidate region extraction network uses two parallel classification branches and regression branches to perform binary classification and bounding box regression on the intermediate feature map to obtain the anchor box. The offset of the classification branch is expressed as follows: in, Indicates The classification probability of anchor boxes, represents a fully connected layer network for classification tasks, represents the normalized exponential function; The regression branch performs bounding box regression as follows: in, Represents the anchor box The offset of Includes 4 offsets, namely: anchor box Center point Coordinate offset , Anchor Frame Center point Coordinate offset , Anchor Frame Width offset and anchor box Height offset , represents a fully connected layer network for regression tasks; According to the anchor box The coordinates and anchor box of The candidate area is calculated by the offset, and the calculation formula is: in, Represents the center point of the candidate region coordinate, Represents the center point of the candidate region coordinate, represents the width of the candidate region, represents the height of the candidate region, Represents the natural base.

6. The incremental target detection method based on gradient decomposition and alignment according to claim 3 is characterized in that: The ROI Head unit extracts a feature vector from the candidate region, and classifies the feature vector and fine-tunes the bounding box, 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 original image visual feature map. The extraction formula of the feature vector is: in, Indicates The feature vector corresponding to the candidate region is represents the visual feature map of the original image, Indicates candidate regions, Represents ROI pooling unit; The feature vector is input into the fully connected layer for classification and bounding box fine-tuning, and the formula is: in, Indicates The classification probability of the candidate regions is represents the normalized exponential function, represents a fully connected layer network for classification tasks, represents the fine-tuned bounding box, Represents a fully connected layer network for regression tasks.

7. The incremental target detection method based on gradient decomposition and alignment according to claim 3, characterized in that: Initially training the incremental object detection model using a candidate region extraction network loss and a ROI-Head loss; The candidate region extraction network loss The expression formula is: in, represents the category prediction of the candidate region, represents the true label of the candidate region, represents the bounding box prediction of the candidate region, represents the ground-truth bounding box of the candidate region, represents the normalization factor for the classification task, represents the cross entropy loss, Indicates Category prediction of candidate regions, Indicates The true labels of the candidate regions are , 1 represents the target, 0 represents the background, represents the normalization factor for the regression task, Indicates smoothness loss, Indicates Bounding box prediction for each candidate region, Indicates The ground-truth bounding boxes of the candidate regions; The ROI-Head loss The expression formula is: in, represents the predicted category of the feature vector, represents the true category label of the feature vector, Indicates The predicted category of the feature vector, Indicates The true category labels of the feature vectors are , Represents the background category.

8. The incremental target detection method based on gradient decomposition and alignment according to claim 2, characterized in that: The Gaussian mixture model-based pseudo-label generation and screening module generates and screens pseudo-labels according to the prediction results of the incremental target detection model that has completed initial training, and specifically includes the following steps: The incremental target detection model is used to infer the data in the current stage to generate a set of prediction results, which are: in, represents the set of prediction results of the incremental target detection model, represents the incremental object detection model The bounding box locations of the candidate regions, Indicates The predicted category of the feature vector, represents the incremental object detection model The confidence score of the prediction result is ; Based on the Gaussian mixture model, the set of confidence scores in the prediction result set is modeled to obtain the Gaussian mixture model of the confidence score set, which is expressed as follows: in, A Gaussian mixture model representing a set of confidence scores, It means the mean , the variance is Gaussian distribution, It means the mean , the variance is Gaussian distribution, represents the weight of the region whose confidence score is below the threshold, represents the weight of the region with confidence score above the threshold, , , , , and It is estimated by an unsupervised expectation maximization algorithm; According to the Gaussian mixture model of the confidence score set, the prediction results in the prediction result set are divided into three confidence regions, namely: , is a low confidence area and is directly discarded; among them, represents a super parameter, ; , is the medium confidence region; , which is a high confidence region, is extracted as a pseudo label.

9. The incremental target detection method based on gradient decomposition and alignment according to claim 2, characterized in that: The gradient decomposition and alignment module dynamically decomposes the new label gradient and the pseudo label gradient, specifically including the following steps: In the incremental learning phase of the incremental target detection model, the total loss function of the incremental target detection model is Including the loss of new categories and the loss of old categories, the expression formula is: in, Represents the loss of the new category in the current stage, which is calculated by the prediction results of the incremental target detection model and the true label annotated by the current dataset. represents the loss of pseudo labels, that is, the loss of old categories, which is calculated by 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; The total loss function of the incremental object detection model , the gradient of the incremental target detection model It is decomposed into two parts: new label gradient and pseudo label gradient, and its expression formula is: in, represents the new label gradient, , the new label gradient is the new category gradient, represents the parameters of the target detection model, represents the pseudo-label gradient, , the pseudo-label gradient is the old category gradient.

10. The incremental target detection method based on gradient decomposition and alignment according to claim 9, characterized in that: The gradient decomposition and alignment module performs gradient alignment when the gradient directions conflict, specifically: Incremental object detection model gradients With pseudo-labeled gradient When the directions are inconsistent, When the pseudo-label gradient Mapping to the gradient of the incremental object detection model The vertical direction is expressed as: in, Represents pseudo-label gradient The transpose of represents the norm of the pseudo-label gradient, Represents the gradient direction of the adjusted incremental target detection model; Gradient directions based on the adjusted incremental object detection model The parameters of the incremental target detection model are updated, that is, an incremental target detection model based on gradient decomposition and alignment is obtained.

Images