A curriculum learning distillation-based incremental target detection method, system and device

Abstract

The application discloses a kind of based on course type learning distillation incremental target detection method, system and equipment, it is related to computer vision technical field, including, obtain the example set of old class sample and the new class training data of current stage, and initialize teacher model and student model;Input image is sent into teacher model, the prediction result of teacher model is Top-K screening and IOU conflict suppression, generate high-quality pseudo label, after high-quality pseudo label and real label are combined, participate in the hungarian matching of student model;Based on the prediction of teacher model, generate foreground mask and background mask, in decoder layer based on foreground mask and background mask, foreground guide feature distillation is calculated feature distillation loss;The output logic value of teacher model and student model is extracted, and only the channel feature corresponding to old class is selected.The application realizes that model can efficiently and stably maintain the recognition ability to learned class while learning new class, significantly alleviate the catastrophic forgetting.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an incremental target detection method, system and device based on curriculum-based learning distillation. Background Technology

[0002] In the field of computer vision, object detection technology is quite mature, capable of accurately locating and recognizing objects in images. However, the real world is dynamic and constantly evolving, with new objects constantly emerging. To enable models to continuously learn new categories without retraining on all old data, researchers have proposed class-incremental object detection. Its core challenge is overcoming "catastrophic forgetting," where models rapidly forget old knowledge while learning new knowledge.

[0003] To alleviate this problem, mainstream methods employ a knowledge distillation framework: in the incremental learning phase, a teacher model that has already learned old knowledge is fixed, and a student model is trained. By designing a distillation loss (such as feature alignment loss), the student model's internal representation is made similar to the teacher model when learning new data, thus preserving old knowledge. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a class-based incremental target detection method based on curriculum-based learning distillation to solve the comprehensive problems of pseudo-label noise interference, coarse constraints of new and old knowledge, and imbalance of learning dynamics in the training process of existing knowledge distillation frameworks.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a class incremental target detection method based on curriculum-based learning distillation, which includes acquiring an example set containing old class samples and new class training data in the current stage, and initializing a teacher model and a student model; The input image is fed into the teacher model, and the prediction results of the teacher model are subjected to Top-K screening and IOU conflict suppression to generate high-quality pseudo-labels. The high-quality pseudo-labels are merged with the real labels and then used for Hungarian matching in the student model. Foreground and background masks are generated based on the predictions of the teacher model. At the decoder layer, foreground-guided feature distillation is performed based on the foreground and background masks to calculate the feature distillation loss. Extract the output logic values ​​of the teacher model and the student model, and select only the channel features corresponding to the old category to calculate the KL divergence loss of the teacher model and the student model on the channel features corresponding to the old category; The weights of the basic detection loss, detector knowledge distillation loss, feature distillation loss, and KL divergence loss are dynamically adjusted according to the current training round, and the weighted basic detection loss, detector knowledge distillation loss, feature distillation loss, and KL divergence loss are summed to obtain the total loss. Based on the total loss, the student model is updated using backpropagation.

[0007] As a preferred embodiment of the incremental target detection method based on curriculum-based learning distillation described in this invention, the initialization of the teacher model and the student model specifically involves: the teacher model being the student model that has converged in the previous training phase, and its network parameters being frozen during the current training process; the structure of the student model being exactly the same as that of the teacher model, its network parameters being initialized to the parameters of the teacher model, and being in a trainable state.

[0008] As a preferred embodiment of the incremental target detection method based on curriculum-based learning distillation described in this invention, the specific process of performing Top-K screening and IOU conflict suppression on the prediction results of the teacher model to generate high-quality pseudo-labels includes: The prediction results of the teacher model are sorted in descending order of confidence score, and the top K prediction results are selected to form the initial screening sample set. Calculate the intersection-union ratio (IoU) between each predicted bounding box in the initial filtered sample set and all ground truth bounding boxes in the current input image; Remove boxes whose intersection-union ratio with any real bounding box is greater than a preset conflict threshold. The predicted bounding boxes are used to obtain a set of conflict-free pseudo-labels. The conflict-free pseudo-label set is merged with the real label set of the current input image to form an enhanced supervision label set for Hungarian matching in the student model. Its expression is: ; in, To enhance the supervision of the label set, Pseudo-labels generated for the teacher model A collection of real tags. The conflict threshold, Calculate the intersection-union ratio of the bounding box for the function.

[0009] As a preferred embodiment of the incremental target detection method based on curriculum-based learning distillation described in this invention, the process of generating foreground and background masks based on the teacher model specifically includes: From the prediction results of the teacher model, select the prediction boxes whose predicted scores are higher than the score threshold τ; The selected prediction boxes are subjected to intra-class nonmaximum suppression to obtain the final set of foreground targets; Based on the position of each bounding box in the foreground target set, the corresponding region on the feature map is marked as 1, and a foreground mask is generated. The remaining areas on the feature map are marked as 0 to generate a background mask. ,Right now =1- ; The foreground-guided feature distillation loss The calculation formula is: ; in, and Let represent the output feature maps of the teacher model and the student model at the decoder layer, respectively. Foreground mask, As background mask, Foreground-guided characteristic distillation loss, For the expectation in the spatial dimension, These are the background constraint weight coefficients.

[0010] As a preferred embodiment of the incremental target detection method based on curriculum-based learning distillation described in this invention, the specific process of calculating the KL divergence loss of the teacher model and the student model on the corresponding channel features of the old category includes: The output logic value from the teacher model and the output logic value of the student model In the process, the channel features corresponding to all old categories are extracted to obtain... and ; To each and conduct Normalization is performed to obtain the probability distribution over the old categories; Calculate the Kullback-Leibler divergence between the old class probability distribution of the teacher model and the old class probability distribution of the student model, and use it as the boundary distillation loss. The calculation formula is as follows: ; in, This refers to boundary distillation losses.

[0011] As a preferred embodiment of the incremental target detection method based on curriculum-based learning distillation described in this invention, the specific strategy for dynamically adjusting the weights of various losses according to the current training round is as follows: The total loss function is defined as:

[0012] in Based on loss detection, For detector knowledge distillation loss, and Characteristic distillation losses and boundary distillation loss Dynamic weighting coefficients; Set the starting round of course scheduling With the length of the weight growth interval ; In training rounds < At that time, total loss = ; In training rounds ≥ When needed, the weighting coefficients are dynamically adjusted according to the following formula. and : ; in, This is the preset maximum value for the weight of the corresponding loss term. The (x,0,1) function restricts the input value x to the interval [0,1].

[0013] As a preferred embodiment of the incremental target detection method based on curriculum-based learning distillation described in this invention, after backpropagation updating the student model based on the total loss, the method further includes model iteration and example set update. The student model that converges in the current training phase is used as the teacher model in the next training phase. From the new training data of the current stage, some samples are selected and added to the example set according to a preset strategy, and some old samples are removed from the example set according to the capacity limit, so as to update and maintain the example set for subsequent incremental learning.

[0014] Secondly, the present invention provides a class incremental target detection system based on curriculum-based learning distillation, including a data processing module for acquiring an example set containing old class samples and new class training data in the current stage, and initializing a teacher model and a student model; The collaborative cleanup module, connected to the data processing module, is used to feed the input image into the teacher model, perform Top-K filtering and IOU conflict suppression on the prediction results of the teacher model to generate high-quality pseudo-labels, and merge the high-quality pseudo-labels with the real labels to participate in the Hungarian matching of the student model; at the same time, it generates foreground masks and background masks based on the predictions of the teacher model, and performs foreground-guided feature distillation based on the foreground masks and background masks at the decoder layer to calculate the feature distillation loss; The boundary distillation module, connected to the data processing module, is used to extract the output logic values ​​of the teacher model and the student model, and select only the channel features corresponding to the old category to calculate the KL divergence loss of the teacher model and the student model on the channel features corresponding to the old category. The distillation strategy module, connected to the collaborative purification module and the boundary distillation module, is used to dynamically adjust the weights of the basic detection loss, detector knowledge distillation loss, feature distillation loss and KL divergence loss according to the current training round, and sum the weighted detector knowledge distillation loss, feature distillation loss and KL divergence loss to obtain the total loss; The model training module, connected to the distillation strategy module, is used to update the student model through backpropagation based on the total loss.

[0015] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the incremental target detection method based on curriculum-based learning distillation as described in the first aspect of the present invention.

[0016] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the incremental target detection method based on curriculum-based learning distillation as described in the first aspect of the present invention.

[0017] The beneficial effects of this invention are as follows: By constructing a noise-resistant collaborative cleanup mechanism, high-quality pseudo-labels are generated using Top-K screening and IOU conflict suppression, and combined with real labels to form an enhanced supervision set. Simultaneously, a foreground mask is introduced for foreground-guided feature distillation to suppress background noise. A knowledge-consolidation-based boundary distillation is designed, which precisely constrains and stabilizes the decision boundaries of old classes by separating old class channels and calculating KL divergence loss. Finally, this method employs a curriculum-based dynamic scheduling strategy. In the early stages of training, only basic detection loss and detector distillation loss are used to enable the model to quickly learn new classes and initially retain old classes. Subsequently, the weights of feature distillation and boundary distillation are dynamically and linearly increased to gradually strengthen the retention of old knowledge. This effectively solves the comprehensive problems of pseudo-label noise interference, coarse constraints between new and old knowledge, and dynamic imbalance in the learning of new and old classes in incremental class learning. Through cleanup-focused signal processing and a targeted-temporal knowledge consolidation mechanism, the model can efficiently and stably maintain the recognition ability of learned classes while continuously learning new classes, significantly mitigating catastrophic forgetting. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a class-based incremental target detection method based on curriculum-based learning distillation.

[0020] Figure 2 This is a schematic diagram of an incremental target detection system based on curriculum-based learning distillation. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figures 1-2 This is one embodiment of the present invention, which provides a class incremental target detection method based on curriculum-based learning distillation, including the following steps: S1. Initialize the teacher model and student model.

[0025] S1.1 The teacher model is the student model that has converged in the previous training phase, and its network parameters are frozen during this training process; the structure of the student model is exactly the same as that of the teacher model, its network parameters are initialized to the parameters of the teacher model, and it is in a trainable state.

[0026] Furthermore, at the beginning of incremental learning, a model carrying all learned category knowledge is necessary; this model is named the teacher model. The teacher model originates from the model that completed training and reached optimal performance in the previous incremental learning phase—essentially the student model that converged in the previous phase. During the current training process, all network layer parameters of the teacher model, including the weights and biases of convolutional layers and the parameters of fully connected layers, are set to a non-updateable state, i.e., frozen. Corresponding to the teacher model, a trainable model with an identical structure is constructed; this model is named the student model. The initial network parameters of the student model are obtained by directly copying all parameters from the teacher model, ensuring that the student model inherits the feature representations and knowledge learned by the teacher model from the very beginning of training. Throughout subsequent training iterations, only the parameters of the student model are iteratively updated using the backpropagation algorithm based on the calculated total loss. This setup guarantees a stable foundation for knowledge transfer from the teacher model to the student model. Thus, the roles and states of the teacher and student models are clearly established, providing a fixed source of knowledge and optimizable model entities for subsequent collaborative purification, boundary distillation, and curriculum-based scheduling.

[0027] S2. The prediction results of the teacher model are subjected to Top-K screening and IOU conflict suppression to generate high-quality pseudo-labels.

[0028] S2.1 Sort the prediction results of the teacher model in descending order of confidence score, and select the top K prediction results to form the initial screening sample set.

[0029] Next, the teacher model's prediction output needs to be processed. After the input image is fed into the teacher model, it generates a set of predictions containing bounding box coordinates, class confidence scores, and class labels. These predictions are then sorted in descending order of their confidence scores. After sorting, K predictions are selected from the top of the sorted list; these selected predictions constitute the initial filtered sample set. This operation filters out a large number of low-confidence, potentially unreliable prediction noises, resulting in a preliminary optimized candidate set.

[0030] S2.2 Calculate the intersection-union ratio (IUU) of each predicted bounding box in the initial filtered sample set with all ground truth bounding boxes in the current input image.

[0031] Furthermore, after obtaining the initial screening sample set, the next step is to perform geometric collision detection and screening. For each predicted bounding box in the initial screening sample set, the intersection-union ratio (IUR) between it and all ground truth bounding boxes in the current input image is calculated. The IUR is calculated using the standard formula, which is the ratio of the intersection area to the union area of ​​two bounding boxes. This calculation process generates one or more overlap metrics with ground truth bounding boxes for each predicted bounding box from the teacher model.

[0032] S2.3, Remove bounding boxes whose intersection-union ratio with any real bounding box is greater than a preset conflict threshold. The predicted bounding boxes are used to obtain a set of conflict-free pseudo-labels.

[0033] Furthermore, based on the calculated intersection-union ratio, conflict suppression operations are performed. A preset conflict threshold is set. This threshold is used to determine whether the predicted bounding box has an unacceptable geometrical conflict with the ground truth bounding boxes. It iterates through each predicted bounding box in the initial filtered sample set, checking if its intersection-union ratio (IU) with all ground truth bounding boxes exceeds the conflict threshold. This occurs when the intersection-union ratio (IoU) of a predicted bounding box with any ground truth bounding box is greater than the conflict threshold. If a bounding box is found to be redundant or incorrectly located, it is considered to be a redundant prediction of the real target and is removed from the set. After traversal and conditional judgment, the remaining bounding boxes form a set of conflict-free pseudo-labels, in which the bounding boxes have low spatial conflict with existing real labels.

[0034] S2.4. Merge the conflict-free pseudo-label set with the real label set of the current input image to form an enhanced supervision label set for Hungarian matching of the student model. Its expression is: ; in, To enhance the supervision of the label set, Pseudo-labels generated for the teacher model A collection of real tags. The conflict threshold, Calculate the intersection-union ratio of the bounding box for the function.

[0035] Furthermore, the purified teacher model predictions are fused with manually labeled real data. Specifically, the conflict-free pseudo-label set obtained in the previous step is joined with the original set of real labels attached to the current input image. The merged label set constitutes the enhanced supervision label set used to guide student model training, denoted as... The mathematical expression for this process is: Equal to the set of real labels With the condition " Less than The teacher model generates pseudo-labels. The union of . Among them, This represents the function used to calculate the intersection-union ratio. At this point, an enhanced supervision signal, combining high-quality real-world annotations with rigorously selected, high-confidence, and position-conflict-free pseudo-labels from the teacher model, is ready for use in the Hungarian matching algorithm for the student model.

[0036] S3. The foreground mask and background mask are generated based on the teacher model prediction.

[0037] S3.1 From the prediction results of the teacher model, select the prediction boxes whose predicted scores are higher than the score threshold τ.

[0038] Furthermore, generating the mask required for foreground-guided feature distillation and calculating the loss necessitates locating key target regions from the teacher model's predictions. First, the teacher model's output predictions are processed, containing multiple bounding boxes with prediction scores. A score threshold τ is set to distinguish between high-confidence and low-confidence predictions. All bounding boxes in the teacher model's predictions are iterated through, filtering out those with prediction scores strictly higher than the score threshold τ. These filtered bounding boxes are considered potential high-quality foreground target candidates, providing a foundational set for subsequent refinement.

[0039] S3.2 Perform intra-class nonmaximum suppression on the selected prediction boxes to obtain the final set of foreground targets.

[0040] Furthermore, redundancy removal is performed on the high-scoring predicted bounding boxes selected above. Since multiple highly overlapping bounding boxes may point to the same real-world object in the image, intra-class non-maximum suppression (NMS) is required. The core of this operation is to sort the bounding boxes according to their prediction scores within the same prediction class and calculate the intersection-union (IU) ratio between any two bounding boxes. For any pair of bounding boxes, if their IU ratio exceeds a preset overlap threshold, the bounding box with the higher prediction score is retained, while the bounding box with the lower score is suppressed. After traversing all classes and completing this operation, a set of bounding boxes with redundant boxes removed, where each instance theoretically corresponds to only one most representative bounding box, is obtained. This set is defined as the final foreground object set.

[0041] S3.3. Based on the position of each target box in the foreground target set, mark the corresponding region as 1 on the feature map to generate a foreground mask. The remaining areas on the feature map are marked as 0 to generate a background mask. ,Right now =1- .

[0042] Furthermore, based on the determined set of foreground objects, a binary mask for distinguishing the foreground from the background is generated at the corresponding spatial location of the feature map. Feature maps with the same spatial dimensions, output by the teacher and student models at the decoder layer, are obtained. According to the coordinates of each bounding box in the final set of foreground objects, the mask value is marked as the number 1 in the corresponding spatial region of the feature map. This binary map, covering all foreground object regions, is defined as the foreground mask. Accordingly, in all regions of the feature map not covered by the foreground object, the mask value is marked as the number 0, and this binary map, complementary to the foreground mask, is defined as the background mask. Background mask By adjusting the foreground mask The relationship is obtained by performing a point-by-point calculation of "1 minus the corresponding position value". =1- At this point, the spatial attention mask used to guide feature distillation is ready.

[0043] S3.4, the foreground guiding feature distillation loss The calculation formula is: ; in, and Let represent the output feature maps of the teacher model and the student model at the decoder layer, respectively. Foreground mask, As background mask, Foreground-guided characteristic distillation loss, For the expectation in the spatial dimension, These are the background constraint weight coefficients.

[0044] Furthermore, using the generated foreground and background masks, as well as the teacher and student model features, the foreground-guided feature distillation loss is calculated. The point-by-point differences between the teacher and student model features at all spatial locations are calculated, and the square of their L2 norms is taken. For the foreground region, this squared difference map is multiplied point-by-point with the foreground mask, and the expected value is calculated in the spatial dimension, forming a strong constraint term for the foreground features. For the background region, this squared difference map is multiplied point-by-point with the background mask, and the expected value is calculated in the spatial dimension, then multiplied by a background constraint weight coefficient between 0 and 1, forming a weak constraint term for the background features. Adding the strong and weak constraint terms yields the complete foreground-guided feature distillation loss. This loss function aims to force the student model to closely mimic the feature representation of the teacher model in key foreground target regions, while allowing certain feature differences in non-key background regions, thus achieving selective and focused distillation of feature knowledge.

[0045] S4. The KL divergence loss of the teacher model and the student model on the corresponding channel features of the old category is calculated.

[0046] S4.1, From the output logic value of the teacher model and the output logic value of the student model In the process, the channel features corresponding to all old categories are extracted to obtain... and .

[0047] Furthermore, calculating the boundary distillation loss involves extracting, transforming, and measuring the differences between the teacher and student models' outputs on the old category channels. First, it's necessary to separate the parts related to old knowledge from the final outputs of both models. After processing the input image, the teacher model outputs a tensor representing the raw scores or logistic values ​​for all categories; this tensor is named the teacher model's output logistic value, and its dimension is typically the batch size multiplied by the total number of categories. The student model, after the same forward propagation, also produces a tensor with the exact same structure, named the student model's output logistic value. To perform knowledge distillation for old categories, the channel features corresponding to all old categories must be extracted from the teacher model's output logistic value, based on a predefined list of old category indices; the resulting tensor is named Zolds. Similarly, the corresponding channel features are extracted from the student model's output logistic value, based on the exact same list of old category indices; the resulting tensor is named Zolds. This operation ensures that subsequent comparisons are limited to the model's judgments on previously learned categories and are unrelated to the new categories being learned.

[0048] S4.2, respectively for and conduct Normalization is performed to obtain the probability distribution over the old categories.

[0049] Furthermore, in obtaining and After obtaining these two tensors containing only old category channel features, the next step is to transform these raw logistic values ​​into probability distributions. Specifically, for... and application Normalization process. The function performs calculations along the category channel dimension, mapping the original logistic value of each sample across its old categories to a set of probability values, the sum of which is 1. conduct The processing yields a result representing the teacher model's predicted probability distribution for possible old-class targets in the input image. Perform the same The processing yields a probability distribution representing the student model's predictions for the same old class of targets. At this point, the decision outputs of the teacher and student models on the old class have been transformed into comparable probabilistic forms.

[0050] S4.3 Calculate the Kullback-Leibler divergence between the old class probability distribution of the teacher model and the old class probability distribution of the student model, as the boundary distillation loss. The calculation formula is as follows: ; in, This refers to boundary distillation losses.

[0051] Furthermore, the boundary distillation loss LKL is calculated based on the two probability distributions mentioned above. Specifically, the Kullback-Leibler divergence is used to measure the difference between the old class probability distribution of the teacher model and the old class probability distribution of the student model. The Kullback-Leibler divergence is an asymmetric measure of the difference between one probability distribution and another reference probability distribution. Its input is... Old class channel features of the teacher model after normalization The probability distribution it represents, and after Normalized student model old class channel features The probability distribution represented by the given value. The calculated scalar value is the boundary distillation loss LKL. The core purpose of this loss function is to minimize the difference between the output probability distribution of the student model on the old categories and the corresponding distribution of the teacher model, thereby forcing the student model to maintain a high degree of consistency with the decision boundaries of the learned old categories when incrementally learning new categories, thus achieving targeted consolidation of knowledge.

[0052] S5. The weights of each loss are dynamically adjusted according to the current training round.

[0053] S5.1 The total loss function is defined as follows:

[0054] in Based on loss detection, For detector knowledge distillation loss, and Characteristic distillation losses and boundary distillation loss Dynamic weighting coefficients; Furthermore, to achieve course-based dynamic weight scheduling, it is first necessary to clarify the overall composition of the various losses during training and the adjustment objectives of the weight parameters. After the forward propagation of each training iteration, four core losses are calculated: the basic detection loss for detecting targets, the detector knowledge distillation loss for supervising the overall detection performance of the model, the foreground-guided feature distillation loss for selectively transferring feature space knowledge, and the boundary distillation loss for targeted consolidation of old class knowledge. The total loss function is defined as the weighted sum of these four losses. In this expression, and It determines the characteristic distillation loss and boundary distillation loss The dynamic weight coefficients that contribute to the intensity of the total loss are not fixed in specific values, but are dynamically adjusted according to the training process. The weight of the detector knowledge distillation loss is 1 by default and usually does not participate in dynamic scheduling. The loss calculation framework is built on this basis.

[0055] S5.2 Setting the starting round of course scheduling With the length of the weight growth interval .

[0056] Furthermore, in order to perform dynamic scheduling, the start and step parameters of course learning need to be preset. Define the starting round of course scheduling. This is an integer value representing the training epoch. (In the training epoch counter) The value is less than The entire phase is considered the initial stage of the model's rapid learning of new categories. Simultaneously, a weight growth interval length *r* is defined, a positive number used to control the rate at which the weight coefficients grow from 0 to their preset maximum value, or the number of rounds required. Parameters and Together, they form the time base for the scheduling strategy, ensuring synchronization between weight changes and the training process. S5.3, in training rounds < At that time, total loss = .

[0057] Furthermore, in training rounds Less than the starting round In this phase, the initial scheduling strategy is executed. During this phase, the contributions of feature distillation loss and boundary distillation loss are completely removed when calculating the total loss L. Therefore, the total loss function is... = The goal at this stage is to have the student model in the lead. In round training, the main basis is Based on detection loss and detector knowledge distillation loss (Its core supervisory signal comes from the enhanced supervisory label set after merging real labels and high-quality pseudo labels) to learn, so that it can preferentially and quickly absorb and build new category feature representations and detection capabilities from new data without being constrained by the distillation of structured knowledge such as features and boundaries.

[0058] S5.4, in training rounds ≥ When needed, the weighting coefficients are dynamically adjusted according to the following formula. and .

[0059] ; in, This is the preset maximum value for the weight of the corresponding loss term. The (x,0,1) function restricts the input value x to the interval [0,1].

[0060] Furthermore, when training rounds Reaching or exceeding the starting round At this point, the system enters the dynamic weight enhancement phase. During this phase, the dynamic weight coefficient... and The value is no longer 0, but begins to increase over time according to a preset formula. For each weight coefficient that needs to be dynamically adjusted... molecular( The number of rounds that have occurred since the scheduling began was calculated. (Using () Divide by the length of the weight growth interval This yields a proportional value representing the growth process. The function acts on this ratio, restricting it to a closed interval between 0 and 1, ensuring the numerical stability of the calculation process and defining the saturation point for weight growth. Ultimately, the current weight coefficient ω equals its maximum value. The product of this cropped ratio. Using this formula, in ≥ In subsequent training, and The value will start from 0 and increase with each round. The growth rate increases linearly with the increase of r (the growth slope is controlled by r), until they each approach their respective maximum values. and This dynamic adjustment method allows the influence of feature distillation loss and boundary distillation loss on model optimization to gradually and smoothly increase as training progresses, thereby gradually strengthening the consolidation of old knowledge in the later stages of training and realizing a course-based learning transition from "focusing on learning new things" to "balancing new and old knowledge".

[0061] S6. After backpropagating and updating the student model based on the total loss, model iteration and example set update are also included.

[0062] S6.1. The student model that has converged in the current training phase shall be used as the teacher model in the next training phase.

[0063] Furthermore, after completing the current training phase, the model and data need to be iteratively updated to prepare for the next incremental learning phase. First, the model's role is transformed. When the student model finishes training in this incremental phase—that is, when the student model's detection accuracy on the validation set stabilizes or reaches the preset training epoch limit—it means the current student model has converged. At this point, the overall structure and all weight parameters of this student model, which has completed the current phase of learning and whose parameters have been optimized, are completely saved. At the start of the next incremental learning phase, this saved model will no longer be used as the training object but will be directly loaded and given a new role: the teacher model for the next incremental learning phase. This new teacher model will carry knowledge of all historical categories, including the new learning categories of the current phase, and will keep its parameters frozen in the next training cycle to guide the student model in learning updated knowledge categories in the next phase. Thus, the knowledge transfer chain of model iteration continues.

[0064] S6.2 From the new training data of the current stage, select some samples according to the preset strategy to add to the example set, and remove some old samples from the example set according to the capacity limit, so as to update and maintain the example set for subsequent incremental learning.

[0065] Furthermore, after completing the model role transition, the next step is to update the example set to mitigate catastrophic forgetting. The example set is a data buffer storing a small number of representative samples from each previously learned class, with a fixed total capacity. The update operation involves two parallel steps: sample addition and sample removal. The sample addition step involves selecting a subset of samples from the newly trained data used in the current stage of training, based on a pre-defined strategy. This strategy can be random sampling or selection based on some importance metric (such as sample diversity, difficulty for model decision-making, etc.). These selected new class samples, along with their corresponding true labels, are added to the example set. The sample removal step involves checking whether the total data volume of the example set exceeds its pre-defined fixed capacity limit after adding new samples. If it does, a subset of older samples needs to be removed from the example set according to a pre-defined replacement strategy to free up space. The removal strategy can also be random removal or selection based on sample retention time, class representativeness, etc. By performing addition and removal operations, the sample set is updated: it includes new category samples from the latest stage, allowing for the replay of new knowledge in subsequent learning, while maintaining overall representativeness within capacity constraints by removing some older samples. The updated sample set will be used in the next incremental learning stage along with new training data to help the new student model review and consolidate all learned knowledge, including new categories. At this point, the dynamic data buffer supporting continuous learning is maintained.

[0066] This embodiment also provides a class incremental target detection system based on curriculum-based learning distillation, including: a data processing module, used to acquire an example set containing old class samples and new class training data in the current stage, and initialize the teacher model and student model; The collaborative cleanup module, connected to the data processing module, is used to feed the input image into the teacher model, perform Top-K filtering and IOU conflict suppression on the prediction results of the teacher model to generate high-quality pseudo-labels, and merge the high-quality pseudo-labels with the real labels to participate in the Hungarian matching of the student model; at the same time, it generates foreground masks and background masks based on the predictions of the teacher model, and performs foreground-guided feature distillation based on the foreground masks and background masks at the decoder layer to calculate the feature distillation loss; The boundary distillation module, connected to the data processing module, is used to extract the output logic values ​​of the teacher model and the student model, and select only the channel features corresponding to the old category to calculate the KL divergence loss of the teacher model and the student model on the channel features corresponding to the old category. The distillation strategy module, connected to the collaborative purification module and the boundary distillation module, is used to dynamically adjust the weights of the basic detection loss, detector knowledge distillation loss, feature distillation loss and KL divergence loss according to the current training round, and sum the weighted detector knowledge distillation loss, feature distillation loss and KL divergence loss to obtain the total loss; The model training module, connected to the distillation strategy module, is used to update the student model through backpropagation based on the total loss.

[0067] This embodiment also provides a computer device applicable to the incremental target detection method based on curriculum-based learning distillation, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the incremental target detection method based on curriculum-based learning distillation as proposed in the above embodiment.

[0068] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0069] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the incremental target detection method based on curriculum-based learning distillation as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0070] In summary, this invention constructs a noise-resistant collaborative cleanup mechanism, utilizes Top-K screening and IOU conflict suppression to generate high-quality pseudo-labels, and combines them with real labels to form an enhanced supervision set. Simultaneously, it introduces foreground masks for foreground-guided feature distillation to suppress background noise. A knowledge-consolidation-based boundary distillation is designed, precisely constraining and solidifying the decision boundaries of old classes by separating old class channels and calculating KL divergence loss. Finally, this method employs a curriculum-based dynamic scheduling strategy. In the early stages of training, only basic detection loss and detector distillation loss are used to enable the model to quickly learn new classes and initially retain old classes. Subsequently, the weights of feature distillation and boundary distillation are dynamically and linearly increased to gradually strengthen the retention of old knowledge. This effectively solves the comprehensive problems of pseudo-label noise interference, coarse constraints between new and old knowledge, and dynamic imbalance between new and old class learning in incremental class learning. Through cleanup-focused signal processing and a targeted-temporal knowledge consolidation mechanism, the model can efficiently and stably maintain the recognition ability of learned classes while continuously learning new classes, significantly mitigating catastrophic forgetting.

[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for incremental target detection based on curriculum-based learning distillation, characterized in that: include, Obtain the example set containing old class samples and the new class training data for the current stage, and initialize the teacher model and student model; The input image is fed into the teacher model, and the prediction results of the teacher model are subjected to Top-K screening and IOU conflict suppression to generate high-quality pseudo-labels. The high-quality pseudo-labels are then merged with the real labels and used in the Hungarian matching of the student model. Based on the teacher model, foreground and background masks are generated. At the decoder layer, foreground-guided feature distillation is performed based on the foreground and background masks to calculate the feature distillation loss. Extract the output logic values ​​of the teacher model and the student model, and select only the channel features corresponding to the old category to calculate the KL divergence loss of the teacher model and the student model on the channel features corresponding to the old category; The weights of the basic detection loss, detector knowledge distillation loss, feature distillation loss, and KL divergence loss are dynamically adjusted according to the current training round, and the weighted detector knowledge distillation loss, feature distillation loss, and KL divergence loss are summed to obtain the total loss. Based on the total loss, the student model is updated using backpropagation.

2. The incremental target detection method based on curriculum-based learning distillation as described in claim 1, characterized in that: The initialization of the teacher model and student model specifically involves the following: the teacher model is the student model that has converged in the previous training phase, and its network parameters are frozen during this training process; the structure of the student model is exactly the same as that of the teacher model, its network parameters are initialized to the parameters of the teacher model, and it is in a trainable state.

3. The incremental target detection method based on curriculum-based learning distillation as described in claim 2, characterized in that: The process of performing Top-K filtering and IOU conflict suppression on the prediction results of the teacher model to generate high-quality pseudo-labels includes: The prediction results of the teacher model are sorted in descending order of confidence score, and the top K prediction results are selected to form the initial screening sample set. Calculate the intersection-union ratio (IoU) between each predicted bounding box in the initial filtered sample set and all ground truth bounding boxes in the current input image; Remove boxes whose intersection-union ratio with any real bounding box is greater than a preset conflict threshold. The predicted bounding boxes are used to obtain a set of conflict-free pseudo-labels. The conflict-free pseudo-label set is merged with the real label set of the current input image to form an enhanced supervision label set for Hungarian matching in the student model. Its expression is: ; in, To enhance the supervision of the label set, Pseudo-labels generated for the teacher model A collection of real tags. The conflict threshold, Calculate the intersection-union ratio of the bounding box for the function.

4. The incremental target detection method based on curriculum-based learning distillation as described in claim 3, characterized in that: The process of generating foreground and background masks based on the teacher model includes: From the prediction results of the teacher model, select the prediction boxes whose predicted scores are higher than the score threshold τ; The selected prediction boxes are subjected to intra-class nonmaximum suppression to obtain the final set of foreground targets; Based on the position of each bounding box in the foreground target set, the corresponding region on the feature map is marked as 1, and a foreground mask is generated. The remaining areas on the feature map are marked as 0 to generate a background mask. ,Right now =1- ; The foreground-guided feature distillation loss The calculation formula is: ； in, and Let represent the output feature maps of the teacher model and the student model at the decoder layer, respectively. Foreground mask, As background mask, Foreground-guided characteristic distillation loss, For the expectation in the spatial dimension, These are the background constraint weight coefficients.

5. The incremental target detection method based on curriculum-based learning distillation as described in claim 4, characterized in that: The specific process for calculating the KL divergence loss of the teacher model and the student model on the corresponding channel features of the old categories includes: The output logic value from the teacher model and the output logic value of the student model In the process, the channel features corresponding to all old categories are extracted to obtain... and ; To each and conduct Normalization is performed to obtain the probability distribution over the old categories; Calculate the Kullback-Leibler divergence between the old class probability distribution of the teacher model and the old class probability distribution of the student model, and use it as the boundary distillation loss. The calculation formula is as follows: ； in, This refers to boundary distillation losses.

6. The incremental target detection method based on curriculum-based learning distillation as described in claim 5, characterized in that: The specific strategy for dynamically adjusting the weights of various losses based on the current training round is as follows: The total loss function is defined as: in Based on loss detection, For detector knowledge distillation loss, and Characteristic distillation losses and boundary distillation loss Dynamic weighting coefficients; Set the starting round of course scheduling With the length of the weight growth interval ; In training rounds < At that time, total loss = ; In training rounds ≥ When needed, the weighting coefficients are dynamically adjusted according to the following formula. and : ； in, This is the preset maximum value for the weight of the corresponding loss term. The (x,0,1) function restricts the input value x to the interval [0,1].

7. The incremental target detection method based on curriculum-based learning distillation as described in claim 6, characterized in that: Following the backpropagation update of the student model based on the total loss, the process also includes model iteration and example set update: The student model that converges in the current training phase is used as the teacher model in the next training phase. From the new training data of the current stage, some samples are selected and added to the example set according to a preset strategy, and some old samples are removed from the example set according to the capacity limit, so as to update and maintain the example set for subsequent incremental learning.

8. A class incremental target detection system based on curriculum-based learning distillation, based on the class incremental target detection method based on curriculum-based learning distillation as described in any one of claims 1 to 7, characterized in that: This includes a data processing module, used to acquire an example set containing old class samples and new class training data for the current stage, and to initialize the teacher model and student model; The collaborative purification module, connected to the data processing module, is used to send the input image into the teacher model, perform Top-K filtering and IOU conflict suppression on the prediction results of the teacher model to generate high-quality pseudo-labels, and merge the high-quality pseudo-labels with the real labels to participate in the Hungarian matching of the student model. Simultaneously, foreground and background masks are generated based on the predictions of the teacher model, and foreground-guided feature distillation is performed at the decoder layer based on the foreground and background masks to calculate the feature distillation loss; The boundary distillation module, connected to the data processing module, is used to extract the output logic values ​​of the teacher model and the student model, and select only the channel features corresponding to the old category to calculate the KL divergence loss of the teacher model and the student model on the channel features corresponding to the old category. The distillation strategy module, connected to the collaborative purification module and the boundary distillation module, is used to dynamically adjust the weights of the basic detection loss, detector knowledge distillation loss, feature distillation loss and KL divergence loss according to the current training round, and sum the weighted basic detection loss, detector knowledge distillation loss, feature distillation loss and KL divergence loss to obtain the total loss. The model training module, connected to the distillation strategy module, is used to update the student model through backpropagation based on the total loss.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the incremental target detection method based on curriculum-based learning distillation as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the incremental target detection method based on curriculum-based learning distillation as described in any one of claims 1 to 7.

Images