Fully supervised target detection model construction method and device

Abstract

The present application relates to a method and apparatus for constructing a fully supervised target detection model. Based on a fully supervised dataset, the method significantly improves the training efficiency and detection performance of the model in this field and reduces the resources required for training. In addition, the method uses a recursive reused teacher model for self-training, associates contextual information, and provides high-quality pseudo-labels for the student model. Furthermore, the Transformer module helps the model improve learning stability by adding additional box regression tasks.

Description

Technical Field

[0001] The present application relates to the field of computer vision technology, and in particular to a method and device for constructing a fully supervised target detection model. Background Art

[0002] With the rapid development of computer vision technology, object detection has a wide range of applications in traffic management, monitoring systems, medicine and other fields. Most of the current detectors have achieved success on datasets such as MS-COCO and PASCAL VOC, which relies on large-scale standard detection annotations (categories and bounding boxes). At this stage, with the increasing demand for more powerful and accurate detection models, the need to collect and label more data is also increasing. However, manually annotating class labels and bounding boxes for large-scale datasets is a very expensive and tedious task, which is not cost-effective in practical applications. As a remedy, weakly supervised (WSOD), semi-supervised (SSOD) and fully supervised object detection (OSOD) algorithms have received widespread attention in recent years.

[0003] SSOD enhances model performance by leveraging additional unannotated data, while WSOD utilizes a proprietary network designed for image-level annotation. Compared to SSOD, OSOD leverages instance information from additional weak labels, improving pseudo-label quality and reducing sample noise. In contrast to WSOD, OSOD is not limited to a single weak label format but accommodates nearly all weak labels, such as tags, points, and boxes. However, OSOD suffers from long training times and poor detection accuracy. Summary of the Invention

[0004] In order to overcome at least one of the shortcomings of the prior art, the present application provides a method and apparatus for constructing a fully supervised target detection model.

[0005] In a first aspect, a method for constructing a fully supervised target detection model is provided, comprising:

[0006] Step 1: Build a target detection dataset. The target detection dataset includes a standard annotation dataset and a weak annotation dataset. The weak annotation dataset includes multiple sample weak annotation types.

[0007] Step 2: Train the teacher model based on the standard labeled dataset to obtain the trained teacher model;

[0008] Step 3: Input the weakly labeled dataset into the trained teacher model, and the trained teacher model becomes the current teacher model to obtain the prediction result;

[0009] Step 4: Filter the prediction results by confidence threshold and weak label information to generate pseudo labels, and use the pseudo labels to update the weak label dataset to obtain an updated weak label dataset;

[0010] Step 5: Train the student model based on the standard labeled dataset and the updated weakly labeled dataset to obtain a trained student model;

[0011] Step 6: Update the weight parameters of the teacher model according to the weight parameters of the trained student model to obtain an updated teacher model;

[0012] Step 7: Determine whether the conditions for the end of iterative training are met. If so, output the updated teacher model, which is the final target detection model. If not, use the updated teacher model as the current teacher model and return to step 3.

[0013] In one embodiment, the teacher network and the student network have the same structure. The teacher network includes a residual network and a Transformer module connected in sequence. The residual network includes multiple residual blocks, each residual block includes multiple convolutional layers, and the input data outputs a tensor after passing through the residual block; the Transformer module includes an encoder and a decoder, and the decoder includes a Querises module and a Learnable Anchors module. The Querises module is used to obtain the box offset and category, and the Learnable Anchors module is used to obtain the updated learnable box. The box offset and the updated learnable box are added to obtain the predicted box output by the decoder.

[0014] In one embodiment, the residual block specifically implements the function, which is expressed by the following formula:

[0015] I l+1 = l +[θ l+1 *(I l · l )]

[0016] k l+1 =[C2*[C1*[θ l+1 *(I l · l )]+k l ]]

[0017] Among them, I l+1 is the feature of the l+1th convolutional layer in the residual block, I l is the feature of the lth convolutional layer in the residual block, θ l+1 is the convolution parameter of the l+1th convolution layer in the residual block, k l is the hidden tensor of the lth convolutional layer in the residual block, k l+1is the hidden tensor of the l+1th convolutional layer in the residual block, C2 and C1 are 1×1 and 3×3 convolutions respectively, * represents the convolution operation, and · represents the concatenation operation.

[0018] In one embodiment, the types of weak annotations of samples in the weakly labeled dataset include at least two of: no annotation, category annotation, annotation of the number of each category, point annotation, point annotation and category annotation, standard box annotation, and noise box annotation.

[0019] In one embodiment, the weight parameters of the teacher model are updated according to the weight parameters of the trained student model using the following formula:

[0020]

[0021] Among them, θ e is the weight parameter of the teacher model updated in the e-th iteration training, μ is the coefficient, θ e-1 The weight parameters of the teacher model updated for the e-1th iteration training, is the weight parameter of the trained student model obtained after the e-th iteration training.

[0022] In one embodiment, the condition for the end of iterative training is that the loss of the updated teacher model tends to be stable.

[0023] In one embodiment, the total loss function used during the training of the teacher model and the student model is:

[0024] L=σL match +γL known

[0025] Among them, L is the total loss function, L match is the matching loss, L known is the known loss, σ is the weight of the matching loss, and γ is the weight of the known loss;

[0026] Known loss L known for:

[0027]

[0028] Among them, i is the i-th prediction result, M is the number of prediction results, c i is the prediction category corresponding to the i-th prediction result, c i The corresponding predicted probability, b i is the prediction box corresponding to the i-th prediction result, b(i) is the real box, L box It is L1 loss or L2 loss.

[0029] In a second aspect, a fully supervised target detection model construction device is provided, comprising:

[0030] The dataset construction module is used to construct the target detection dataset, which includes a standard annotation dataset and a weak annotation dataset; the weak annotation dataset includes a variety of sample weak annotation types;

[0031] The first training module is used to train the teacher model based on the standard labeled data set to obtain the trained teacher model;

[0032] The prediction module is used to input the weakly labeled dataset into the trained teacher model, which is the current teacher model, to obtain the prediction results;

[0033] The weakly labeled dataset update module is used to filter the prediction results by confidence threshold and weakly labeled information, generate pseudo labels, and use the pseudo labels to update the weakly labeled dataset to obtain an updated weakly labeled dataset;

[0034] The second training module is used to train the student model based on the standard labeled dataset and the updated weakly labeled dataset to obtain a trained student model;

[0035] The weight parameter updating module is used to update the weight parameters of the teacher model according to the weight parameters of the trained student model to obtain an updated teacher model;

[0036] The judgment module is used to determine whether the conditions for the end of iterative training have been met. If so, the updated teacher model is output as the final target detection model. If not, the updated teacher model is used as the current teacher model and enters the prediction module.

[0037] In a third aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned method for constructing a fully supervised target detection model.

[0038] In a fourth aspect, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the above-mentioned method for constructing a fully supervised target detection model.

[0039] Compared with the existing technology, the present application has the following beneficial effects: the model construction method of the present application, based on a fully supervised dataset, greatly improves the training efficiency and detection performance of the model in this field, and reduces the resources required for training; in addition, the recursive reuse of the teacher model is used for self-training, and contextual information is associated to provide high-quality pseudo labels for the student model; and the Transformer module helps the model improve learning stability by adding additional box regression tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present application may be better understood by referring to the following description in conjunction with the accompanying drawings, which together with the following detailed description are incorporated into and form a part of this specification. In the drawings:

[0041] Figure 1 A flowchart of a method for constructing a fully supervised object detection model according to an embodiment of the present application is shown;

[0042] Figure 2 A schematic diagram of a teacher model and a student model is shown;

[0043] Figure 3 Schematic diagram of the structure of the residual block is shown;

[0044] Figure 4 shows the structural diagram of the Transformer module;

[0045] Figure 5 A comparison chart of target detection results of the model constructed using the method of the present application and the existing model Omni-DETR is shown. DETAILED DESCRIPTION

[0046] Exemplary embodiments of the present application are described below with reference to the accompanying drawings. For the sake of clarity and conciseness, not all features of actual embodiments are described in this specification. However, it should be understood that in the process of developing any such actual embodiment, many implementation-specific decisions may be made to achieve the developer's specific goals, and these decisions may vary from one implementation to another.

[0047] It is also necessary to explain here that, in order to avoid obscuring the present application due to unnecessary details, the accompanying drawings only show the device structure closely related to the solution according to the present application, while other details that are not closely related to the present application are omitted.

[0048] It should be understood that the present application is not limited to the described embodiments due to the following description with reference to the accompanying drawings. In this document, where feasible, the embodiments may be combined with each other, features between different embodiments may be replaced or borrowed, and one or more features may be omitted in one embodiment.

[0049] This application embodiment provides a method for constructing a fully supervised target detection model. Figure 1 A flowchart of a method for constructing a fully supervised target detection model according to an embodiment of the present application is shown. Figure 1 , methods include:

[0050] Step 1: Build a target detection dataset, which includes a standard annotation dataset and a weak annotation dataset.

[0051] Here, the target detection dataset is represented by D, and the standard annotation dataset is represented by D L Indicates that the weakly labeled dataset is represented by D w Indicates that the standard annotation dataset D L The dataset D includes multiple samples, where the sample can be an image with multiple targets. Each sample is annotated with a standard annotation, including a standard box and a category annotation. w It also includes multiple samples, each of which is weakly labeled. The types of weak labels can include 7 types, namely: the first type: no labeling, the second type: category labeling, the third type: labeling the number of each category, the fourth type: point labeling, the fifth type: point labeling and category labeling, the sixth type: standard box labeling, and the seventh type: noise box labeling.

[0052] Step 2: Train the teacher model based on the standard labeled dataset to obtain the trained teacher model. Figure 2 Figure 2 shows a schematic diagram of the teacher model. Here, the teacher model is trained by optimizing the loss function.

[0053] Step 3: Input the weakly labeled dataset into the trained teacher model, which becomes the current teacher model, and obtain the prediction results. The prediction results here include the predicted box and category.

[0054] Step 4: Filter the prediction results by confidence threshold and weak label information, generate pseudo labels, and use the pseudo labels to update the weak label dataset D w , and get the updated weakly labeled dataset. Here, confidence threshold filtering sets a confidence threshold. When the confidence corresponding to the prediction result is greater than the confidence threshold, the prediction result is retained; otherwise, the prediction result is deleted.

[0055] Here, for the prediction results obtained for samples of different weakly labeled types, the specific implementation method of weakly labeled information filtering is as follows:

[0056] If the sample's annotation type is the first type: no annotation, the confidence threshold filtering result shall prevail.

[0057] If the sample's annotation type is the second type: category annotation, then based on the weakly labeled category, the prediction results that do not belong to the weakly labeled category are removed; and based on the confidence of the retained prediction results, at least one instance (target) is assigned to the unpredicted category in the weakly labeled category.

[0058] If the sample's annotation type is the third type: the number of each category is marked, then the prediction results that do not belong to the weakly labeled category will be removed according to the weakly labeled category; since the category and the corresponding number are known, the corresponding number of instances are assigned to each category according to the confidence level.

[0059] If the annotation type of the sample is the fourth type: point annotation, then according to the point position information of the point annotation, the prediction results whose prediction boxes are outside the point position information of the point annotation are removed.

[0060] If the annotation type of the sample is the fifth type: point annotation and category annotation, then according to the point position information of the point annotation, the prediction results whose prediction boxes are outside the point position information of the point annotation are removed; and the prediction results whose prediction categories do not match the category annotations corresponding to the point annotations are removed.

[0061] If the sample's annotation type is the sixth type: standard box annotation, after filtering with the confidence threshold, for each instance of the standard box annotation, the intersection-over-union (IoU) between all predicted boxes and the annotation box is calculated, the IoUs are sorted, and the prediction result with the largest IoU is selected as the pseudo label for the instance.

[0062] If the sample's annotation type is the seventh type: noise box annotation, the processing method is the same as the sixth type.

[0063] Step 5: Based on the standard annotation dataset D L And the updated weakly labeled dataset are used to train the student model to obtain the trained student model.

[0064] Here, we can perform data enhancement on the samples in the updated weakly labeled dataset to obtain an expanded weakly labeled dataset. Expanding weakly labeled datasets The sample annotations in the dataset have also been adjusted accordingly with data augmentation. Specifically, data augmentation can include occlusion, translation, rotation, flipping, color conversion, color dithering, and / or Gaussian blurring. Occlusion fills randomly selected boxes of a certain size on the image with solid color blocks to act as a shield; translation refers to shifting the entire image a certain distance in all directions; rotation refers to rotating the image clockwise or counterclockwise around the center of the image; flipping refers to flipping the image around the vertical or horizontal axis of the center; color conversion refers to converting the RGB image of the dataset into an HSV or grayscale image; color dithering creates a color crossover effect of adjacent point differences by shifting the hue of the image; and Gaussian blurring refers to convolving the image with a normal distribution to blur the image, which can be regarded as a low-pass filter.

[0065] When training the student model, a certain proportion of the standard annotation dataset D can be selected in each batch. LSamples and expanded weakly labeled datasets The samples are mixed to train the model; the student model is trained by optimizing the loss function, and consistent regularization is used to expand the weak label data set. The loss function generated during training is used as a constraint, and training is iterated for multiple rounds until the model loss function converges to a stable result.

[0066] Step 6: Update the weight parameters of the teacher model according to the weight parameters of the trained student model to obtain an updated teacher model;

[0067] Step 7: Determine whether the conditions for the end of iterative training are met. If so, output the updated teacher model, which is the final target detection model. If not, use the updated teacher model as the current teacher model and return to step 3.

[0068] Here, the condition for the end of iterative training is that the loss of the updated teacher model tends to be stable.

[0069] The above embodiment uses a teacher model for self-training to provide high-quality pseudo labels for the student model. Based on a fully supervised dataset, it greatly improves the training efficiency and detection performance of the model in this field and reduces the resources required for training.

[0070] In one embodiment, the teacher network and the student network have the same structure. The teacher network includes a residual network ResNet50 and a Transformer module connected in sequence. The residual network includes multiple residual blocks. The residual network is a feature extraction network. Each residual block includes multiple convolutional layers. The input data outputs a tensor after passing through the residual block. Figure 3 The figure shows the structure of the residual block, which realizes recursive reuse of features by recursively aggregating and transferring feature information of each layer. The Transformer module includes an encoder and a decoder. The decoder includes a querise module and a learnable anchor module. The querise module is used to obtain the box offset and category, and the learnable anchor module is used to obtain the updated learnable box. The box offset and the updated learnable box are added to obtain the predicted box output by the decoder. Figure 4 Figure 2 shows a schematic diagram of the structure of the Transformer module.

[0071] The specific implementation function of the residual block can be expressed by the following formula:

[0072] I l+1 = l +[θ l+1 *(I l · l )]

[0073] k l+1 =[C2*[C1*[θ l+1 *(I l · l )]+k l ]]

[0074] Among them, I l+1 is the feature of the l+1th convolutional layer in the residual block, I l is the feature of the lth convolutional layer in the residual block, θ l+1 is the convolution parameter of the l+1th convolution layer in the residual block, k l is the hidden tensor of the lth convolutional layer in the residual block, k l+1 is the hidden tensor of the l+1th convolutional layer in the residual block, C2 and C1 are 1×1 and 3×3 convolutions respectively, * represents the convolution operation, and · represents the concatenation operation. Figure 3 The purple blocks in the figure represent tensors, and the green blocks represent hidden tensors.

[0075] In the above embodiment, the original Transformer module directly learns instance information through the queries module, and decodes the queries information through the mlp in the Transformer module to obtain the instance category and location information (box). In this embodiment, the task is decomposed into two parts: box prediction and instance prediction (box offset and category). The learnable anchors module is used to obtain the updated learnable box, and the queries module decodes the box offset and category. The box offset is added to the learnable box to obtain the final predicted box. Since the box information is decoupled, it can be used through the auxiliary learning task ( Figure 4 The auxiliary module skips the bipartite matching, and all losses are generated by the box offset and category components, balancing the impact of inconsistent matching.

[0076] In one embodiment, in step 6, the weight parameters of the teacher model are updated according to the weight parameters of the trained student model using the following formula:

[0077]

[0078] Among them, θ e is the weight parameter of the teacher model updated in the e-th iteration training, μ is the coefficient, θ e-1 The weight parameters of the teacher model updated for the e-1th iteration training, is the weight parameter of the trained student model obtained after the e-th iteration training.

[0079] In one embodiment, the total loss function used during the training of the teacher model and the student model is:

[0080] L=σL match +γL known

[0081] Among them, L is the total loss function, L match is the matching loss, L known is the known loss, σ is the weight of the matching loss, and γ is the weight of the known loss; here, L match This includes the MSE loss used to measure the target classification prediction, and the L1 or L2 loss used to measure the size accuracy of the target position box prediction.

[0082] Known loss L known for:

[0083]

[0084] Among them, i is the i-th prediction result, M is the number of prediction results, c i is the prediction category corresponding to the i-th prediction result, c i The corresponding predicted probability, b i is the prediction box corresponding to the i-th prediction result, b(i) is the real box, L box It is L1 loss or L2 loss.

[0085] Based on the same inventive concept as the method for constructing a fully supervised target detection model, this embodiment also provides a corresponding device for constructing a fully supervised target detection model, the device comprising:

[0086] The dataset construction module is used to construct the target detection dataset, which includes the standard annotation dataset and the weak annotation dataset;

[0087] The first training module is used to train the teacher model based on the standard labeled data set to obtain the trained teacher model;

[0088] The prediction module is used to input the weakly labeled dataset into the trained teacher model, which is the current teacher model, to obtain the prediction results;

[0089] The weakly labeled dataset update module is used to filter the prediction results by confidence threshold and weakly labeled information, generate pseudo labels, and use the pseudo labels to update the weakly labeled dataset to obtain an updated weakly labeled dataset;

[0090] The second training module is used to train the student model based on the standard labeled dataset and the updated weakly labeled dataset to obtain a trained student model;

[0091] The weight parameter updating module is used to update the weight parameters of the teacher model according to the weight parameters of the trained student model to obtain an updated teacher model;

[0092] The judgment module is used to determine whether the conditions for the end of iterative training have been met. If so, the updated teacher model is output as the final target detection model. If not, the updated teacher model is used as the current teacher model and enters the prediction module.

[0093] The fully supervised target detection model construction device of this embodiment has the same inventive concept as the fully supervised target detection model construction method mentioned above. Therefore, the specific implementation method of the device can be seen in the embodiment part of the fully supervised target detection model construction method mentioned above, and its technical effect corresponds to the technical effect of the above method, which will not be repeated here.

[0094] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned fully supervised target detection model construction method.

[0095] An embodiment of the present application provides a computer program product, including a computer program / instruction, which implements the above-mentioned fully supervised target detection model construction method when the computer program / instruction is executed by a processor.

[0096] In order to further verify the effectiveness of the model constructed in this application, a target detection experiment was conducted. Figure 5 The comparison chart of the target detection results of the model constructed by the method of this application and the existing model Omni-DETR is shown. Figure 5 It can be seen that it is difficult to detect close instances, resulting in problems such as missed detection or confusion. The model constructed by the method of the present application demonstrates advanced detection capabilities, can clearly distinguish cross-samples, can effectively accelerate the model convergence speed, and improve the model detection performance.

[0097] The above descriptions are merely examples of various embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any modifications or substitutions that can be readily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for constructing a fully supervised target detection model, characterized in that: include: Step 1: construct a target detection dataset, which includes a standard annotation dataset and a weak annotation dataset; The weakly labeled dataset includes multiple sample weakly labeled types; The standard annotation dataset and the weak annotation dataset both include multiple samples, each of which is an image with multiple objects; Step 2: training a teacher model based on the standard labeled dataset to obtain a trained teacher model; Step 3: Input the weakly labeled dataset into the trained teacher model, where the trained teacher model is the current teacher model, to obtain a prediction result; Step 4: Filter the prediction results by using a confidence threshold and weakly labeled information to generate pseudo labels, and use the pseudo labels to update the weakly labeled dataset to obtain an updated weakly labeled dataset. The weak annotation information filtering includes: If the sample's annotation type is the first type: no annotation, it is filtered based on the confidence threshold; If the sample's annotation type is the second type: category annotation, then based on the weakly labeled category, the prediction results that do not belong to the weakly labeled category are removed; and based on the confidence of the retained prediction results, at least one instance is assigned to the unpredicted category in the weakly labeled category; If the sample is labeled as the third type: the number of each category is marked, then the prediction results that do not belong to the weakly labeled category are removed according to the weakly labeled category; and the corresponding number of instances are assigned to each category according to the confidence level; If the annotation type of the sample is the fourth type: point annotation, then according to the point position information of the point annotation, the prediction results whose prediction boxes are outside the point position information of the point annotation are removed; If the sample's annotation type is the fifth type: point annotation and category annotation, then based on the point location information of the point annotation, remove the prediction results whose prediction boxes are outside the point location information of the point annotation; and remove the prediction results whose predicted categories do not match the category annotation corresponding to the point annotation; If the sample's annotation type is the sixth type: standard box annotation, then after filtering with the confidence threshold, for each instance of the standard box annotation, calculate the intersection-over-union ratio between all predicted boxes and the annotation box, sort the intersection-over-union ratios, and select the prediction result with the largest intersection-over-union ratio as the pseudo label of the instance; If the sample's annotation type is the seventh type: noise box annotation, it is processed in the same way as the sixth type; Step 5: training a student model based on the standard labeled dataset and the updated weakly labeled dataset to obtain a trained student model; Step 6: updating the weight parameters of the teacher model according to the weight parameters of the trained student model to obtain an updated teacher model; Step 7, determine whether the conditions for the end of iterative training are met. If so, output the updated teacher model, which is the final target detection model. If not, use the updated teacher model as the current teacher model and return to step 3.

2. The method according to claim 1, wherein The teacher model and the student model have the same structure. The teacher model includes a residual network and a Transformer module connected in sequence. The residual network includes multiple residual blocks, each residual block includes multiple convolutional layers, and the input data outputs a tensor after passing through the residual block; the Transformer module includes an encoder and a decoder, and the decoder includes a Querises module and a LearnableAnchors module. The Querises module is used to obtain the box offset and category, and the Learnable Anchors module is used to obtain the updated learnable box. The box offset and the updated learnable box are added to obtain the predicted box output by the decoder.

3. The method according to claim 2, wherein The residual block specifically implements the function, which is expressed by the following formula: I l+1 =I l +[θ l+1 *(I l ·k l )] k l+1 =[C2*[C1*[θ l+1 *(I l ·k l )]+k l ]] Among them, I l+1 is the feature of the l+1th convolutional layer in the residual block, I l is the feature of the lth convolutional layer in the residual block, θ l+1 is the convolution parameter of the l+1th convolution layer in the residual block, k l is the hidden tensor of the lth convolutional layer in the residual block, k l+1 is the hidden tensor of the l+1th convolutional layer in the residual block, C2 and C1 are 1×1 and 3×3 convolutions respectively, * represents the convolution operation, and · represents the concatenation operation.

4. The method according to claim 1, wherein The types of weak annotations of samples in the weakly labeled dataset include at least two of: no annotation, category annotation, annotation of the number of each category, point annotation, point annotation and category annotation, standard box annotation and noise box annotation.

5. The method according to claim 1, wherein in, The weight parameters of the teacher model are updated according to the weight parameters of the trained student model using the following formula: Among them, θ e is the weight parameter of the teacher model updated in the e-th iteration training, μ is the coefficient, θ e-1 The weight parameters of the teacher model updated for the e-1th iteration training, is the weight parameter of the trained student model obtained after the e-th iteration training.

6. The method according to claim 1, wherein The condition for the iterative training to end is that the loss of the updated teacher model tends to be stable.

7. The method according to claim 1, wherein During the training of the teacher model and the student model, the total loss function used is: L=σL match +γL known Among them, L is the total loss function, L match is the matching loss, L known is the known loss, σ is the weight of the matching loss, and γ is the weight of the known loss; Known loss L known for: Among them, i is the i-th prediction result, M is the number of prediction results, c i is the prediction category corresponding to the i-th prediction result, c i The corresponding predicted probability, b i is the prediction box corresponding to the i-th prediction result, b(i) is the real box, L box It is L1 loss or L2 loss.

8. A fully supervised target detection model construction device, characterized in that: include: A data set construction module is used to construct a target detection data set, wherein the target detection data set includes a standard annotation data set and a weak annotation data set; The weakly labeled dataset includes multiple sample weakly labeled types; the standard labeled dataset and the weakly labeled dataset both include multiple samples, each of which is an image with multiple objects; A first training module is used to train a teacher model based on the standard labeled data set to obtain a trained teacher model; A prediction module is used to input the weakly labeled data set into the trained teacher model, where the trained teacher model is the current teacher model, to obtain a prediction result; A weakly labeled dataset updating module is used to perform confidence threshold filtering and weakly labeled information filtering on the prediction results, generate pseudo labels, and use the pseudo labels to update the weakly labeled dataset to obtain an updated weakly labeled dataset; The weak annotation information filtering includes: If the sample's annotation type is the first type: no annotation, it is filtered based on the confidence threshold; If the sample's annotation type is the second type: category annotation, then based on the weakly labeled category, the prediction results that do not belong to the weakly labeled category are removed; and based on the confidence of the retained prediction results, at least one instance is assigned to the unpredicted category in the weakly labeled category; If the sample is labeled as the third type: the number of each category is marked, then the prediction results that do not belong to the weakly labeled category are removed according to the weakly labeled category; and the corresponding number of instances are assigned to each category according to the confidence level; If the annotation type of the sample is the fourth type: point annotation, then according to the point position information of the point annotation, the prediction results whose prediction boxes are outside the point position information of the point annotation are removed; If the sample's annotation type is the fifth type: point annotation and category annotation, then based on the point location information of the point annotation, remove the prediction results whose prediction boxes are outside the point location information of the point annotation; and remove the prediction results whose predicted categories do not match the category annotation corresponding to the point annotation; If the sample's annotation type is the sixth type: standard box annotation, then after filtering with the confidence threshold, for each instance of the standard box annotation, calculate the intersection-over-union ratio between all predicted boxes and the annotation box, sort the intersection-over-union ratios, and select the prediction result with the largest intersection-over-union ratio as the pseudo label of the instance; If the sample's annotation type is the seventh type: noise box annotation, it is processed in the same way as the sixth type; A second training module is used to train a student model based on the standard labeled dataset and the updated weakly labeled dataset to obtain a trained student model; A weight parameter updating module, configured to update the weight parameters of the teacher model according to the weight parameters of the trained student model to obtain an updated teacher model; The judgment module is used to judge whether the conditions for the end of iterative training are met. If so, the updated teacher model is output as the final target detection model. If not, the updated teacher model is used as the current teacher model and enters the prediction module.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for constructing a fully supervised target detection model according to any one of claims 1 to 7.

10. A computer program product, characterized in that The method comprises a computer program / instruction, which, when executed by a processor, implements the method for constructing a fully supervised target detection model according to any one of claims 1 to 7.

Images