Agricultural material label character detection method and system based on multi-scale self-distillation mechanism
By training a teacher-student model using a multi-scale self-distillation mechanism, the problems of recognition accuracy and robustness in small fonts and complex backgrounds in agricultural input label recognition were solved, achieving efficient text detection and recognition results.
Patent Information
- Application Number
- CN202511935343.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-22
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies for agricultural input label recognition suffer from limitations in detection accuracy, reduced model generalization ability, unstable training, and difficulty in deployment with complex structures, especially in the case of small fonts and complex backgrounds where recognition performance is poor.
A multi-scale self-distillation mechanism is adopted, which optimizes the student model parameters through adaptive knowledge distillation training of the teacher-student model and combines multi-scale data augmentation strategy to improve the recognition accuracy and robustness of small fonts and complex backgrounds.
While maintaining the model's lightweight design, it improves the recognition accuracy of small fonts and complex texture labels, reduces text boundary breakage and character splitting, and enhances the model's robustness in complex environments.
Smart Images
Figure CN121725489A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of label text detection technology, and in particular relates to a method and system for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism. Background Technology
[0002] Traditional text recognition methods primarily rely on image preprocessing, edge extraction, and manually designed features for classification. Their algorithmic structures are relatively fixed, exhibiting poor robustness in handling complex backgrounds, multiple fonts, and multi-angle scenes, making them unsuitable for accurate recognition in natural scenes and industrial applications. In recent years, with the development of deep learning technology, Convolutional Neural Networks (CNNs) and attention-based network structures have been widely applied to text detection and recognition tasks. In text detection, algorithms such as EAST, DB, and CRAFT have effectively improved the accuracy and stability of text region localization by introducing Feature Pyramid Networks (FPNs) and differentiable binarization modules. In text recognition, networks such as CRNN, RARE, and SVTR have achieved an end-to-end recognition process from image input to character output by combining sequence modeling and convolutional feature extraction. In recent years, the Transformer structure has been further introduced into OCR tasks (such as TrOCR and SVTR-Large), demonstrating stronger modeling capabilities in long texts and complex layout scenarios.
[0003] In specific application scenarios (such as agricultural packaging, industrial labels, and commercial invoices), text detection still faces several technical challenges. The text in these scenarios has the following characteristics: significant differences in font size, with some labels containing extremely small or dense text; irregular character arrangement, including multi-line, multi-angle, and curved surface printing; and complex backgrounds, often containing interfering information such as patterns, textures, and reflective materials.
[0004] To address the above characteristics, existing technologies often employ transfer learning or fine-tuning training based on general detection networks (such as YOLO and PP-OCR) to enhance small font detection capabilities. However, these methods generally suffer from the following shortcomings: (1) Limited detection accuracy. While enhancing small font samples can improve local performance, it can easily lead to errors in the recognition of large font regions or incomplete text boundary segmentation. (2) Decreased model generalization ability. Direct fine-tuning training on specific domain data can easily lead to catastrophic forgetting, resulting in a decrease in the model's recognition performance in general scenarios. (3) Insufficient knowledge transfer. Traditional fine-tuning methods only utilize the supervision signals of new samples and fail to fully utilize the feature knowledge of existing high-performance models, leading to unstable training and difficulty in model convergence. (4) Complex end-to-end structure. Although some end-to-end recognition methods (such as Mask TextSpotter and ABINet) possess unified detection and recognition capabilities, their complex network structures and large number of parameters make them difficult to deploy in embedded devices or edge computing environments. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a method and system for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism.
[0006] Firstly, the method includes, Collect agricultural input label images of various types and sizes, preprocess the agricultural input label images, and train a high-performance model based on the preprocessed agricultural input label images; Use the high-performance model after training as the teacher model and freeze the parameters of the teacher model; Construct a student model and train it using the weights of the teacher model as initial parameters; Adaptive self-knowledge distillation training is performed on the teacher model and the trained student model; The teacher model and the trained student model are trained iteratively multiple times to continuously optimize the parameters of the student model. The student model, which has been trained iteratively, is applied to agricultural input label images to detect text regions in the labels, obtain detection results, and then perform post-processing and text content recognition on the detection results.
[0007] Furthermore, the various types and sizes of agricultural input label images specifically include image samples with different shooting angles, lighting conditions, and background complexity.
[0008] Furthermore, the preprocessing of the label images specifically includes image decoding, normalization, scale transformation, affine perturbation, and multi-scale cropping preprocessing of the acquired label images.
[0009] Furthermore, the adaptive self-knowledge distillation training of the teacher model and the student model specifically includes inputting the teacher model and the student model with the same preprocessed label image, extracting intermediate features and detection results respectively, and calculating the distillation loss function.
[0010] Furthermore, the calculation of the distillation loss function specifically includes calculating the distillation master detection loss, calculating the distillation boundary constraint loss, and calculating the distillation feature distance loss.
[0011] Secondly, the system includes: a data processing module, a teacher model freezing module, a student model construction module, a knowledge distillation training module, an iterative training module, and a text recognition module; The data processing module is used to collect agricultural input label images of various types and sizes, preprocess the label images, and train a high-performance model based on the preprocessed label images. The teacher model freezing module uses the high-performance model after training as the teacher model and freezes the parameters of the teacher model. The student model building module is used to build a student model and train the student model using the weights of the teacher model as initial parameters. The knowledge distillation training module is used to perform adaptive self-knowledge distillation training on the teacher model and the trained student model. The iterative training module is used to perform multiple iterative trainings on the teacher model and the trained student model to continuously optimize the student model parameters. The text recognition module is used to apply the iteratively trained student model to agricultural material label images, perform text region detection in the labels, obtain detection results, and perform post-processing and text content recognition on the detection results.
[0012] Furthermore, the various types and sizes of agricultural input label images specifically include image samples with different shooting angles, lighting conditions, and background complexity.
[0013] Furthermore, the data processing module is specifically used to perform image decoding, normalization, scale transformation, affine perturbation, and multi-scale cropping preprocessing on the collected label images.
[0014] Furthermore, the knowledge distillation training module is specifically used to input the same preprocessed label image into the teacher model and the student model, extract intermediate features and detection results respectively, and calculate the distillation loss function.
[0015] Furthermore, the knowledge distillation training module is also used to calculate the distillation master detection loss, the distillation boundary constraint loss, and the distillation feature distance loss.
[0016] Thirdly, a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any of the steps of the above-described method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism.
[0017] Fourthly, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus. Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements any of the steps of the above-described method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism.
[0018] Compared with the prior art, the present invention has the following advantages: 1. This invention proposes a self-distillation training framework for agricultural input label recognition tasks. By introducing a teacher-student dynamic distillation mechanism in the same structure, the model can improve the recognition accuracy of small font and complex texture labels while maintaining lightweight design.
[0019] 2. The composite distillation loss function of this invention achieves multi-layer feature alignment, effectively reducing text boundary breakage and character splitting phenomena, improving the integrity of detection results, and combined with an adaptive multi-scale data augmentation strategy, enabling the model to adapt to small characters, high density and large characters, sparse scenes, improving the robustness of the model in complex environments.
[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 The diagram shows a flowchart of a method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism according to the present invention.
[0023] Figure 2 The diagram shows a schematic of a text detection system module for agricultural input labels based on a multi-scale self-distillation mechanism according to the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] like Figure 1 As shown, this invention proposes a text detection method for agricultural input labels based on a multi-scale self-distillation mechanism. By training a student model through self-distillation under the feature guidance of a teacher model, the student model maintains the global generalization ability of the teacher model during fine-tuning while enhancing detection accuracy for small fonts and complex labels. This allows for stable text detection and recognition under different sizes, densities, and background conditions, improving detection robustness and applicability. The steps include... S1. Collect agricultural input label images of various types and sizes, preprocess the label images, and train a high-performance model based on the processed label images.
[0026] In this embodiment, the present invention collects agricultural input label images of various types and sizes, including image samples with different shooting angles, lighting conditions and background complexity.
[0027] In this embodiment, the present invention also combines boundary map and shrinkage map generation mechanisms, enabling the model to have stronger small text detection and complex background adaptation capabilities while maintaining semantic consistency.
[0028] S2. Use the high-performance model after training as the teacher model and freeze the parameters of the teacher model.
[0029] In this embodiment, the present invention selects a high-performance model that has been trained on a general text detection task as the teacher model, freezes its parameters to maintain its high generalization ability in general scenarios, and provides stable feature guidance and prediction output reference for subsequent student models.
[0030] S3. Construct a student model and train it using the weights of the teacher model as initial parameters.
[0031] In this embodiment, the structure of the student model is kept consistent with that of the teacher model, and the weights of the teacher model are used as initial parameters. The student model participates in gradient updates during training to adapt to small fonts, sparse text, and complex background features in agricultural input label images.
[0032] S4. Perform adaptive self-knowledge distillation training on the teacher model and the trained student model.
[0033] In this embodiment, the present invention performs adaptive self-knowledge distillation training on the teacher model and the student model. Specifically, this includes inputting the same preprocessed labeled image into both the teacher model and the student model, extracting intermediate features and detection results respectively, and calculating the distillation loss function, which includes... 1. Calculate the distillation master detection loss to guide the student model in learning text detection regions, specifically the text region probability map (or segmentation map) predicted by the teacher model. Correspondence between student model predictions and the graph Calculate the differences between them; commonly used losses include: Kullback-Leibler (KL) divergence loss : The student model's predicted probability distribution is encouraged to match the teacher's smoothed distribution. The teacher model's output is typically softened with a higher temperature parameter to include information about inter-class relationships. Mean Squared Error (MSE) Loss : This directly reverts to the soft objective of the teacher model.
[0034] 2. Calculate the distillation boundary constraint loss to improve the accuracy of text edge detection; specifically, this includes: a. Utilize the Ground Truth Boundary Map generated during the preprocessing stage. As a supplementary supervisory function.
[0035] b. Extract the feature layers with strong edge responses from the intermediate feature maps of the teacher model and student model respectively (or decode the predicted boundary map from the main features through a small side-channel convolutional network). and ).
[0036] c. Calculate triple loss: Loss from student boundary prediction versus true boundary: ; Loss for teacher boundary prediction and true boundary (for reference only, not for gradient update of teacher model). Distillation boundary loss: (L1 loss, more sensitive to edges); forces the student model to mimic the teacher model's ability to perceive text edges.
[0037] 3. Calculate the distillation feature distance loss to maintain consistency between the feature spaces of the student model and the teacher model. Specifically, this includes... a. Select L intermediate layers corresponding to the teacher and student models (e.g., the outputs of the last three stages of the backbone network).
[0038] b. For each pair of corresponding feature maps and (Normalization or channel alignment via an adapter layer may be required first), calculate the distance between them.
[0039] Optionally, the measurement methods used in this invention include: L2 distance (MSE), cosine similarity loss, and attention shift loss, etc.
[0040] In this embodiment, the present invention adjusts the gradient according to the weights of each loss and introduces an adaptive adjustment mechanism, so that the student model can maintain the stability of large character detection while learning small character features.
[0041] Optionally, an adaptive adjustment mechanism may be included, specifically including: a. Monitoring and Evaluation: During training, continuously monitor the changing trends of various losses on the validation set, as well as their impact on the performance of the student model on the "small-character validation subset" and the "large-character validation subset".
[0042] b. Dynamic weight adjustment: If it is found that the small character detection performance of the student model improves slowly, while the large character detection performance has saturated, the boundary constraint loss can be appropriately increased. The weight β is adjusted because small character detection relies more on accurate edges; conversely, if the model starts to overfit or perform inconsistently on small characters, the feature distance loss can be increased. The weight γ is used to pull back the teacher model to a stable feature space.
[0043] c. Gradient modulation: Modulate the gradient level, for example, normalize or scale the gradients from different loss terms to prevent one loss term from dominating the optimization direction and ensure the model develops in a balanced manner.
[0044] S5. Iterate and train the teacher model and the trained student model multiple times to continuously optimize the student model parameters.
[0045] In this embodiment, the present invention continuously optimizes the parameters of the student model through multiple rounds of iterative training; wherein, during the training process, the loss weights and validation set accuracy are dynamically monitored, and when the detection performance of the student model under the new data distribution is stable, the training is stopped and the model is saved.
[0046] S6. Apply the iteratively trained student model to agricultural input label images to detect text regions in the labels, obtain detection results, and perform post-processing and text content recognition on the detection results.
[0047] In this embodiment, the trained student model is applied to agricultural input label images to achieve high-precision detection of text regions in the labels; the detection results are then post-processed and the text content is recognized to generate standardized label information.
[0048] like Figure 2 As shown, this invention proposes an agricultural input label text detection system based on a multi-scale self-distillation mechanism, which includes: a data processing module, a teacher model freezing module, a student model construction module, a knowledge distillation training module, an iterative training module, and a text recognition module.
[0049] 1. The data processing module is used to collect agricultural input label images of various types and sizes, perform preprocessing operations on the images, and train a high-performance model based on the processed images.
[0050] In this embodiment, the data processing module acquires agricultural input label images of various types and sizes, including samples with different shooting angles, lighting conditions, and background complexity; and performs preprocessing operations such as image decoding, normalization, scale transformation, affine perturbation, and multi-scale cropping on the acquired images.
[0051] In this embodiment, the data processing module also incorporates boundary map and shrinkage map generation mechanisms, enabling the model to have stronger small text detection and complex background adaptation capabilities while maintaining semantic consistency.
[0052] 2. The teacher model freeze module is used to use the high-performance model after training as the teacher model and freeze the parameters of the teacher model.
[0053] In this embodiment, the teacher model freezing module selects a high-performance model that has been trained on a general text detection task as the teacher model, freezes its parameters to maintain its high generalization ability in general scenarios, and provides stable feature guidance and prediction output reference for subsequent student models.
[0054] 3. The student model building module is used to build student models and train them based on the weights of the teacher model as initial parameters.
[0055] In this embodiment, the student model building module keeps the structure of the student model consistent with that of the teacher model and uses the weights of the teacher model as the initial parameters; wherein, the student model participates in gradient updates during training to adapt to small fonts, sparse text and complex background features in agricultural material label images.
[0056] 4. The knowledge distillation training module is used to perform adaptive self-knowledge distillation training on the teacher model and the trained student model.
[0057] In this embodiment, the loss function calculation module inputs the teacher model and student model into the same batch of training images, extracts intermediate features and detection results respectively, and calculates the distillation loss function, which includes: (1) Calculate the distillation master detection loss, which is used to guide the student model to learn the text detection region, specifically the text region probability map (or segmentation map) predicted by the teacher model. Correspondence between student model predictions and the graph Calculate the differences between them; commonly used losses include: Kullback-Leibler (KL) divergence loss : The student model's predicted probability distribution is encouraged to match the teacher's smoothed distribution. The teacher model's output is typically softened with a higher temperature parameter to include information about inter-class relationships. Mean Squared Error (MSE) Loss : This directly reverts to the soft objective of the teacher model.
[0058] (2) Calculate the distillation boundary constraint loss to improve the accuracy of text edge detection; specifically, this includes, a. Utilize the Ground Truth Boundary Map generated during the preprocessing stage. As a supplementary supervisory function.
[0059] b. Extract the feature layers with strong edge responses from the intermediate feature maps of the teacher model and student model respectively (or decode the predicted boundary map from the main features through a small side-channel convolutional network). and ).
[0060] c. Calculate triple loss: Loss from student boundary prediction versus true boundary: ; Loss for teacher boundary prediction and true boundary (for reference only, not for gradient update of teacher model). Distillation boundary loss: (L1 loss, more sensitive to edges); forces the student model to mimic the teacher model's ability to perceive text edges.
[0061] (4) Calculate the distillation feature distance loss to maintain consistency between the feature spaces of the student model and the teacher model, specifically including: a. Select L intermediate layers corresponding to the teacher and student models (e.g., the outputs of the last three stages of the backbone network).
[0062] b. For each pair of corresponding feature maps and (Normalization or channel alignment via an adapter layer may be required first), calculate the distance between them.
[0063] Optionally, the measurement methods used in this invention include: L2 distance (MSE), cosine similarity loss, and attention shift loss, etc.
[0064] In this embodiment, the present invention adjusts the gradient according to the weights of each loss and introduces an adaptive adjustment mechanism, so that the student model can maintain the stability of large character detection while learning small character features.
[0065] Optionally, an adaptive adjustment mechanism may be included, specifically including: a. Monitoring and Evaluation: During training, continuously monitor the changing trends of various losses on the validation set, as well as their impact on the performance of the student model on the "small-character validation subset" and the "large-character validation subset".
[0066] b. Dynamic weight adjustment: If it is found that the small character detection performance of the student model improves slowly, while the large character detection performance has saturated, the boundary constraint loss can be appropriately increased. The weight β is adjusted because small character detection relies more on accurate edges; conversely, if the model starts to overfit or perform inconsistently on small characters, the feature distance loss can be increased. The weight γ is used to pull back the teacher model to a stable feature space.
[0067] c. Gradient modulation: Modulate the gradient level, for example, normalize or scale the gradients from different loss terms to prevent one loss term from dominating the optimization direction and ensure the model develops in a balanced manner.
[0068] 5. The iterative training module is used to perform multiple iterations of training on the teacher model and the trained student model, continuously optimizing the student model parameters.
[0069] In this embodiment, the present invention continuously optimizes the parameters of the student model through multiple rounds of iterative training; wherein, during the training process, the loss weights and validation set accuracy are dynamically monitored, and when the detection performance of the student model under the new data distribution is stable, the training is stopped and the model is saved.
[0070] 6. The text recognition module is used to apply the iteratively trained student model to agricultural material label images, perform text region detection in the labels, obtain detection results, and perform post-processing and text content recognition on the detection results.
[0071] In this embodiment, the trained student model is applied to agricultural input label images to achieve high-precision detection of text regions in the labels; the detection results are then post-processed and the text content is recognized to generate standardized label information.
[0072] Based on the above disclosure, the present invention also provides an electronic device. The electronic device of this embodiment includes at least one processor and at least one storage medium electrically connected to the processor. The storage medium is electrically connected to the processor, wherein the storage medium stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.
[0073] Based on the same inventive concept, the present invention also provides a storage medium storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described above.
[0074] The foregoing description and accompanying drawings fully illustrate embodiments of the invention to enable those skilled in the art to practice them. Other embodiments may include structural and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Some portions and features of some embodiments may be included or substituted for portions and features of other embodiments. Embodiments of the invention are not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from their scope. The scope of the invention is limited only by the appended claims.
Claims
1. A method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism, characterized in that, The method includes, Collect agricultural input label images of various types and sizes, preprocess the agricultural input label images, and train a high-performance model based on the preprocessed agricultural input label images; Use the high-performance model after training as the teacher model and freeze the parameters of the teacher model; Construct a student model and train it using the weights of the teacher model as initial parameters; Adaptive self-knowledge distillation training is performed on the teacher model and the trained student model; The teacher model and the trained student model are trained iteratively multiple times to continuously optimize the parameters of the student model. The student model, which has been trained iteratively, is applied to agricultural input label images to detect text regions in the labels, obtain detection results, and then perform post-processing and text content recognition on the detection results.
2. The method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism according to claim 1, characterized in that, The various types and sizes of agricultural input label images specifically include image samples with different shooting angles, lighting conditions, and background complexity.
3. The method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism according to claim 1, characterized in that, The preprocessing of agricultural input label images specifically includes image decoding, normalization, scale transformation, affine perturbation, and multi-scale cropping preprocessing of the collected label images.
4. The method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism according to claim 1, characterized in that, The adaptive self-knowledge distillation training of the teacher model and the student model specifically includes inputting the teacher model and the student model with the same preprocessed label image, extracting intermediate features and detection results respectively, and calculating the distillation loss function.
5. The method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism according to claim 4, characterized in that, The calculation of the distillation loss function specifically includes calculating the distillation master detection loss, calculating the distillation boundary constraint loss, and calculating the distillation feature distance loss.
6. A text detection system for agricultural input labels based on a multi-scale self-distillation mechanism, characterized in that, The system includes: a data processing module, a teacher model freezing module, a student model construction module, a knowledge distillation training module, an iterative training module, and a text recognition module; The data processing module is used to collect agricultural input label images of various types and sizes, preprocess the label images, and train a high-performance model based on the preprocessed label images. The teacher model freezing module uses the high-performance model after training as the teacher model and freezes the parameters of the teacher model. The student model building module is used to build a student model and train the student model using the weights of the teacher model as initial parameters. The knowledge distillation training module is used to perform adaptive self-knowledge distillation training on the teacher model and the trained student model. The iterative training module is used to perform multiple iterative trainings on the teacher model and the trained student model to continuously optimize the student model parameters. The text recognition module is used to apply the iteratively trained student model to agricultural material label images, perform text region detection in the labels, obtain detection results, and perform post-processing and text content recognition on the detection results.
7. The agricultural input label text detection system based on a multi-scale self-distillation mechanism according to claim 6, characterized in that, The various types and sizes of agricultural input label images specifically include image samples with different shooting angles, lighting conditions, and background complexity.
8. The agricultural input label text detection system based on a multi-scale self-distillation mechanism according to claim 6, characterized in that, The data processing module is specifically used to perform image decoding, normalization, scale transformation, affine perturbation, and multi-scale cropping preprocessing on the collected label images.
9. The agricultural input label text detection system based on a multi-scale self-distillation mechanism according to claim 6, characterized in that, The knowledge distillation training module is specifically used to input the same preprocessed label image into the teacher model and the student model, extract intermediate features and detection results respectively, and calculate the distillation loss function.
10. The agricultural input label text detection system based on a multi-scale self-distillation mechanism according to claim 9, characterized in that, The knowledge distillation training module is also used to calculate the distillation master detection loss, the distillation boundary constraint loss, and the distillation feature distance loss.
11. 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 steps of the agricultural input label text detection method based on a multi-scale self-distillation mechanism as described in any one of claims 1-5.
12. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the steps of the method for detecting text on agricultural input labels based on a multi-scale self-distillation mechanism as described in any one of claims 1-5.