Electric power vision large model multi-scale semi-supervised target detection method and system
By constructing a multi-scale visual codebook with enhanced power semantics and an adaptive threshold mechanism, and using unlabeled data for semi-supervised fine-tuning, the problem of scarce labeled data and high computational complexity in the power field of large-scale power vision models is solved, and efficient multi-scale defect detection and edge device deployment are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID ELECTRIC POWER RES INST
- Filing Date
- 2026-02-09
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies in the power sector suffer from problems such as scarce labeled data, semantic gaps in models, and high computational complexity, resulting in insufficient accuracy of large-scale power vision models in detecting multi-scale defects and difficulty in deploying them on edge devices.
A multi-scale semi-supervised target detection method is adopted. A multi-scale visual codebook with power semantic enhancement is constructed through a vector quantization knowledge distillation framework. Combined with mask image modeling and adaptive thresholding mechanism, semi-supervised fine-tuning is performed using unlabeled data. The SMLS fine-tuning strategy is adopted to reduce the number of parameters and achieve efficient deployment.
It significantly improves the accuracy and robustness of multi-scale defect detection, reduces annotation costs and computational overhead, enables efficient deployment on edge devices, and alleviates the class imbalance problem.
Smart Images
Figure CN122067192A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a target detection method and system for large visual models, and more particularly to a multi-scale semi-supervised target detection method and system for large visual models of electric power systems. Background Technology
[0002] With the advancement of new power system construction, the scale and complexity of facilities such as transmission lines and substations continue to increase, posing serious challenges to the efficiency, accuracy, and safety of traditional manual inspection methods. Utilizing computer vision technology to achieve intelligent inspection has become an industry trend. Target detection, as a core technology, can automatically locate and identify equipment defects (such as insulator cracks or broken conductor strands) or safety violations (such as not wearing a safety helmet) in images.
[0003] In recent years, large visual models based on the Transformer architecture (such as the ViT and BEiT series) have learned powerful general visual representation capabilities through self-supervised pre-training on large-scale datasets (such as masked image modeling MIM). However, directly applying them to the highly specialized power field presents the following problems: (1) Scarcity and high cost of labeled data: The identification of defects in power equipment requires deep domain knowledge, resulting in highly specialized, long-term, and extremely costly data labeling, making it difficult to obtain a large amount of high-quality labeled data. (2) Domain semantic gap: The visual concepts (such as "animals" and "vehicles") learned by large models pre-trained on general datasets (such as ImageNet) differ greatly from the fine-grained semantics (such as "loose equalizing ring" and "corroded vibration hammer") specific to the power field, resulting in insufficient generalization ability of the model. (3) Multi-scale feature perception problem: Power inspection images simultaneously contain large-scale overall structures (such as the entire conductor and tower) and small-scale local defects (such as fine cracks and missing pins). Existing methods typically use feature fusion at a single scale or a fixed pyramid, which makes it difficult to adaptively and accurately model such cross-scale semantic associations. (4) High model computational complexity: Visual large models have a huge number of parameters, high computational and storage overhead, and are difficult to deploy directly to edge devices with limited computing power (such as inspection drones and mobile terminals) for real-time inference.
[0004] In existing technologies, semi-supervised learning (SSL) methods such as FixMatch, FlexMatch, and FreeMatch attempt to alleviate label dependency by utilizing large amounts of unlabeled data. Among them, FreeMatch's adaptive thresholding (SAT) and fairness regularization (SAF) mechanisms are highly effective in scenarios with low labeling rates. However, these methods are mainly designed for image classification tasks, and their thresholding mechanisms are image-level or category-level, which cannot directly handle the confidence evaluation problem of targets at different scales and locations in object detection. Meanwhile, parameter efficient fine-tuning (PEFT) techniques such as LoRA, although they can effectively reduce the fine-tuning cost of large models, do not consider the different sensitivity of targets at different scales to features in visual tasks due to the design of their low-rank adaptation matrices.
[0005] Therefore, it is of great significance to provide a multi-scale semi-supervised target detection method and system based on a large-scale power vision model. Summary of the Invention
[0006] Purpose of the invention: The purpose of this invention is to provide a multi-scale semi-supervised target detection method and system for large-scale power vision models, which solves the problems of scarce labeled data, semantic gap in models and high computational complexity in power scenarios, significantly improves the detection accuracy and robustness of the model for multi-scale defects, and reduces the model's dependence on power labeled data and the deployment cost on edge devices.
[0007] Technical solution: The present invention provides a multi-scale semi-supervised target detection method using a large-scale power vision model, comprising:
[0008] Acquire power image data, including labeled and unlabeled power images;
[0009] A multi-scale visual word segmenter is trained based on a vector quantization knowledge distillation framework to construct a multi-scale visual codebook for power semantic enhancement. The multi-scale visual codebook includes a top-level codebook for encoding the global structure of power equipment and a bottom-level codebook for encoding local features of equipment defects.
[0010] Based on the multi-scale visual codebook, a large visual model is pre-trained on unlabeled power images using a masked image modeling task.
[0011] Semi-supervised fine-tuning of the large visual model was performed using labeled and unlabeled power image data; the parameters of the large visual model were updated using the SMLS fine-tuning strategy.
[0012] The image of the power source to be detected is input into the updated visual large model, and the detection results of multi-scale targets are output.
[0013] Furthermore, the multi-scale visual word segmenter includes a visual encoder and two quantizers. The specific steps for training the multi-scale visual word segmenter encoder are as follows:
[0014] Image data from the power sector was used as input to a visual teacher model for pre-training, and multi-level features were extracted.
[0015] The visual encoder maps the input power field image data into a sequence of feature vectors; the feature vectors are matched to the top-level codebook and the bottom-level codebook respectively through two quantizers to obtain discrete top-level visual labels and bottom-level visual labels.
[0016] The decoder reconstructs the multi-level features of the teacher model based on the discrete labels.
[0017] Furthermore, the semi-supervised fine-tuning of the large visual model includes setting and adjusting adaptive thresholds for detection targets at different scales, specifically including:
[0018] For the s-th scale, its global threshold The prediction confidence of the large visual model for unlabeled power image data at this scale is updated using an exponential moving average.
[0019] For the c-th category at the s-th scale, its local threshold The predicted probability of this category is normalized and then multiplied by the global threshold to obtain the result.
[0020] When the prediction confidence of unlabeled power image data at the s-th scale is higher than the local threshold of its corresponding category At that time, pseudo-labels for the data are generated to calculate the unsupervised loss at that scale.
[0021] Furthermore, the specific steps for updating the parameters of the large visual model using the SMLS fine-tuning strategy are as follows:
[0022] The large visual model is a Transformer backbone network. For any linear projection layer weight matrix in the Transformer backbone network, its update amount is decomposed into the product of three low-rank matrices.
[0023] The detection targets are divided into three categories according to scale: small, medium, and large, and are associated with small, medium, and large scale detection branches respectively.
[0024] During backpropagation, the gradients generated by each scale detection branch are used only to update specific matrices in their associated low-rank matrices, achieving scale-decoupled parameter updates.
[0025] Furthermore, the division of the detected targets by scale is specifically as follows: based on the ratio of the area of the target detection box to the image area in the target detection task, targets with a ratio less than 0.01 are classified as small-scale, targets with a ratio between 0.01 and 0.1 are classified as medium-scale, and targets with a ratio greater than 0.1 are classified as large-scale.
[0026] Furthermore, the method also includes a fairness regularization method, specifically:
[0027] The distribution of pseudo-labels selected at each scale was statistically analyzed;
[0028] The probability distribution of predictions for unlabeled data at various scales using a large computational vision model;
[0029] A fairness loss is constructed by minimizing the cross-entropy between the predicted distribution and the target distribution normalized by the pseudo-label distribution.
[0030] Furthermore, the pre-training of a large visual model on unlabeled power images using the masked image modeling task also includes a patch aggregation strategy, specifically:
[0031] The patch features of the intermediate layers of the Transformer backbone network are concatenated with the [CLS] token features of the final layer; a mask label prediction is performed on the concatenated features through a lightweight Transformer decoder; thus enabling the [CLS] token to aggregate global information.
[0032] The present invention discloses a large-scale, multi-scale, semi-supervised target detection system for power vision, comprising:
[0033] Data acquisition module: Acquires power image data, including labeled and unlabeled power images;
[0034] Multi-scale visual codebook construction module: Based on the vector quantization knowledge distillation framework, a multi-scale visual word segmenter is trained to construct a multi-scale visual codebook for power semantic enhancement. The multi-scale visual codebook includes a top-level codebook for encoding the global structure of power equipment and a bottom-level codebook for encoding the local features of equipment defects.
[0035] Pre-training module: Based on the multi-scale visual codebook, a large visual model is pre-trained on unlabeled power images using a masked image modeling task;
[0036] Semi-supervised fine-tuning module: Uses labeled and unlabeled power image data to perform semi-supervised fine-tuning on the large visual model; SMLS fine-tuning strategy is used to update the parameters of the large visual model;
[0037] Detection result output module: Input the power image to be detected into the updated visual large model and output the detection results of multi-scale targets.
[0038] The computer device of the present invention includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the multi-scale semi-supervised target detection method for a large-scale power vision model.
[0039] The present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of a multi-scale semi-supervised target detection method for a large-scale power vision model as described above.
[0040] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:
[0041] (1) Significantly improves detection accuracy and robustness: The multi-scale codebook with power semantic enhancement provides domain-related discrete semantic supervision for model pre-training, effectively bridging the semantic gap between general large models and power-specific scenarios. Combined with a multi-scale adaptive threshold mechanism, higher-quality pseudo-labels can be selected for targets of different sizes, thereby making full use of unlabeled data in semi-supervised fine-tuning and significantly improving the detection accuracy and robustness of various defects such as insulator cracks, conductor strand breaks, and missing hardware.
[0042] (2) Significantly reduce annotation costs and data dependence: The multi-scale semi-supervised fine-tuning framework requires only a small amount of labeled data to start. It can safely utilize massive amounts of unlabeled power images through a dynamic threshold strategy, greatly reducing the dependence on expensive and time-consuming professional data annotation.
[0043] (3) Achieving efficient deployment and resource conservation: The proposed SMLS parameter efficient fine-tuning strategy reduces the number of parameters that need to be trained to less than 1% of the total parameter fine-tuning while maintaining model performance, significantly reducing the memory usage and computational overhead during the fine-tuning stage. Combined with model pruning or quantization techniques, it can be easily deployed to edge devices to meet the real-time requirements of power line inspection, saving computing resources and energy consumption.
[0044] (4) Mitigating class imbalance problem: Class fairness regularization is integrated into the multi-scale adaptive threshold framework. By monitoring and adjusting the generation ratio of pseudo-labels of different classes, the "long tail distribution" problem commonly found in power defect data is effectively mitigated, and the recall rate of the model for rare defect classes is improved. Attached Figure Description
[0045] Figure 1 This is a flowchart of the method described in this invention.
[0046] Figure 2 This is a system structure diagram of the present invention. Detailed Implementation
[0047] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0048] Example 1
[0049] This embodiment provides a multi-scale semi-supervised target detection method and system for large-scale power vision models, including the following steps:
[0050] (11) Acquire power image data, including labeled power images and unlabeled power images;
[0051] (12) A multi-scale visual word segmenter is trained based on the vector quantization knowledge distillation framework to construct a multi-scale visual codebook for power semantic enhancement. The multi-scale visual codebook includes a top-level codebook for encoding the global structure of power equipment and a bottom-level codebook for encoding the local features of equipment defects.
[0052] The multi-scale visual segmenter comprises a shared visual Transformer encoder and two independent quantizers. The encoder maps the input power image into a sequence of feature vectors, which are then mapped to two independent codebooks via a top-level quantizer and a bottom-level quantizer, respectively. The top-level codebook encodes the global structural semantics of power equipment (e.g., "insulator string," "conductor"), while the bottom-level codebook encodes the local feature semantics of equipment defects (e.g., "crack texture," "rust spots"). The training objective is to reconstruct the corresponding-level features extracted by a pre-trained teacher model (e.g., CLIP) through the decoder, enabling the visual codebook to learn multi-level discrete semantic concepts relevant to the power field.
[0053] (13) Based on the multi-scale visual codebook, a large visual model is pre-trained on unlabeled power images using a masked image modeling task;
[0054] Using the multi-scale visual segmenter trained in step (12) as the reconstruction target, a visual Transformer backbone network (such as ViT) is pre-trained on a large number of unlabeled power images using the Masked Image Modeling (MIM) task. Randomly masking portions of the image blocks, the Transformer backbone network predicts the visual labels corresponding to the masked positions based on contextual information. These visual labels come from either the top-level or bottom-level codebook. This allows the large visual model to learn the essential features and structural patterns of the power images.
[0055] (14) Semi-supervised fine-tuning of the large visual model was performed using labeled and unlabeled power image data; the parameters of the large visual model were updated using the SMLS fine-tuning strategy.
[0056] Semi-supervised fine-tuning of a pre-trained large-scale visual model is performed using a small amount of labeled power data and a large amount of unlabeled power data. Within the fine-tuning framework, targets are categorized into three scales—small (ratio < 0.01), medium (0.01 ≤ ratio ≤ 0.1), and large (ratio > 0.1)—based on the ratio of the target detection box area to the image area in the target detection task. For each scale s:
[0057] global threshold Based on the prediction confidence of the model for all unlabeled samples at this scale within the current batch, the model is dynamically updated using the exponential moving average (EMA) to reflect the overall learning status of the model at this scale.
[0058] Local threshold For each category c, the predicted probability is normalized and then multiplied by a global threshold. This threshold takes into account the differences in learning difficulty between different categories at different scales.
[0059] When an unlabeled sample is predicted to be class c at scale s, and its confidence level is higher than 100%, the unlabeled sample is predicted to be class c at scale s. Only when the time is right will pseudo-labels be generated to calculate the unsupervised loss at that scale.
[0060] This embodiment employs the SMLS (Small-Medium-Large Scale Adaptation) fine-tuning strategy. All parameters of the pre-trained Transformer backbone network are frozen, and only in the multi-scale detection heads connected to the Transformer backbone network, a set of trainable low-rank adaptation matrices is introduced for each scale branch. For a weight matrix W ∈ R^{d×k} in the original detection head, its update ΔW is decomposed into the product of three consecutive low-rank matrices: ,in , , , and rank During backpropagation, a gradient routing mechanism is designed so that the loss gradients from the small, medium, and large-scale detection branches are used only to update their respective S, M, and L matrices, thereby achieving scale decoupling and efficient utilization of parameter updates.
[0061] (15) Input the power image to be detected into the updated visual large model and output the detection results of multi-scale targets.
[0062] The power inspection image to be detected is input into the finely adjusted and updated visual large model. The model outputs prediction boxes and their categories at all scales. After non-maximum suppression (NMS), the final multi-scale target detection results are obtained.
[0063] Example 2
[0064] This embodiment takes the detection of insulator cracks and conductor strand breaks in transmission line inspection images as an example to illustrate the specific implementation process of the target detection method of the present invention.
[0065] (21) Data Acquisition: Collect power scene image data and annotate 10,000 images to obtain labeled images, including bounding boxes and category labels such as "insulator", "conductor", "insulator-crack", and "conductor-broken strand". There are 400,000 unlabeled inspection images. All images are uniformly scaled to a resolution of 224×224 or 384×384.
[0066] (22) Constructing a multi-scale visual codebook:
[0067] Teacher model: The CLIP-ViT-B / 16 model was selected as the teacher model.
[0068] The word segmenter architecture is a multi-scale visual word segmenter encoder using ViT-B / 16. The top-level codebook size is 8192, the bottom-level codebook size is 8192, and the embedding dimension is 32 for both.
[0069] Training: Using an unlabeled electricity dataset, the optimization objective is the VQ-KD loss as shown in the formula below.
[0070]
[0071] in Indicates the input image; This represents the training dataset; This indicates the number of blocks the image is divided into (for 2^24). 224 images, ) Let i be the output vector at the i-th position of the decoder; This is the feature vector extracted by the teacher model (such as CLIP) for the i-th image patch; Indicates cosine similarity; This is the output vector at the i-th position of the visual word segmenter encoder; For the first in the codebook Each codebook vector; This indicates the codebook index selected through nearest neighbor lookup; express Normalization operation; This indicates that the gradient operation is stopped (and treated as a constant during backpropagation). express The square of the norm.
[0072] After training for 100 epochs, the word segmenter maps global structures such as tower bodies and insulator strings in the power image to the top-level codebook, and maps local features such as crack textures and metallic luster to the bottom-level codebook.
[0073] (23) Visual large model pre-training:
[0074] Backbone network: ViT-B / 16 is used as the backbone.
[0075] Pre-training task: A Masked Image Modeling (MIM) task is used, randomly masking 40% of the image patches. For each masked location, if it is located in the center region of an object in the image, its corresponding top-level codebook label is predicted; if it is located on the edge of an object or in a region of textural detail, its corresponding bottom-level codebook label is predicted. Pre-training is performed for 300 epochs on 400,000 unlabeled images.
[0076] (24) Multi-scale adaptive threshold semi-supervised fine-tuning:
[0077] Scale division: Based on the ratio of the area of the target detection box to the area of the image in the target detection task, the target is divided into three scales: small (ratio < 0.01), medium (0.01 ≤ ratio ≤ 0.1), and large (ratio > 0.1).
[0078] Threshold initialization: Global thresholds at all scales are initialized to 0.5. Local thresholds are initialized based on the initial model's predictions on labeled data.
[0079] Pseudo-label generation: For each unlabeled image, weak enhancement (random cropping and flipping) and strong enhancement (RandAugment) are performed. After the weakly enhanced image is predicted by the large visual model, if the confidence of the predicted bounding box in the corresponding scale and category is higher than the current threshold τ_t^s(c), it is retained as a pseudo-label.
[0080] Loss calculation: Total loss L = L_s + λ_u * L_u + λ_f * L_f.
[0081] Where L_s represents the standard detection loss for labeled data (e.g., Focal Loss). L_u is the unsupervised loss, which calculates the detection loss between strongly enhanced image predictions and pseudo-labels, and is only calculated for predictions above a threshold. L_f is the fairness loss, calculated in the same way as SAF in FreeMatch, but with separate statistics for the class distributions at three scales. λ_u and λ_f represent the corresponding loss weights, typically set to 1 and 0.01, respectively.
[0082] (25) Efficient parameter fine-tuning
[0083] SMLS fine-tuning strategy implementation: In the Faster R-CNN style detector head, inject low-rank matrices into the branch linear layers of each scale in the RPN and ROI Head. Set the rank m=8.
[0084] Gradient routing: During backpropagation, gradient separation is achieved through custom gradient hooks. For example, the loss gradient generated by a small-scale target is only used to update the S matrix, which is labeled "small-scale specific," while the gradients of the M and L matrices are set to zero. The same applies to medium-scale and large-scale targets.
[0085] Training configuration: Using the AdamW optimizer, initial learning rate 5e-4, batch size 32. Only the parameters introduced by SMLS and the newly added layers in the detector head are trained; the backbone network remains frozen, and fine-tuning is performed for 50 epochs.
[0086] (26) Target detection reasoning
[0087] The inspection image to be detected is input into the finely adjusted and updated large-scale visual model, which outputs multi-scale prediction results. Cross-scale nonmaximum suppression is applied to finally output the defect location and category, which are then visualized and overlaid on the original image.
[0088] Example 3
[0089] This embodiment provides a large-scale, semi-supervised target detection system for power vision, comprising:
[0090] Data acquisition module: Acquires power image data, including labeled and unlabeled power images;
[0091] Multi-scale visual codebook construction module: Based on the vector quantization knowledge distillation framework, a multi-scale visual word segmenter is trained to construct a multi-scale visual codebook for power semantic enhancement. The multi-scale visual codebook includes a top-level codebook for encoding the global structure of power equipment and a bottom-level codebook for encoding the local features of equipment defects.
[0092] Pre-training module: Based on the multi-scale visual codebook, a large visual model is pre-trained on unlabeled power images using a masked image modeling task;
[0093] Semi-supervised fine-tuning module: Uses labeled and unlabeled power image data to perform semi-supervised fine-tuning on the large visual model; SMLS fine-tuning strategy is used to update the parameters of the large visual model;
[0094] Detection result output module: Input the power image to be detected into the updated visual large model and output the detection results of multi-scale targets.
[0095] This embodiment is based on the same inventive concept as Embodiment 1, and will not be repeated here.
[0096] Example 4
[0097] A computer device includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the multi-scale semi-supervised target detection method and system for large-scale power vision models as described above.
[0098] Example 5
[0099] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned power vision large-scale multi-scale semi-supervised target detection method and system.
Claims
1. A multi-scale semi-supervised target detection method using a large-scale power vision model, characterized in that, include: Acquire power image data, including labeled and unlabeled power images; A multi-scale visual word segmenter is trained based on a vector quantization knowledge distillation framework to construct a multi-scale visual codebook for power semantic enhancement. The multi-scale visual codebook includes a top-level codebook for encoding the global structure of power equipment and a bottom-level codebook for encoding local features of equipment defects. Based on the multi-scale visual codebook, a large visual model is pre-trained on unlabeled power images using a masked image modeling task. Semi-supervised fine-tuning of the large visual model was performed using labeled and unlabeled power image data; the parameters of the large visual model were updated using the SMLS fine-tuning strategy. The image of the power source to be detected is input into the updated visual large model, and the detection results of multi-scale targets are output.
2. The method for multi-scale semi-supervised target detection using a large-scale power vision model according to claim 1, characterized in that, The multi-scale visual word segmenter includes a visual encoder and two quantizers. The specific steps for training the multi-scale visual word segmenter encoder are as follows: Image data from the power sector was used as input to a visual teacher model for pre-training, and multi-level features were extracted. The visual encoder maps the input power image data into a sequence of feature vectors; The feature vectors are matched to the top-level codebook and the bottom-level codebook respectively by two quantizers to obtain discrete top-level visual labels and bottom-level visual labels; The decoder reconstructs the multi-level features of the teacher model based on the discrete labels.
3. The method for multi-scale semi-supervised target detection using a large-scale power vision model according to claim 1, characterized in that, The semi-supervised fine-tuning of the large visual model includes setting and adjusting adaptive thresholds for detection targets at different scales, specifically including: For the s-th scale, its global threshold The prediction confidence of the large visual model for unlabeled power image data at this scale is updated using an exponential moving average. For the c-th category at the s-th scale, its local threshold The predicted probability of this category is normalized and then multiplied by the global threshold to obtain the result. When the prediction confidence of unlabeled power image data at the s-th scale is higher than the local threshold of its corresponding category At that time, pseudo-labels for the data are generated to calculate the unsupervised loss at that scale.
4. The multi-scale semi-supervised target detection method for a large-scale power vision model according to claim 1, characterized in that, The specific steps for updating the parameters of the large visual model using the SMLS fine-tuning strategy are as follows: The large visual model is a Transformer backbone network. For any linear projection layer weight matrix in the Transformer backbone network, its update amount is decomposed into the product of three low-rank matrices. The detection targets are divided into small-scale, medium-scale, and large-scale according to their scale, and are respectively associated with small-scale, medium-scale, and large-scale detection branches; During backpropagation, the gradients generated by each scale detection branch are used only to update specific matrices in their associated low-rank matrices, achieving scale-decoupled parameter updates.
5. The multi-scale semi-supervised target detection method for a large-scale power vision model according to claim 4, characterized in that, The division of the detected targets by scale is specifically as follows: based on the ratio of the area of the target detection box to the area of the image in the target detection task, targets with a ratio less than 0.01 are classified as small scale, targets with a ratio between 0.01 and 0.1 are classified as medium scale, and targets with a ratio greater than 0.1 are classified as large scale.
6. The method for multi-scale semi-supervised target detection using a large-scale power vision model according to claim 1, characterized in that, The method also includes a fairness regularization method, specifically: The distribution of pseudo-labels selected at each scale was statistically analyzed; The probability distribution of predictions for unlabeled data at various scales using a large computational vision model; A fairness loss is constructed by minimizing the cross-entropy between the predicted distribution and the target distribution normalized by the pseudo-label distribution.
7. The method for multi-scale semi-supervised target detection using a large-scale power vision model according to claim 1, characterized in that, The task of pre-training a large visual model on unlabeled power images using masked image modeling also includes a patch aggregation strategy, specifically: The patch features of the intermediate layers of the Transformer backbone network are concatenated with the [CLS] token features of the final layer; a mask label prediction is performed on the concatenated features through a lightweight Transformer decoder; thus enabling the [CLS] token to aggregate global information.
8. A multi-scale semi-supervised target detection system for a large-scale vision model of power systems, characterized in that, include: Data acquisition module: Acquires power image data, including labeled and unlabeled power images; Multi-scale visual codebook construction module: Based on the vector quantization knowledge distillation framework, a multi-scale visual word segmenter is trained to construct a multi-scale visual codebook for power semantic enhancement. The multi-scale visual codebook includes a top-level codebook for encoding the global structure of power equipment and a bottom-level codebook for encoding the local features of equipment defects. Pre-training module: Based on the multi-scale visual codebook, a large visual model is pre-trained on unlabeled power images using a masked image modeling task; Semi-supervised fine-tuning module: Uses labeled and unlabeled power image data to perform semi-supervised fine-tuning on the large visual model; SMLS fine-tuning strategy is used to update the parameters of the large visual model; Detection result output module: Input the power image to be detected into the updated visual large model and output the detection results of multi-scale targets.
9. A computer device, characterized in that, The method includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of a multi-scale semi-supervised object detection method for a large-scale power vision model as described in any one of claims 1-7.
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 multi-scale semi-supervised target detection method for a large-scale power vision model as described in any one of claims 1-7.