Image annotation model construction method and system for industrial scene
By adopting the RAM model-based image annotation model construction method in industrial scenarios, problems such as high cost and unstable image annotation in industrial scenarios are solved, and efficient and accurate image annotation is achieved, providing support for industrial inspection and quality control.
Patent Information
- Application Number
- CN202411828241.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-04-29
AI Technical Summary
In the prior art, image labeling has problems such as high manual labeling cost, unstable labeling quality, limitations of labeling tools, poor data quality and difficult to obtain large-scale labeling data in industrial scenarios.
The image annotation model construction method based on RAM model is used to preprocess the image data of industrial scenarios and their text description data, and divide them into training sets and test sets. The training set is input into the pre-constructed RAM model to identify anything for training, obtain the trained image annotation model, and verify the model through the test set, and finally obtain the image annotation model used for industrial scenarios.
It improves the accuracy and efficiency of image labeling, and can automatically generate high-quality labeling data in complex industrial environments, and is suitable for application scenarios such as industrial inspection and quality control.
Smart Images
Figure CN119904710A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image annotation, and in particular to a method and system for constructing an image annotation model for industrial scenarios. Background Art
[0002] The development of modern industry aims to build a new industrial ecosystem oriented towards intelligent factory production and based on the automatic flow of data. Machine vision, as an important link connecting people, data and machines in modern industry, is an important technology for industrial transformation, bringing simultaneous improvements in production cycle, quality and efficiency to the manufacturing industry. With the development of information technology, fieldbus technology and deep learning technology, the proportion of visual data in industrial data continues to increase. The development of visual technology plays a key role in mining industrial data and obtaining useful information.
[0003] With the breakthrough development of artificial intelligence technology in recent years, especially the emergence of large natural language models, the processing capabilities of multimodal data including visual data have been improved and the application boundaries have been extended. The industrial vertical domain visual model is a model trained with general industrial scenarios (general domains) to process visual tasks in specific industrial fields or vertical fields. The combination of industrial big data and artificial intelligence technology helps to realize the digital and intelligent management of the entire life cycle of industrial production. By identifying, positioning, detecting and analyzing various target objects in the production process, problems can be discovered in a timely manner and optimized and adjusted to improve production efficiency and quality. Existing image annotation methods include manual annotation and automatic annotation methods, but they still have some technical drawbacks. These problems may limit their effectiveness and efficiency in practical applications, mainly including:
[0004] 1. High cost of manual labeling: Manual labeling requires a lot of time and manpower, especially in the case of large-scale data sets; and high-quality labeling usually requires professional labelers, which increases the overall cost of the project.
[0005] 2. Unstable labeling quality: Manual labeling is highly subjective, and different labelers may have different understandings of the same image, resulting in inconsistent labeling results. At the same time, long-term labeling work may cause labelers to become fatigued and negligent, affecting the labeling quality.
[0006] 3. Limitations of annotation tools and platforms: Existing annotation tools may lack advanced features, such as automatic annotation suggestions, real-time quality checks, etc.; and the user interfaces of some annotation tools are not user-friendly, which affects annotation efficiency.
[0007] 4. In practical applications, due to the particularity of industrial scenarios, obtaining high-quality industrial image data may be challenging. Poor data quality and inaccurate annotations may lead to a decrease in the generalization performance of the annotation model in actual scenarios.
[0008] 5. Since target detection has high requirements for data, large-scale labeled data is needed to train the model, which may not be easy to obtain in industrial scene tasks. In addition, in the industrial field, there are often problems such as small size and complex shape of the detection targets, which may lead to poor performance of the labeled model in small target detection and complex scenes. Summary of the invention
[0009] The purpose of the present invention is to provide a method and system for constructing an image annotation model for industrial scenarios, aiming to solve the above-mentioned problems in the prior art.
[0010] An embodiment of the present invention provides a method for constructing an image annotation model for industrial scenarios, comprising:
[0011] Acquire image data and corresponding text description data in an industrial scene, preprocess the image data and the corresponding text description data, and divide the processed data into a training set and a test set; and
[0012] The training set is input into a pre-built RAM model for recognizing anything for training to obtain a trained image annotation model, and the trained image annotation model is verified using the test set to obtain a final image annotation model.
[0013] An embodiment of the present invention provides an image annotation model construction system for industrial scenarios, including:
[0014] A data module, used to obtain image data and corresponding text description data in an industrial scene, preprocess the image data and the corresponding text description data, and divide the processed data into a training set and a test set; and
[0015] A construction module is used to input the training set into a pre-built RAM model for recognizing anything for training to obtain a trained image annotation model, and to verify the trained image annotation model using the test set to obtain a final image annotation model.
[0016] An embodiment of the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the above-mentioned method for constructing an image annotation model for industrial scenarios when executed by the processor.
[0017] An embodiment of the present invention further provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps of the above-mentioned method for constructing an image annotation model for industrial scenarios are implemented.
[0018] The use of the embodiments of the present invention may include the following beneficial effects: the embodiments of the present invention construct an image annotation model for industrial scenarios based on the RAM model in the vertical domain vision model, thereby effectively solving specific visual tasks in this field and improving the accuracy and efficiency of detection, recognition or prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0020] Figure 1 is a flow chart of a method for constructing an image annotation model for industrial scenarios according to an embodiment of the present invention;
[0021] Figure 2 It is a technical roadmap of the vertical domain visual RAM model of an embodiment of the present invention;
[0022] Figure 3 Schematic diagram of an image annotation model building system for industrial scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION
[0023] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.
[0024] Method Embodiment
[0025] According to an embodiment of the present invention, a method for constructing an image annotation model for industrial scenes is provided. Figure 1 is a flow chart of a method for constructing an image annotation model for industrial scenarios according to an embodiment of the present invention. Figure 1 As shown, the method for constructing an image annotation model for industrial scenes according to an embodiment of the present invention specifically includes:
[0026] Step S101, obtaining image data and corresponding text description data in an industrial scene, preprocessing the image data and corresponding text description data, and dividing the processed data into a training set and a test set, specifically includes:
[0027] Performing data cleaning, data enhancement and standardization processing on the image data;
[0028] Performing word segmentation, stop word removal and stemming processing on the corresponding text description data;
[0029] Wherein, the data cleaning includes removing duplicate data, removing data with poor image quality, and excluding images that are not relevant to the annotation task;
[0030] The data enhancement includes random rotation and flipping, scale transformation and random cropping, illumination transformation, color transformation, Mosaic enhancement and Mixup enhancement;
[0031] Step S102, inputting the training set into a pre-built RAM model for recognizing anything for training to obtain a trained image annotation model, and using the test set to verify the trained image annotation model to obtain a final image annotation model, specifically includes:
[0032] The training set is input into a pre-built RAM model, image data is feature extracted through an image encoder to generate an image feature vector, text data is feature extracted through a pre-trained word embedding model in a text encoder to obtain a label embedding vector, the image feature vector and the label embedding vector are feature fused through an image-label interactive encoder using a multimodal fusion technique, and the fused features are mapped to a label space through a decoder to generate a final labeling result;
[0033] The method further comprises:
[0034] The final image annotation model is embedded into the industrial inspection system, and the industrial image annotation task is performed by using the final image annotation model.
[0035] The following is a specific example of the method for constructing an image annotation model for an industrial scene according to an embodiment of the present invention. Figure 2 As shown, the above technical solution of the embodiment of the present invention is described in detail.
[0036] Image annotation is an important task in industrial inspection and quality control. Traditional manual annotation methods are time-consuming and error-prone, and the performance of existing automated annotation methods in complex industrial scenarios is often unsatisfactory. Therefore, an efficient and accurate image annotation model construction method is needed to meet the needs of industrial scenarios. An embodiment of the present invention relates to a method for constructing an image annotation model for industrial scenarios based on a RAM (Recognize Anything Model) model, aiming to improve the accuracy and efficiency of image annotation. Through its powerful multimodal understanding and generation capabilities, the RAM model can automatically generate high-quality annotation data in complex industrial environments, thereby providing support for application scenarios such as industrial inspection and quality control.
[0037] RAM (Recognize Anything Model) is a model that can recognize anything, supporting Chinese and English, that is, a powerful basic model for image annotation. The RAM model is a multimodal deep learning model designed to extract information from images and text and generate high-quality annotations. RAM has taken an important step for large models in computer vision, demonstrating the ability to recognize any common category with high accuracy in zero shots. The RAM model introduces a new paradigm for image annotation, using large-scale image-text pairs for training instead of manual annotation. The development of RAM includes four key steps. First, through automatic text semantic parsing, large-scale image labels without annotations are obtained; then, a preliminary model is trained by unified title and labeling tasks, supervised by original text and parsed labels respectively; third, a data engine is used to generate additional annotations and clean up erroneous annotations; finally, the model is retrained with the processed data and fine-tuned with a smaller but higher quality dataset.
[0038] RAM uses text semantic parsing to extract image tags, providing a large number of tags without the need for expensive manual annotation. RAM's overall architecture is similar to Tag2Text and includes three key modules: an image encoder for feature extraction, an image tag recognition decoder for tagging, and a text encoder-decoder for text generation.
[0039] RAM uses the swing-Transformer as the image encoder because it outperforms the standard Vision Transformer (ViT) in the visual language and tagging domain. The text generation encoder-decoder consists of 12 layers of Transformers, and the label recognition decoder consists of 2 layers of Transformers.
[0040] RAM also uses CLIP’s off-the-shelf text encoder to generate text label queries through prompt integration. CLIP’s image encoder is used to extract image features, and the model’s recognition ability for unseen categories is enhanced by aligning image-text features.
[0041] Specifically, the RAM model used in the embodiment of the present invention is a powerful multimodal model that can recognize and understand various types of data, especially images and text. It consists of multiple key components that work together to enable the model to generate high-quality labels and descriptions. The following are the main components of the RAM model:
[0042] 1. Image Encoder: Extract high-level features from the input image, which can reflect the main content and structure of the image, and generate feature vectors of the image, which contain the visual information of the image.
[0043] 2. Tag Embedding: Convert tags into embedding vectors that can represent the semantic information of tags. That is, use pre-trained word embedding models (such as Word2Vec, GloVe, or BERT) to generate embedding vectors of tags, which contain the semantic information of tags.
[0044] 3. ImageTag Interaction Encoder: Captures the interaction between image features and tag embeddings to enhance the model’s understanding of image content. Uses multimodal fusion techniques (such as concatenation, weighted summation, and attention mechanisms) to combine image features and tag embeddings to generate an interaction matrix that represents the interaction between image features and tag embeddings.
[0045] 4. Decoder: Located in the output layer of the model, it is responsible for converting image features and label embeddings (text features) into high-quality human-readable labels and description texts through multimodal fusion technology. Sequence generation models (such as RNN, LSTM, Transformer) are usually used to generate image labels and description texts.
[0046] The specific steps of the embodiment of the present invention are as follows:
[0047] 1. Data preparation
[0048] ① Collect data: Collect a large amount of images and corresponding text description data. These data can be obtained from open source data sets, obtain relevant images from the Internet through crawler technology, or negotiate with partners and suppliers to obtain image data in professional fields. In an embodiment of the present invention, it is necessary to collect a large amount of industrial scene image data and corresponding text descriptions. The images include different types of defect images, normal images, etc., so as to ensure the diversity and representativeness of the data to cover various industrial inspection scenarios. For example, obtain a picture of a factory production line and its text description "On the factory production line, technical workers are busy operating various computers, conveyors and other electronic devices and equipment."
[0049] ②Data preprocessing:
[0050] a. Data cleaning is a key step to ensure data quality. This includes removing duplicate data, processing samples with poor image quality, and excluding images that are not relevant to the task, which helps improve the model's adaptability to real-world scenarios.
[0051] Among them, image preprocessing: standardization of images, including resizing, cropping, normalization, etc.
[0052] Text preprocessing: segment the text, remove stop words, stem the text, and generate a vocabulary.
[0053] b. Image data enhancement plays a vital role in training models, especially in the development of large vertical models. A good data enhancement strategy can effectively improve the generalization ability and robustness of the model, making it perform better in real scenes. The data enhancement method used in the embodiment of the present invention is as follows:
[0054] (1) Random rotation and flipping
[0055] Randomly rotating and flipping images helps the model learn objects at different angles and directions, which is crucial for the diversity of objects in real scenes. By introducing such transformations, the model can better adapt to object detection from different perspectives.
[0056] (2) Scaling and cropping
[0057] Introducing scale transformation and random cropping helps simulate the different sizes and positions of targets in the image. This data augmentation method makes the model more adaptable to targets of different scales and spatial ranges.
[0058] (3) Lighting transformation
[0059] Simulate target detection scenarios under different lighting conditions, including changes in brightness and contrast. By introducing lighting changes, the model can better adapt to lighting changes in different environments and improve the robustness of the model.
[0060] (4) Color change
[0061] By adjusting the colors of images, such as changing hue, saturation, and brightness, we can increase the diversity of the data. This transformation helps the model better adapt to different visual environments and conditions.
[0062] (5) Mosaic enhancement
[0063] Mosaic data augmentation is a method of splicing multiple images into a new image to increase the diversity of training data. Specifically, Mosaic randomly crops and splices several different images together to form a mixed image. The advantage of this method is that the model needs to process information from multiple different scenes at the same time, so as to better learn rich feature representations. Mosaic provides more scene changes, allowing the model to better adapt to various complex scenes and better cope with complex situations in the real world.
[0064] (6) Mixup enhancement
[0065] Mixup is a data augmentation method that mixes two images together to create new training samples by linearly interpolating images and labels. The core idea of this method is to mix different samples at the image and label level to generate training data with richer information. Mixup introduces differences between samples, which helps to improve the generalization ability of the model; in addition, Mixup has a certain degree of robustness against adversarial attacks, which can improve the security of the model.
[0066] The embodiment of the present invention constructs a rich, high-quality data set by comprehensively applying the above data processing and image data enhancement methods, providing sufficient and diverse training data for the RAM model, thereby achieving better performance in real scenarios; and Mosaic and Mixup, as advanced image data enhancement technologies, provide more diverse and more robust data for model training. Their advantages are increasing the diversity of training data and improving the generalization ability of the model, so that the RAM model can achieve more outstanding performance in real scenarios.
[0067] 2. Model architecture design
[0068] ①Multimodal encoder:
[0069] Image encoder: Use swing-Transformer to encode the image and extract image features.
[0070] Text encoder: Use pre-trained word embedding models (such as BERT, RoBERTa, DeBERTa, T5, etc.) to encode text and extract text features.
[0071] ② Multimodal fusion layer: fuses image features and text features, such as splicing, weighted summation, attention mechanism, etc.
[0072] ③Decoder: Map the fused features to the annotation space and generate the final annotation results.
[0073] 3. RAM model training process
[0074] A. Initialization parameters: Initialize the parameters of the RAM model image encoder and text encoder.
[0075] B. Define loss function: Select appropriate loss function according to application scenario or data type, such as cross entropy loss, mean square error loss, etc., to measure the difference between the annotation predicted by the model and the actual annotation. Or introduce alignment loss function, such as contrastive loss, to enhance the alignment effect between different modalities.
[0076] C. Forward propagation: Input image and text data, extract features through the multimodal encoder; pass the features to the multimodal fusion layer to generate fused features; generate annotation results through the decoder.
[0077] D. Calculate loss: Compare the generated annotation results with the true annotations and calculate the loss value.
[0078] E. Back propagation: Based on the loss value, the model parameters are updated through the back propagation algorithm to optimize the model performance.
[0079] F. Iterative training: Repeat the process of forward propagation, loss calculation, and backpropagation until the model converges or reaches a predetermined number of training rounds.
[0080] 4. Annotation model evaluation and optimization
[0081] ① Validation set evaluation: Evaluate the performance of the model on the validation set, using evaluation indicators such as accuracy, recall, F1 score, etc.
[0082] ② Hyperparameter tuning: Adjust hyperparameters such as learning rate, batch size, and number of layers to optimize model performance.
[0083] ③Regularization: Use regularization techniques (such as L1 and L2 regularization) to prevent overfitting.
[0084] ④ Early stopping strategy: If the performance on the validation set no longer improves, training can be stopped early to avoid overfitting of the model.
[0085] 5. Annotation model deployment
[0086] ① Model saving: Save the trained annotation model weights for subsequent use.
[0087] ② Reasoning: Deploy the model to practical applications and annotate new industrial images.
[0088] ③Continuous optimization: Based on the feedback from the actual application of the annotation model in industrial scenarios, the annotation model is continuously optimized to improve the accuracy and efficiency of annotation. For example, an online learning mechanism is introduced to enable the model to be continuously updated and optimized based on user feedback, and / or a reinforcement learning method is combined to enable the model to continuously learn and improve during the interaction process.
[0089] Preferably, the embodiments of the present invention can also improve the RAM model according to the actual application scenario, for example, by introducing multi-head attention or adaptive attention mechanism to better capture the relationship between different modalities. Introduce a dynamic fusion mechanism to dynamically adjust the fusion strategy according to the characteristics of the input data to improve the flexibility and robustness of the model. Or reduce the number of parameters and computational complexity of the model through model pruning and quantization techniques so that it can run efficiently on resource-constrained devices. It is also possible to design a lightweight model architecture, such as using lightweight networks such as MobileNet and ShuffleNet to reduce inference time and memory usage.
[0090] When processing sequence data, consider introducing historical context information, such as the images of the previous frames or the text of the previous sentences, to enhance the model's spatiotemporal understanding ability. And combine environmental information, such as geographic location, time, etc., to provide richer background knowledge.
[0091] System Example
[0092] According to an embodiment of the present invention, a system for constructing an image annotation model for industrial scenes is provided. Figure 3 is a schematic diagram of an image annotation model construction system for industrial scenarios according to an embodiment of the present invention. Figure 3 As shown, the image annotation model construction system for industrial scenes according to an embodiment of the present invention specifically includes:
[0093] The data module 30 is used to obtain image data and corresponding text description data in the industrial scene, preprocess the image data and the corresponding text description data, and divide the processed data into a training set and a test set, specifically for:
[0094] Performing data cleaning, data enhancement and standardization processing on the image data;
[0095] Performing word segmentation, stop word removal and stemming processing on the corresponding text description data;
[0096] Wherein, the data cleaning includes removing duplicate data, removing data with poor image quality, and excluding images that are not relevant to the annotation task;
[0097] The data enhancement includes random rotation and flipping, scale transformation and random cropping, illumination transformation, color transformation, Mosaic enhancement and Mixup enhancement;
[0098] The construction module 32 is used to input the training set into the pre-constructed RAM model for recognizing anything for training, to obtain a trained image annotation model, and to verify the trained image annotation model using the test set to obtain a final image annotation model, which is specifically used for:
[0099] The training set is input into a pre-built RAM model, image data is feature extracted through an image encoder to generate an image feature vector, text data is feature extracted through a pre-trained word embedding model in a text encoder to obtain a label embedding vector, the image feature vector and the label embedding vector are feature fused through an image-label interactive encoder using a multimodal fusion technique, and the fused features are mapped to a label space through a decoder to generate a final labeling result;
[0100] The system further comprises:
[0101] The application module is used to embed the final image annotation model into the industrial inspection system and perform industrial image annotation tasks through the final image annotation model.
[0102] The embodiment of the present invention is a system embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.
[0103] In summary, the embodiment of the present invention provides a method for constructing an image annotation model for industrial scenarios based on a RAM model. By combining multimodal understanding and generation technology, efficient and accurate image annotation is achieved, providing strong support for industrial inspection and quality control. The specific beneficial effects include the following:
[0104] 1. High accuracy: Through the multimodal understanding and generation capabilities of the RAM model, the annotation model can automatically generate high-quality annotation data, significantly improving the accuracy of annotation.
[0105] 2. Efficiency: Automated labeling methods greatly reduce the time and cost of manual labeling and improve work efficiency.
[0106] 3. Strong adaptability: The annotation model can demonstrate good performance in a variety of industrial scenarios and has strong adaptability and robustness.
[0107] 4. Scalability: The method framework proposed in the embodiment of the present invention is flexible and can be expanded and optimized according to specific needs, and is suitable for different industrial detection tasks.
[0108] Device Example 1
[0109] An embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps described in the method embodiment when executed by the processor.
[0110] Device Example 2
[0111] An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps described in the method embodiment are implemented.
[0112] The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk or optical disk, etc.
[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for constructing an image annotation model for industrial scenes, characterized in that include: Acquire image data and corresponding text description data in an industrial scene, preprocess the image data and the corresponding text description data, and divide the processed data into a training set and a test set; as well as The training set is input into a pre-built RAM model for recognizing anything for training to obtain a trained image annotation model, and the trained image annotation model is verified using the test set to obtain a final image annotation model.
2. The method according to claim 1, characterized in that The method further comprises: The final image annotation model is embedded into the industrial inspection system, and the industrial image annotation task is performed by using the final image annotation model.
3. The method according to claim 1, characterized in that Preprocessing the image data and its corresponding text description data specifically includes: Performing data cleaning, data enhancement and standardization processing on the image data; Performing word segmentation, stop word removal and stemming processing on the corresponding text description data; Wherein, the data cleaning includes removing duplicate data, removing data with poor image quality, and excluding images that are not relevant to the annotation task; The data enhancement includes random rotation and flipping, scale transformation and random cropping, illumination transformation, color transformation, mosaic enhancement and mixup enhancement.
4. The method according to claim 1, characterized in that: Inputting the training set into the pre-built RAM model for identifying anything for training specifically includes: The training set is input into a pre-built RAM model, and image data is subjected to feature extraction through an image encoder to generate an image feature vector. The text data is subjected to feature extraction through a pre-trained word embedding model in a text encoder to obtain a label embedding vector. The image feature vector and the label embedding vector are subjected to feature fusion by using an image-label interactive encoder using a multimodal fusion technique, and the fused features are mapped to the annotation space through a decoder to generate a final annotation result.
5. A system for building an image annotation model for industrial scenes, characterized in that include: A data module, used to obtain image data and corresponding text description data in an industrial scene, preprocess the image data and the corresponding text description data, and divide the processed data into a training set and a test set; as well as A construction module is used to input the training set into a pre-built RAM model for recognizing anything for training to obtain a trained image annotation model, and to verify the trained image annotation model using the test set to obtain a final image annotation model.
6. The system according to claim 5, characterized in that The system further comprises: The application module is used to embed the final image annotation model into the industrial inspection system and perform industrial image annotation tasks through the final image annotation model.
7. The system according to claim 5, characterized in that The data module is specifically used for: Performing data cleaning, data enhancement and standardization processing on the image data; Performing word segmentation, stop word removal and stemming processing on the corresponding text description data; Wherein, the data cleaning includes removing duplicate data, removing data with poor image quality, and excluding images that are not relevant to the annotation task; The data enhancement includes random rotation and flipping, scale transformation and random cropping, illumination transformation, color transformation, mosaic enhancement and mixup enhancement.
8. The system according to claim 5, characterized in that The building blocks are specifically used for: The training set is input into a pre-built RAM model, and image data is subjected to feature extraction through an image encoder to generate an image feature vector. The text data is subjected to feature extraction through a pre-trained word embedding model in a text encoder to obtain a label embedding vector. The image feature vector and the label embedding vector are subjected to feature fusion by using an image-label interactive encoder using a multimodal fusion technique, and the fused features are mapped to the annotation space through a decoder to generate a final annotation result.
9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for constructing an image annotation model for industrial scenes as described in any one of claims 1 to 4.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an implementation program for information transmission, and when the program is executed by the processor, the steps of the method for constructing an image annotation model for industrial scenarios as described in any one of claims 1 to 4 are implemented.