Bronze ware texture identification method based on multi-principle data annotation and deep learning
By employing multi-principle data annotation and a deep learning object detection network, the problems of low accuracy and poor generalization ability in bronze artifact texture recognition were solved. This enabled the construction of a high-quality training dataset and accurate model recognition, making it suitable for image recognition of various cultural heritages.
Patent Information
- Application Number
- CN202511040778.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-11-18
AI Technical Summary
Existing technologies lack a unified "truth value" standard for bronze artifact texture recognition. The unique texture features and poor image quality result in low-quality training data for the model, low recognition accuracy, and poor generalization ability.
A high-quality training dataset is constructed using a multi-principle data annotation strategy. This dataset is combined with a deep learning object detection network, including annotations based on principles of intuitiveness, labelability, scalability, wholeness, and multi-dimensionality. Residual networks and spatial pyramid pooling modules are used for feature extraction and prediction.
It significantly improves the accuracy and robustness of bronze texture recognition, solves multi-scale and detail problems, enhances the model's generalization ability, and is suitable for recognition tasks in different scenarios.
Smart Images
Figure CN120976575A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of computer vision and artificial intelligence application, and relates to an image recognition method applied to digital protection of cultural heritage, in particular, a bronze ware texture recognition method based on multi-principle data labeling and deep learning. BACKGROUND
[0002] As an important cultural heritage, the surface texture of bronze ware (such as beast face pattern, cloud and thunder pattern, and dragon pattern) carries rich historical and artistic information. Accurate identification and classification of these textures are of great significance for archaeological research, digital filing and active utilization of cultural relics.
[0003] However, automatic recognition of bronze ware texture faces many challenges. First, the naming and classification standards of texture differ in different literature, and there is a lack of unified "ground truth" standard. Second, the texture features themselves have uniqueness: the size span is extremely large, from a few centimeters to several meters; the lines are extremely delicate, up to millimeter level; and the image quality is poor due to factors such as rust, damage, shooting angle, and light, resulting in blurred features.
[0004] Existing general object detection algorithms, such as YOLO and ResNet, although perform well in regular image recognition tasks, will not work well if directly applied to bronze ware texture recognition due to low quality of training data. For example, inconsistent labeling will cause the model to learn incorrect features; for textures that are small in size or large in area, simple rectangular box labeling will introduce a lot of background noise or cause information loss, making it difficult for the model to converge, resulting in low recognition accuracy and poor generalization ability. Therefore, how to construct a high-quality training dataset for such special recognition objects and combine it with a deep learning model for effective recognition is a difficult problem to be solved in the current technical field. SUMMARY
[0005] In view of the deficiencies of the prior art, the present application provides a bronze ware texture recognition method based on multi-principle data labeling and deep learning, which significantly improves the accuracy and robustness of bronze ware texture recognition.
[0006] To achieve the above purpose, the present application adopts the following technical solutions:
[0007] A bronze ware texture recognition method based on multi-principle data labeling and deep learning includes the following steps:
[0008] S1, obtaining an image containing bronze ware texture, and constructing an original image dataset;
[0009] S2, labeling the texture in the aforementioned original image dataset according to a preset multi-principle data labeling strategy, and generating a high-quality labeled dataset;
[0010] S3, input the high-quality labeled data set into a preset deep learning target detection network model for training and evaluation to obtain a mature model capable of identifying textures of bronze wares;
[0011] S4, processing a new bronze ware image by using the mature model to output category and position information of textures contained in the image.
[0012] As a preferred embodiment of the above scheme, in step S1, the image data can be obtained through on-site shooting in a museum and collecting on an Internet platform.
[0013] As a preferred embodiment of the above scheme, in step S2, the multi-principle data labeling strategy includes:
[0014] Intuitive principle: only the textures that can be directly recognized by the naked eye are labeled, and the textures that cannot be directly judged due to blurring, obstruction, etc. are not labeled, so as to ensure the direct correspondence between the labeled data and the pixel features;
[0015] Identification principle: only the textures with individualized and unambiguous features are labeled, and the features with too strong commonality or easy to produce ambiguity are excluded, so as to avoid the model learning ambiguous or wrong features;
[0016] Expansion principle: when labeling incomplete textures, if the remaining part contains key identification features or the completeness is higher than a preset threshold, the category of the complete form is labeled; or when labeling low-recognition textures, the labeling range is appropriately expanded, and the local structure of the object is also framed, so as to introduce context auxiliary information;
[0017] Holistic principle: for textures with repetitive and continuous distribution, a single labeling box is formed by overall labeling instead of labeling one by one according to the smallest unit, so as to reduce redundant labeling and retain the overall layout information of the textures;
[0018] Multi-dimensional principle: ensure that the labeled data used for training is diversified, including samples with different shooting angles, lighting conditions, background environments and image resolutions, so as to enhance the generalization ability of the model and prevent overfitting.
[0019] As a preferred embodiment of the above scheme, in step S3, the labeled data set is randomly divided into a training set and a test set according to a certain proportion, the training is completed on the training set, and then the evaluation is performed on the test set to prove the effectiveness of the method.
[0020] As a preferred embodiment of the above scheme, in step S3, the structure of the deep learning target detection network model includes a backbone network for extracting high-level features of an image, a neck network for fusing multi-scale features, and a decoupled detection head for predicting categories and positions.
[0021] As a preferred solution of the above-mentioned scheme, the backbone network is a residual network, the neck network is a spatial pyramid pooling module, and the decoupled detection head inputs the feature map to the classification branch and the positioning branch in parallel for independent prediction.
[0022] A bronze ware texture recognition system based on multi-principle data labeling and deep learning, comprising:
[0023] A data collection and arrangement module is responsible for obtaining image data containing bronze ware texture;
[0024] A data labeling module labels the original image data set according to a multi-principle data labeling strategy using a labeling tool;
[0025] A model training module builds a deep learning target detection network and trains a mature model capable of identifying bronze ware texture;
[0026] A model evaluation and application module is used to test and evaluate the effectiveness of the method and process new bronze ware images to output the category and location information of the texture contained in the image.
[0027] Due to the above structure, the beneficial effects of the present application are:
[0028] 1. Significantly improve the recognition accuracy: By implementing the five labeling principles of intuitiveness, identification, expansiveness, integrity and multidimensionality, the quality of the training set is guaranteed from the data source, effectively reducing the interference of noise and ambiguous information on model training, enabling the model to learn more accurate and recognizable texture features, thereby greatly improving the recognition accuracy.
[0029] 2. Effectively solve the problem of multi-scale and details: The expansiveness principle and the integrity principle are designed specifically for the characteristics of large size span, rich details or strong repetition of texture. Through contextual association and holistic labeling, the model can better capture and understand the features of complex texture, solving the problem that traditional labeling methods cannot balance macro and micro.
[0030] 3. Enhance the generalization ability of the model: The multidimensionality principle ensures the diversity of training data, enabling the trained model to adapt to different recognition tasks in different scenarios, avoiding the overfitting problem caused by single training scenario, and showing stronger robustness in actual application.
[0031] 4. The method has universality: The multi-principle data labeling strategy proposed in the present application is not only suitable for the bronze ware of Sanxingdui, but also can be applied to other cultural heritage image recognition tasks with similar characteristics (such as fine, multi-scale, and existence of defects), and has good application prospect. BRIEF DESCRIPTION OF DRAWINGS
[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.
[0033] Figure 1 This is a flowchart of the process of the present invention;
[0034] Figure 2 This is a schematic diagram of data annotation in this invention;
[0035] Figure 3 This is a schematic diagram illustrating the identification principle of the present invention;
[0036] Figure 4 , Figure 5 This is a schematic diagram of the model evaluation results of the present invention. Detailed Implementation
[0037] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0038] like Figure 1 As shown, this embodiment provides a method for bronze artifact texture recognition based on multi-principle data annotation and deep learning, including the following steps:
[0039] S1. Data Collection and Processing
[0040] Images containing bronze artifact textures were acquired to construct the original image dataset. In this embodiment, image data was acquired through various methods, including on-site photography at museums and collection from internet platforms. Approximately 4,900 images containing Sanxingdui bronze artifacts and their textures were obtained (of which, self-taken images accounted for about 15%, with dimensions of approximately 2,000 pixels; images collected from the internet ranged in size from 200 to 2,000 pixels, covering various angles and lighting conditions), forming the original image dataset.
[0041] S2. Data labeling based on multiple principles
[0042] Using the open-source annotation tool LabelImg, and based on a pre-defined multi-principle data annotation strategy, the textures in the aforementioned original image dataset are annotated (e.g., ...). Figure 2 (As shown), a high-quality labeled dataset is generated. The multi-principle data labeling strategy includes:
[0043] Intuitive principle: Only the textures that can be directly recognized by the naked eye in the image are labeled, and the textures that cannot be directly judged due to blurring, obstruction, etc. are not labeled, so as to ensure the direct correspondence between the labeled data and the pixel features. Specifically, in a partially blurred bronze statue image, the beast face pattern on its abdomen is clearly visible and is labeled; the texture of the ring foot part cannot be recognized due to rust and low light, and is not labeled.
[0044] Identifying principle: Only textures with personalized and unambiguous features are labeled, and features with too strong commonality or prone to ambiguity (such as “bronze eye bubble”) are excluded, so as to avoid the model learning ambiguous or incorrect features. Specifically, for the bronze portrait, the “braids” and “cloud thunder pattern vertical ears” have unique and unambiguous shapes, and are labeled. As a basic component of all eye structures, “bronze eye bubble” has too strong commonality and is not labeled as an independent texture (such as Figure 3
[0045] Expansion principle: When labeling incomplete textures, if the remaining part contains key identification features or the completeness is higher than the preset threshold (such as 60%), the class of its complete form is labeled; or when labeling low-recognition textures, the labeling range is appropriately expanded, and the local structure of the object where the texture is located is also selected, so as to introduce context auxiliary information. Specifically, for a broken bronze snake, more than 60% of the “diamond cloud pattern” features are still retained on its cross section, and the class is labeled as “lingxingyunwen”. For the “rolled cloud pattern” on the top of the bronze bird head, the entire feather crest part containing the texture is selected and labeled, and the label is set as “juanyunwen”, so that the model learns auxiliary features from the overall shape of the feather crest. The specific labeling list is as follows in Table 1.
[0046] Table 1 Labeling list
[0047]
[0048]
[0049]
[0050]
[0051] Integrity principle: For textures with repetitive and continuous distribution (such as fish scale-shaped feather patterns), the whole is labeled, forming a single labeling box, rather than being labeled one by one according to the smallest unit, so as to reduce redundant labeling and preserve the overall layout information of the texture. Specifically, for the “fish scale-shaped feather pattern” on the bronze chicken body, instead of labeling each “scale” one by one, the entire feather pattern area is labeled as a whole, and the label is set as “yulinzhuangyuwen”.
[0052] Multi-dimension principle: Ensure the diversity of the labeled data used for training, including samples with different shooting angles, lighting conditions, background environments, and image resolutions, to enhance the generalization ability of the model and prevent overfitting. Specifically, ensure that images of the same texture (such as "beast face pattern") in the dataset have both official high-definition museum images and tourist-photographed oblique and partial close-up images, as well as morphological variants from different vessels (zun, lei).
[0053] Through the above labeling, a high-quality labeled dataset (PASCAL VOC format) containing 81 types of vessels and texture labels is generated.
[0054] S3, Model training and evaluation
[0055] The high-quality labeled dataset is input into a pre-set deep learning object detection network model for training and evaluation, obtaining a mature model capable of recognizing bronze vessel textures. Specifically, it includes:
[0056] S31, Build a deep learning object detection network based on the PyTorch framework. The network uses a residual network ResNet-18 as the backbone network to extract high-level features of images, and a pyramid pooling module SPP as the neck network to process and fuse feature maps of different scales. Finally, a decoupled detection head is connected, which parallelly inputs the feature map into the classification branch and the positioning branch to predict the target class and the bounding box position respectively.
[0057] S32, Randomly divide the labeled dataset into training set and test set according to a certain proportion (such as 80%:20% proportion). Set the training parameters: the optimizer is SGD, the initial learning rate is 0.001, the momentum is 0.9, and the training rounds are 150. The loss function is composed of three parts: confidence loss, classification loss, and bounding box regression loss. After training, evaluate on the test set, and the average precision mean (MAP) of the model reaches 87.5%, proving the effectiveness of this method. For obvious features such as "bronze figure with gold mask" and "bronze beast face", the accuracy rate is more than 99% (as shown in Figure 4 、 Figure 5 ).
[0058] For "cloud pattern", "diamond cloud pattern" and other difficult-to-identify textures, acceptable recognition results are also obtained, and the model evaluation results are as shown in Table 2.
[0059] Table 2 Evaluation results
[0060]
[0061]
[0062]
[0063]
[0064] S4, model application
[0065] The mature model is used for processing a new bronze image, and the category and position information of the texture contained in the image are output.
[0066] The embodiment also provides a bronze texture recognition system based on multi-principle data labeling and deep learning, comprising:
[0067] A data collection and arrangement module is responsible for acquiring image data containing bronze textures;
[0068] A data labeling module labels original image data sets according to a multi-principle data labeling strategy by using a labeling tool;
[0069] A model training module builds a deep learning target detection network and trains a mature model capable of recognizing bronze textures;
[0070] A model evaluation and application module is used for testing and evaluation to prove the effectiveness of the method, and processing a new bronze image, and outputting the category and position information of the texture contained in the image.
[0071] The above is only a preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A method for bronze artifact texture recognition based on multi-principle data annotation and deep learning, characterized in that: Includes the following steps: S1. Obtain images containing bronze artifact textures and construct the original image dataset; S2. Based on the preset multi-principle data annotation strategy, the textures in the aforementioned original image dataset are annotated to generate a high-quality annotated dataset. S3. Input the high-quality labeled dataset into a preset deep learning object detection network model for training and evaluation to obtain a mature model that can recognize the texture of bronze artifacts. S4. Use the mature model to process the new bronze artifact image and output the category and location information of the texture contained in the image.
2. The method for bronze artifact texture recognition based on multi-principle data annotation and deep learning according to claim 1, characterized in that: In step S1, image data can be obtained through various means, including taking photos on-site at the museum and collecting data from internet platforms.
3. The method for bronze artifact texture recognition based on multi-principle data annotation and deep learning according to claim 1, characterized in that: In step S2, the multi-principle data annotation strategy includes: The principle of intuitiveness: only textures that can be directly identified by the naked eye in the image are labeled. Textures that cannot be intuitively judged due to blurring, occlusion, or other reasons are not labeled, so as to ensure a direct correspondence between the labeled data and pixel features. The principle of labeling: only textures with personalized and unambiguous features are labeled, and features that are too common or easily ambiguous are removed, so as to avoid the model learning vague or incorrect features. Extensibility principle: When annotating incomplete textures, if the remaining part contains key identification features or the completeness is higher than a preset threshold, it should be annotated according to the category of its complete form; or when annotating low-recognition textures, the annotation range should be appropriately expanded to include the local structure of the object in which it is located, so as to introduce contextual auxiliary information. Holistic principle: For textures with repetitive and continuous distribution, the whole is labeled to form a single label box, rather than labeling each smallest unit one by one, in order to reduce redundant labeling and retain the overall layout information of the texture. Multi-dimensional principle: Ensure that the labeled data used for training comes from diverse sources, including samples with different shooting angles, lighting conditions, background environments, and image resolutions, in order to enhance the generalization ability of the model and prevent overfitting.
4. The method for bronze artifact texture recognition based on multi-principle data annotation and deep learning according to claim 1, characterized in that: In step S3, the labeled dataset is randomly divided into a training set and a test set according to a certain ratio. After training on the training set, the dataset is evaluated on the test set to prove the effectiveness of the method.
5. The method for bronze artifact texture recognition based on multi-principle data annotation and deep learning according to claim 1, characterized in that: In step S3, the deep learning object detection network model adopts a structure including a backbone network for extracting high-level image features, a neck network for fusing multi-scale features, and a decoupled detection head for predicting category and location.
6. The method for bronze artifact texture recognition based on multi-principle data annotation and deep learning according to claim 5, characterized in that: The backbone network is a residual network, the neck network is a spatial pyramid pooling module, and the decoupled detection head inputs the feature map in parallel to the classification branch and the localization branch for independent prediction.
7. A bronze artifact texture recognition system based on multi-principle data annotation and deep learning according to any one of claims 1 to 6, characterized in that: include: Data collection and processing module: responsible for acquiring image data containing bronze artifact textures; Data annotation module: Uses annotation tools to annotate the original image dataset according to a multi-principle data annotation strategy; Model training module: Build a deep learning object detection network and train it to obtain a mature model that can recognize the texture of bronze artifacts; Model Evaluation and Application Module: Used to test and evaluate the effectiveness of the method, process new bronze artifact images, and output the category and location information of the textures contained in the image.