Intelligent authenticity identification method for oracle inscriptions based on deep learning
Through the deep learning OracleNet model and Gaussian displacement field generation algorithm, the problem of subjectivity and low efficiency of traditional oracle artifact identification methods is solved, and efficient and accurate identification of authenticity of oracle artifacts is achieved, which is suitable for practical application scenarios.
Patent Information
- Application Number
- CN202510538191.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-27
- Publication Date
- 2025-08-08
AI Technical Summary
Traditional oracle bone artifact authenticity identification methods rely on experience and technical equipment, have strong subjectivity, low identification efficiency, and difficult to identify complex counterfeiting methods. The existing AI technology lacks special solutions.
The intelligent identification method of oracle artifacts based on deep learning is adopted, and the OracleNet model is constructed, combined with the Gaussian displacement field generation algorithm to simulate the natural deformation of oracle bone, enhance data diversity, and use the ResNet18 network to extract features to achieve lossless identification.
It improves the accuracy and robustness of authenticity identification of oracle bone artifacts, reduces the dependence on physical contact of cultural relics, and provides an efficient and lightweight model deployment solution.
Smart Images
Figure CN120451659A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intersection between artificial intelligence and cultural relic identification, and in particular to an intelligent authenticity identification method for oracle bone cultural relics. Background Art
[0002] As an important representative of ancient Chinese characters, oracle bone inscriptions carry a wealth of historical and cultural information. However, with the development of the cultural relics market, the counterfeiting of oracle bone relics has become increasingly serious. Traditional methods for identifying the authenticity of oracle bone relics mainly rely on the experience and subjective judgment of experts, which has problems such as low identification efficiency, strong subjectivity, and limitations on the knowledge level and experience of experts. The fragility of oracle bone inscriptions and their nature as treasured cultural relics make it difficult to directly access the oracle bones. The scarcity of cultural relic image samples can easily lead to the problem of model overfitting. In addition, traditional methods are difficult to accurately identify some subtle features and complex counterfeiting methods. Therefore, there is an urgent need for an objective, efficient, and accurate method for identifying the authenticity of oracle bone relics. Liu Yushuang proposed a method for identifying the authenticity of oracle bone relics from the perspective of bronze rust in the article "New Trends in Oracle Bone Inscription Counterfeiting - Bronze Rust Oracle Bone Inscriptions" (Oracle Bone Inscriptions and Yin and Shang History, 2020, (00): 611-620): Specifically, the authenticity of oracle bone relics can be identified through dye or chemical raw material identification, high-imitation bronze rust identification, bronze rust transplantation identification, etc. However, the oracle bone identification method in this article has the following major flaws: 1) The identification method relies on experience and technical equipment. The operation of equipment such as Charles filters and microscopes requires professional knowledge and experience. Different people have different judgment standards, which may lead to misjudgment. In addition, some counterfeiting methods are sophisticated, and existing equipment and technology are difficult to fully detect. 2) The identification features are subjective. For example, the judgment of the color and gloss of copper rust is greatly influenced by personal aesthetics, cognition and experience. The characteristics of authentic oracle bones in different regions and different preservation environments are different, which can easily be confused with counterfeit features and difficult to accurately identify. Existing AI technologies (such as quantum cloud code traceability and three-dimensional display enhancement) are only applicable to modern commodities or display scenarios, and there is a lack of dedicated solutions for the authenticity identification of oracle bone artifacts. Summary of the Invention
[0003] In response to the problems that traditional identification methods rely on experience and technical equipment, identification features are subjective, and identification efficiency is low, the present invention proposes an intelligent authenticity identification method for oracle bone artifacts based on deep learning. It innovatively introduces elastic deformation data enhancement, improves the accuracy and robustness of authenticity identification of oracle bone artifacts, reduces dependence on physical contact with artifacts, realizes non-destructive identification, and provides an efficient and lightweight model deployment solution to adapt to actual application scenarios.
[0004] In order to achieve the above object, the technical solution of the present invention is achieved as follows:
[0005] A method for intelligently identifying the authenticity of oracle bone artifacts based on deep learning, comprising the following steps:
[0006] S1: Obtain oracle bone image sets including training sets and validation sets;
[0007] S2: Preprocessing the training set and validation set, and constructing an OracleNet model based on ResNet18, wherein the preprocessing uses a Gaussian displacement field generation algorithm to simulate the natural deformation of oracle bones;
[0008] S3: train the OracleNet model using the training set and the validation set;
[0009] S4: Pre-process the oracle bone image to be predicted and input it into the trained model to obtain the predicted probability result. The authenticity of the oracle bone artifact is judged based on the predicted probability result.
[0010] Furthermore, the method for preprocessing the training set is as follows: using a Gaussian displacement field generation algorithm to simulate the natural deformation of oracle bones, simulating the deformation characteristics of oracle bones in a natural environment by controlling deformation and continuity, and ensuring the deformation consistency of each RGB channel by using channel-independent deformation technology, and finally generating a diversified training set image tensor by adding random noise;
[0011] The method for preprocessing the validation set is as follows: bilinear interpolation is used to resize the validation set images to a uniform size; center cropping is performed on the resized images, and the size of the cropped images is consistent with the model input size to remove useless information at the edge of the image while ensuring that the center part of the image is retained; the cropped images are normalized to normalize the pixel values of the images to the range of [0,1] to accelerate the convergence of the model.
[0012] Furthermore, the Gaussian displacement field generation algorithm is implemented as follows: a random displacement field is generated for each pixel point (i, j) in each channel of the image in the training set; the new coordinates of each pixel point are calculated using the displacement field; pixel values are sampled from the original image through a bilinear interpolation method, and a deformed image is generated based on the new coordinates.
[0013] Furthermore, the method for generating the random displacement field is:
[0014]
[0015] Among them, ∈ x and ∈ y is an independent and uniformly distributed random noise matrix, G σ is a Gaussian filter with a standard deviation of σ, which is used to control the degree of smooth random noise in the displacement field. α is the displacement amplitude scaling factor, which is used to control the deformation amplitude.
[0016] Furthermore, the ResNet18-based OracleNet model includes an input image node, a feature extraction backbone node, a global average pooling layer node, a flattening operation node and a classification head improvement node connected in sequence; the input image node is used to load the input image, the feature extraction backbone node is used to extract features of different levels and types in the input image, the global average pooling layer node is used to perform a global average pooling operation on the feature map output by the feature extraction backbone node, and the classification head improvement node is used to predict the category to which the image belongs and output the probability value of the image belonging to each category.
[0017] Furthermore, the feature extraction backbone network includes an initial convolution layer, a maximum pooling layer and a multi-layer residual block structure connected in sequence; the initial convolution layer is used to extract low-level features of the input image and obtain a low-level feature map; the maximum pooling layer is used to further reduce the resolution of the feature map, enhance translation invariance, and retain significant features; the multi-residual block structure gradually extracts high-level semantic features by stacking multiple residual blocks to obtain a high-level semantic feature map.
[0018] Furthermore, the multi-layer residual block structure includes Layer 1, Layer 2, Layer 3 and Layer 4 layers connected in sequence, and each layer adopts a residual function composed of 3×3 convolutional layers.
[0019] Furthermore, the calculation process of the multi-layer residual block structure is as follows: the input feature map is calculated through the Layer1 layer, and the output dimension remains unchanged. The output feature map y1 is used as the input of the Layer2 layer. The Layer2 layer reduces the spatial resolution of the output feature map y1 by half and doubles the number of channels, and introduces 1×1 convolution for dimension matching, and outputs the feature map y2. The Layer3 layer reduces the spatial resolution of the output feature map y2 by half and doubles the number of channels, and introduces 1×1 convolution for dimension matching, and outputs the feature map y3. The Layer4 layer reduces the spatial resolution of the output feature map y3 by half and doubles the number of channels, and introduces 1×1 convolution for dimension matching, and outputs the feature map y4.
[0020] Furthermore, the classification head improvement node includes a Dropout layer, a fully connected layer and a LogSoftmax layer; the Dropout layer is used to randomly discard a percentage of neurons to prevent overfitting and enhance the generalization ability of the model; the fully connected layer is used to map high-dimensional features to binary classification results; the LogSoftmax layer converts the binary classification result vector into logarithmic probability.
[0021] Furthermore, the calculation formula of the fully connected layer is:
[0022] y q =Wx q +b(W∈R 2×512, b∈R2 )
[0023] Among them, x q is the input feature vector, W is the weight matrix, b is the bias vector, y q is the output feature vector;
[0024] The calculation formula of the LogSoftmax layer is:
[0025]
[0026] Where C is the number of categories, x i is the i-th element in the input vector, x j Represents the jth element in the input vector.
[0027] The beneficial effects of the present invention are:
[0028] By learning and training a large number of oracle bone images through deep learning models, the potential features in the oracle bone images can be excavated, avoiding the subjectivity of traditional identification methods and improving the accuracy and efficiency of authenticity identification of oracle bone artifacts.
[0029] By designing a Gaussian displacement field generation algorithm that can control deformation intensity, the diversity of data is increased, which helps the model learn richer features and improves model performance. At the same time, it solves the problem of model overfitting under small sample data and effectively improves the model's generalization ability for oracle bone texture features. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0031] Figure 1 It is the overall flow chart of the present invention.
[0032] Figure 2 This is a schematic diagram of the OracleNet model structure based on ResNet18 of the present invention. DETAILED DESCRIPTION
[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.
[0034] A deep learning-based intelligent authentication method for oracle bone artifacts, such as Figure 1 As shown, the steps include:
[0035] S1: Obtain the oracle bone image set including the training set and the validation set.
[0036] Obtain an oracle bone image set and divide it into a training set and a validation set in a ratio of 8:2. The training set is used for model training, and the validation set is used for model evaluation and tuning. The training set contains 80 real oracle bone images and 120 pseudo oracle bone images, and the validation set contains 20 real oracle bone images and 30 pseudo oracle bone images.
[0037] S2: Preprocess the training set and validation set, and build an OracleNet model based on ResNet18.
[0038] Training set preprocessing: A Gaussian displacement field generation algorithm was used to simulate the natural deformation of oracle bones. By controlling deformation and continuity, the deformation characteristics of oracle bones in their natural environment were simulated to ensure that the enhanced image retained the original texture characteristics. Furthermore, channel-independent deformation technology was used to ensure deformation consistency across RGB channels to avoid color distortion. Random noise was added to generate a diverse training set image tensor. This helps the model learn more generalizable features and enhances its robustness to natural deformations such as cracking and distortion of oracle bones.
[0039] Specifically, for each pixel (i, j) in each channel of the image in the training set, a random displacement field is generated:
[0040]
[0041] Among them, ∈ x and ∈ y is an independent and uniformly distributed random noise matrix with a value range of (-1,1), G σ is a Gaussian filter with a standard deviation of σ, which is used to control the degree of smooth random noise in the displacement field, and the value of σ is 8. α is a displacement amplitude scaling factor, which is 16 in this embodiment and is used to control the deformation amplitude.
[0042] Furthermore, the new coordinates of each pixel are calculated using the displacement field:
[0043] (i′,j′)=(i+Δx(i,j),j+Δy(i,j))
[0044] Furthermore, pixel values are sampled from the original image by a bilinear interpolation method, and a deformed image is generated based on the new coordinates.
[0045] Furthermore, random noise (Gaussian noise with a mean of 0 and a variance of 0.1 in this embodiment) is added to the deformed image to increase data diversity. The random noise can be Gaussian noise or salt and pepper noise. Different image interference conditions can be simulated by controlling the intensity and distribution of the noise.
[0046] Validation set preprocessing: Resizing, center cropping, and normalization operations are applied to the validation set. These operations convert the original image data into validation set image tensors suitable for model input, ensuring the stability and comparability of the validation results.
[0047] Specifically, an image scaling algorithm (in this embodiment, bilinear interpolation is used) is used to resize the validation set images to a uniform size (224×224).
[0048] Furthermore, center cropping is performed on the resized image to remove useless information at the edge of the image while ensuring that the center of the image is retained. The size of the cropped image is consistent with the model input size.
[0049] Furthermore, the cropped image is normalized (Min–Max normalization is used in this embodiment) to normalize the pixel values of the image to the range of [0, 1] to accelerate the convergence of the model.
[0050] The ResNet18-based OracleNet model includes an input image node, a feature extraction backbone node, a global average pooling layer node, a flattening operation node, and a classification head improvement node, which are connected in sequence.
[0051] The input image node is used to load the input image;
[0052] The feature extraction backbone node is used to extract features of different levels and types from the input image, such as natural deformation features such as cracks and distortions in oracle bone inscriptions. The feature extraction backbone node includes an initial convolutional layer, a maximum pooling layer, and a four-layer residual block structure (Layer 1 to Layer 4) connected in sequence.
[0053] The initial convolution layer is used to extract low-level features of the input image and obtain a low-level feature map; the maximum pooling layer is used to further reduce the resolution of the feature map, enhance translation invariance, and retain salient features; the residual block structure gradually extracts high-level semantic features by stacking multiple residual blocks to obtain a high-level semantic feature map.
[0054] Specifically, in this embodiment, the initial convolution layer uses a convolution kernel size of kernel_size of 7×7, a stride of 2, and a padding of 3 pixels. The convolution operation performs low-level feature extraction on the 3-channel 224×224 resolution input image and quickly reduces the image size, outputting 64-channel 112×112 resolution image data.
[0055] The calculation formula is:
[0056]
[0057] Among them, H in is the input image height, H out is the output image height, kernel_size is the convolution kernel size, stride is the step size for downsampling. padding is the padding pixel used to maintain the input boundary information. Indicates the floor symbol.
[0058] The calculation in the width direction is similar. The results in both the height and width directions are 112, and the final resolution is 112×112.
[0059] Preferably, the maximum pooling layer uses a 3×3 pooling window with a step size of 2 and a padding pixel of 1 to perform pooling on the image data output by the initial convolutional layer, further reducing the feature map resolution and outputting 64-channel 56×56 resolution image data.
[0060] The residual block structured processing layer is preferably used to address the vanishing gradient problem in deep networks and allows the model to more effectively learn high-dimensional feature representations. The input is an image vector of dimension (64, 56, 56) after processing by the max pooling layer, and the output is an image vector of dimension (512, 7, 7) after structuring through four layers (Layer 1 to Layer 4) of residual blocks.
[0061] The structured processing of residual blocks from Layer 1 to Layer 4 all follows the residual connection mechanism, which is expressed as follows:
[0062] y=F(x,{W i})+x
[0063] Among them, y is the output feature map, x is the input feature map, F(x,{W i}) represents the residual function composed of two 3×3 convolutional layers, {W i} represents the weight of layer i, and + denotes a skip connection, which allows x to be added directly to the result of the residual function F(·). This skip connection mechanism ensures that information can be passed to the next layer without any transformation, thereby simplifying the learning process and alleviating the problem of vanishing gradients. This helps alleviate the degradation problem in deep networks and makes it possible to train deeper networks.
[0064] Layer-specific processing details:
[0065] Layer 1: The input feature map has a dimension of (64, 56, 56). After processing with the residual block, the output feature map y1 maintains its dimension, also (64, 56, 56). This layer does not reduce the spatial size, mainly to extract local features without changing the spatial resolution of the input. Its mathematical expression can be expressed as:
[0066] y1=F(x1,{W i1})+x1
[0067] Among them, x1 is the input data of Layer1, y1 is the output feature map of Layer1, F(·) is the residual function composed of two 3×3 convolutional layers, and W i1 is the weight parameter of Layer1.
[0068] Layer 2: The input is the output feature map y1 of Layer 1, which has a dimension of (64, 56, 56). After downsampling with a convolution operation with a stride of 2, the output feature map y1 has a dimension of (128, 28, 28). This step halves the spatial resolution of the feature map and doubles the number of channels, enhancing the model's ability to capture complex patterns.
[0069] Its mathematical expression is as follows:
[0070] y2=F(y1,{W i2})+W s1 y1
[0071] Among them, y1 is the output feature map of Layer2, W i2 is the weight parameter of Layer2, W s1 is the weight parameter corresponding to the 1×1 convolution.
[0072] The special thing here is that when the dimensions of the input and output are inconsistent (for example, the number of channels increases from 64 to 128), an additional 1×1 convolution is used to adjust the dimension of the data y1 so that it can be consistent with F(y1,{W i2}) and add the results.
[0073] Layer 3: Continuing with a similar approach to Layer 2, Layer 3 also uses a convolution operation with a stride of 2 to further reduce the feature map size and double the number of channels. The input is the Layer 2 output feature map y2, with dimensions (128, 28, 28). After processing, the output feature map y3 has dimensions (256, 14, 14). Its mathematical expression is:
[0074] y3=F(y2,{W i3})+W s2 y2
[0075] Among them, y3 is the output feature map of Layer3, W i3 is the weight parameter of Layer3, W s2 is the weight parameter corresponding to the 1×1 convolution.
[0076] Layer 4: Continuing with a similar approach to Layer 2, the input is the Layer 3 output feature map y3, with dimensions (256, 14, 14). The final output feature map y4 has dimensions (512, 7, 7). Convolution with a stride of 2 is again used to reduce the feature map size and increase the number of channels. This completes the process of extracting features from coarse to fine, adapting to the subsequent global average pooling and classification tasks.
[0077] Its mathematical expression is:
[0078] y4=F(y3,{W i4})+W s3 y3
[0079] y4 is the output feature map of Layer4, W i4 is the weight parameter of Layer3, W s3 is the weight parameter corresponding to the 1×1 convolution.
[0080] The weight parameter W here s3 For matching feature map y3 and F(y3,{W i4}) so that the two can be added together.
[0081] The global average pooling layer node is used to perform a global average pooling operation on the feature map output by the feature extraction backbone node. Each feature map is averaged in the spatial dimensions (width and height) to obtain a feature vector of fixed length. This can reduce the number of model parameters and prevent overfitting while retaining the global information of the feature map.
[0082] Specifically, in this embodiment, the global average pooling layer performs feature aggregation on the 512-channel 7×7 resolution feature map output by the residual block structured processing layer, takes the average of all pixels in each channel, and outputs a 512-channel 1×1 resolution feature vector.
[0083] The calculation formula is:
[0084]
[0085] Among them, x c,i,j is the value of the cth channel at position (i, j) of the input feature map, y c is the value of the cth channel of the output feature vector. H×W represents the spatial resolution (height×width) of the input feature map.
[0086] The flatten operation node is used to convert the multi-dimensional feature vector output by the global average pooling layer into a one-dimensional vector so that it can be input into the subsequent fully connected layer for processing. It plays the role of data dimension conversion so that the data format meets the input requirements of the classification head.
[0087] In this embodiment, after being processed by the flatten operation node, the output is a 1-dimensional (1*512) image vector.
[0088] The calculation formula is:
[0089] y z =Flatten(x z )
[0090] Among them, x z is the input feature map, y z is the flattened vector.
[0091] The classification head improvement node is used to predict the category to which the image belongs and output the probability value of the image belonging to each category. The classification head improvement node includes a dropout layer, a fully connected layer, and a probability normalization (LogSoftmax) layer.
[0092] The Dropout layer is used to randomly drop a percentage of neurons (50% in this example) to prevent overfitting and enhance the generalization ability of the model. The fully connected layer is used to map high-dimensional features to binary classification results. The LogSoftmax layer converts the binary classification result vector into logarithmic probability.
[0093] The calculation formula of the fully connected layer is:
[0094] y q =Wx q +b(W∈R 2×512, b∈R 2 )
[0095] Among them, x q is the input feature vector, W is the weight matrix, b is the bias vector, y q is the output feature vector.
[0096] The calculation formula for probability normalization is:
[0097]
[0098] Where C is the number of categories, which is 2 in this embodiment, and x i is the i-th element in the input vector, x j Represents the jth element in the input vector.
[0099] S3: The OracleNet model is trained using the training and validation sets. After training, the model's state dictionary file is saved. Adam is used as the optimizer, and gradient clipping is employed during training to prevent gradient explosion, ensure training stability, and enable model convergence. The learning rate scheduler uses a cosine annealing learning rate adjustment strategy, dynamically adjusting the learning rate based on the number of training rounds. This allows the model to update parameters with appropriate step sizes at different stages, improving training efficiency and model performance.
[0100] S4: Pre-process the oracle bone image to be predicted and input it into the trained model to obtain the predicted probability result. The authenticity of the oracle bone artifact is judged based on the predicted probability result.
[0101] Specifically, the predicted oracle bone image is preprocessed in the same way as the validation set, and the trained model is input. After feature extraction, global average pooling, flattening operation and classification head processing, the logarithmic probability value is output. The authenticity of the oracle bone artifact is judged based on the logarithmic probability value. If the probability of belonging to a real oracle bone is greater than the preset threshold (0.5), it is judged to be a real oracle bone, otherwise it is judged to be a fake oracle bone.
[0102] When dealing with the problem of sample imbalance, the present invention uses an elastic deformation data augmentation method to increase the diversity of training data, simulate more different image deformation situations, help the model learn richer features, and improve the robustness of the model, so that it can perform better when facing images with different deformations.
[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for intelligent authenticity identification of oracle bone artifacts based on deep learning, characterized in that: Including steps: S1: Obtain oracle bone image sets including training sets and validation sets; S2: Preprocessing the training set and validation set, and constructing an OracleNet model based on ResNet18, wherein the preprocessing uses a Gaussian displacement field generation algorithm to simulate the natural deformation of oracle bones; S3: train the OracleNet model using the training set and the validation set; S4: Pre-process the oracle bone image to be predicted and input it into the trained model to obtain the predicted probability result. The authenticity of the oracle bone artifact is judged based on the predicted probability result.
2. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to claim 1 is characterized in that: The method for preprocessing the training set is as follows: using a Gaussian displacement field generation algorithm to simulate the natural deformation of oracle bones, simulating the deformation characteristics of oracle bones in a natural environment by controlling deformation and continuity, and ensuring the deformation consistency of each RGB channel by using channel-independent deformation technology, and finally generating a diversified training set image tensor by adding random noise; The validation set was preprocessed by resizing the validation set images to a uniform size using bilinear interpolation. Center-cropping was performed on the resized images to match the model input size, removing useless information from the image edges while preserving the image center. The cropped image is normalized to normalize the pixel values of the image to the range of [0, 1] to speed up the convergence of the model.
3. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to claim 2 is characterized in that: The Gaussian displacement field generation algorithm is implemented as follows: for each pixel (i, j) in each channel of the image in the training set, Generate a random displacement field; use the displacement field to calculate the new coordinates of each pixel; sample pixel values from the original image using a bilinear interpolation method, and generate a deformed image based on the new coordinates.
4. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to claim 3 is characterized in that: The method for generating a random displacement field is: Among them, ∈ x and ∈ y is an independent and uniformly distributed random noise matrix, G σ is a Gaussian filter with a standard deviation of σ, which is used to control the degree of smooth random noise in the displacement field. α is the displacement amplitude scaling factor, which is used to control the deformation amplitude.
5. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to any one of claims 1 to 4, characterized in that: The OracleNet model based on ResNet18 includes an input image node, a feature extraction backbone node, a global average pooling layer node, a flattening operation node and a classification head improvement node connected in sequence; the input image node is used to load the input image, the feature extraction backbone node is used to extract features of different levels and types in the input image, the global average pooling layer node is used to perform a global average pooling operation on the feature map output by the feature extraction backbone node, and the classification head improvement node is used to predict the category to which the image belongs and output the probability value of the image belonging to each category.
6. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to claim 5 is characterized in that: The feature extraction backbone network includes an initial convolution layer, a maximum pooling layer and a multi-layer residual block structure connected in sequence; the initial convolution layer is used to extract low-level features of the input image and obtain a low-level feature map; the maximum pooling layer is used to further reduce the resolution of the feature map, enhance translation invariance, and retain significant features; The multi-residual block structure gradually extracts high-level semantic features by stacking multiple residual blocks to obtain a high-level semantic feature map.
7. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to claim 6 is characterized in that: The multi-layer residual block structure includes Layer 1, Layer 2, Layer 3 and Layer 4 layers connected in sequence, and each layer adopts a residual function composed of 3×3 convolution layers.
8. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to claim 7 is characterized in that: The calculation process of the multi-layer residual block structure is as follows: the input feature map is calculated through the Layer1 layer, and the output dimension remains unchanged. The output feature map y1 is used as the input of the Layer2 layer. The Layer2 layer reduces the spatial resolution of the output feature map y1 by half and doubles the number of channels, and introduces 1×1 convolution for dimension matching, and outputs the feature map y2. The Layer3 layer reduces the spatial resolution of the output feature map y2 by half and doubles the number of channels, and introduces 1×1 convolution for dimension matching, and outputs the feature map y3. The Layer4 layer reduces the spatial resolution of the output feature map y3 by half and doubles the number of channels, and introduces 1×1 convolution for dimension matching, and outputs the feature map y4.
9. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to any one of claims 5 to 8, characterized in that: The classification head improvement nodes include a Dropout layer, a fully connected layer, and a LogSoftmax layer; the Dropout layer is used to randomly discard a percentage of neurons to prevent overfitting and enhance the generalization ability of the model; the fully connected layer is used to map high-dimensional features to binary classification results; and the LogSoftmax layer converts the binary classification result vector into a logarithmic probability.
10. The method for intelligent authenticity identification of oracle bone relics based on deep learning according to claim 9 is characterized in that: The calculation formula of the fully connected layer is: y q =Wx q +b(W∈R 2×512, b∈R 2 ) Among them, x q is the input feature vector, W is the weight matrix, b is the bias vector, y q is the output feature vector; The calculation formula of the LogSoftmax layer is: Where C is the number of categories, x i is the i-th element in the input vector, x j Represents the jth element in the input vector.