Ceramic authenticity detection method based on deep learning

Abstract

The invention relates to the technical field of ceramic detection, in particular to a ceramic authenticity detection method and system based on deep learning, and the method comprises the steps: S1, obtaining and preprocessing a ceramic image; s2, constructing triple training data according to the preprocessed ceramic image; s3, constructing a ceramic authenticity detection model based on a deep metric learning framework, inputting the triple training data into the ceramic authenticity detection model, and performing optimization training on the ceramic authenticity detection model by using a triple loss function; and S4, evaluating the trained ceramic authenticity detection model to obtain a trained ceramic authenticity detection model. According to the invention, the trained ceramic authenticity detection model is constructed to realize automatic, high-efficiency and high-precision detection of ceramic authenticity.

Description

Technical Field

[0001] The present invention relates to the field of ceramic detection technology, and in particular to a ceramic authenticity detection method based on deep learning. Background Art

[0002] In the long-standing field of ceramic art, distinguishing authentic ceramics from imitations has always been a key concern for professional appraisers and collectors. Traditional ceramic authenticity testing relies primarily on expert experience, conducted through visual observation of glaze color, patterns, and modeling features, as well as examination of microstructures using a magnifying glass. This approach is subject to significant subjectivity, low efficiency, and difficulty in standardization. With the advancement of technology, instruments such as fluorescence spectroscopy, X-ray diffraction, and electron microscopy have been introduced into the field of ceramic testing, enabling more objective data on elemental composition and microstructure. However, these instruments are often bulky, expensive, and complex to operate, making them difficult to meet the demands of rapid on-site identification and unable to effectively respond to evolving counterfeiting techniques and the growing demand for authentication.

[0003] Chinese patent CN202310819883.8 provides a method and device for detecting ancient ceramics based on a graph neural network model. The core of the method is to collect information on the element content of ancient ceramic fluorescence spectra, construct a graph data sample composed of porcelain body nodes, porcelain glaze nodes, and glaze color nodes fully connected by undirected edges, and use graph neural networks to distinguish authenticity and age. Although this method has certain portability and discrimination reliability, it still has obvious technical defects: first, the method still relies on fluorescence spectrum element content measurement equipment, which increases the detection cost; second, although the graph neural network model can combine multi-element features, it cannot effectively capture the micro-texture and detailed features of the ceramic surface, and its ability to distinguish high-quality imitations is limited; in addition, this method fails to give full play to the advantages of deep learning in feature extraction and pattern recognition, especially the lack of a special optimization mechanism for ceramic image features, and cannot adapt to complex and changing identification environments. Summary of the Invention

[0004] In view of this, the present invention proposes a ceramic authenticity detection method based on deep learning to solve the technical problem of insufficient detection precision and accuracy of ceramic detection methods in the existing technology.

[0005] The technical solution of the present invention is implemented as follows: The present invention provides a ceramic authenticity detection method based on deep learning, comprising the following steps:

[0006] S1, acquiring ceramic images and performing preprocessing;

[0007] S2, constructing triplet training data based on the preprocessed ceramic images;

[0008] S3. Build a ceramic authenticity detection model based on a deep metric learning framework, input the triplet training data into the ceramic authenticity detection model, and optimize the training of the ceramic authenticity detection model using the triplet loss function;

[0009] S4. Evaluate the trained ceramic authenticity detection model to obtain a trained ceramic authenticity detection model.

[0010] Based on the above technical solution, preferably, step S1 specifically includes:

[0011] S11, acquiring ceramic images, including authentic ceramic images and imitation ceramic images, and performing data enhancement on the ceramic images, wherein the data enhancement includes geometric transformation, color perturbation, and texture simulation;

[0012] S12. Normalize the ceramic image after data enhancement and standardize it using the pre-calculated mean and standard deviation to obtain a pre-processed ceramic image.

[0013] Based on the above technical solution, preferably, in step S12, the calculation formulas for the pre-calculated mean and standard deviation are as follows:

[0014]

[0015] Where c∈{R,G,B} represents the color channel of the image, N is the total number of images, H is the height of the image, and W is the width of the image. Represents the pixel value of color channel c at position (x, y) in the i-th image.

[0016] Based on the above technical solution, preferably, step S2 includes:

[0017] S21, constructing a triplet function, and using the triplet function to construct samples of the preprocessed ceramic images in batches to form a triplet set, where each triplet includes an anchor sample, a positive sample of the same type as the anchor sample, and a negative sample of a different type from the anchor sample;

[0018] S22. Input the triplet set into the data generator to generate triplet training data.

[0019] Based on the above technical solution, preferably, the triple function is as follows:

[0020] Satisfy condition a i ,p i ∈C k ,

[0021] Among them, C krepresents the set of genuine ceramics of type k, M is the total number of valid triples, a i represents the anchor sample, p i represents the positive sample, n i represents negative samples.

[0022] Based on the above technical solution, preferably, the network structure of the ceramic authenticity detection model in step S3 includes:

[0023] The input layer is used to receive the input ceramic image;

[0024] For the feature extraction backbone, we use a pre-trained ResNet50 network and freeze the first 30% of the convolutional layers.

[0025] A global average pooling layer is used to convert the feature map into a 2048-dimensional vector;

[0026] The first fully connected layer contains 128 neurons and uses the ReLU activation function;

[0027] Dropout layer, with a dropout rate of 0.5;

[0028] The second fully connected layer contains 64 neurons and uses the ReLU activation function;

[0029] Normalization layer normalizes the output features to obtain the normalized feature vector.

[0030] Based on the above technical solution, preferably, in step S3, the calculation formula of the triple loss function is:

[0031]

[0032] Where s(·) is the cosine similarity measurement function, Represents the feature vector of anchor sample a in the i-th triple; represents the feature vector of the positive sample p in the i-th triplet, represents the feature vector of negative sample n in the i-th triplet, margin is the set boundary parameter, and its value is 0.8; B is the batch size.

[0033] Based on the above technical solution, preferably, in step S4, evaluating the trained ceramic authenticity detection model includes:

[0034] Calculate one or more evaluation metrics selected from accuracy, precision, recall, F1 score, and AUC value;

[0035] and / or,

[0036] Perform feature space visualization analysis to observe the distribution of samples of different categories in the feature space, and perform similarity distribution analysis on the extracted feature vectors.

[0037] On the basis of the above technical solution, preferably, the hard example mining principle is adopted when constructing the triple set in step S21, specifically including:

[0038] First, randomly select anchor sample a, which is a genuine ceramic image;

[0039] Randomly select a positive sample p from the same authentic ceramic images as the anchor sample a;

[0040] According to the principle of hard example mining, negative sample n is selected. By calculating the feature similarity between the anchor sample a under the current model and each sample in the candidate negative sample set, the sample with the highest feature similarity to the anchor sample a is preferentially selected as the negative sample n. The similarity is calculated using the cosine similarity metric function.

[0041] The selected anchor sample a, positive sample p, and negative sample n are combined into a triplet, and the triplet is added to the triplet set.

[0042] On the basis of the above technical solution, preferably, in step S3, when optimizing and training the ceramic authenticity detection model, a difficult example mining strategy is adopted, specifically including:

[0043] During the optimization training process, the data generator is used to dynamically generate data before each training batch:

[0044] Calculate the feature representation of samples in the batch based on the current model parameters; dynamically select difficult triplet samples that meet the condition s(a,p)>s(a,n); and update the triplet data composition in the training batch in real time;

[0045] Calculate the triplet loss for the difficult triplet samples in each training batch and update the model parameters by gradient descent;

[0046] As the model parameters are continuously updated, the difficult example mining strategy continues to select difficult samples under the current model state for training until the model converges and the training ends.

[0047] The ceramic authenticity detection method and system based on deep learning of the present invention have the following advantages over the prior art:

[0048] (1) The present invention realizes the automated, efficient and high-precision detection of ceramic authenticity by constructing a trained ceramic authenticity detection model, overcoming the limitations of traditional expert identification, which is highly subjective and inefficient, and existing instrument analysis, which is expensive, complex to operate and difficult to apply on site.

[0049] (2) By adopting a pre-trained deep convolutional neural network as the basis of the deep metric learning model and freezing some of its early convolutional layers and loading pre-trained weights during initialization, the general underlying image feature extraction capabilities learned by the pre-trained model from large-scale datasets can be effectively utilized; the convergence speed of the model on a specific ceramic dataset is accelerated, and the dependence on massive labeled ceramic data is reduced, thereby improving the extraction efficiency of diverse ceramic sample features and the accuracy of the final detection.

[0050] (3) In the training process of the deep metric learning model, a triplet loss function is adopted. This loss function is based on the feature vectors of anchor samples, positive samples and negative samples. By calculating the cosine similarity between them and introducing boundary parameters for optimization, combined with the online difficult example mining strategy, it can guide the model to learn a more discriminative feature space, so that the feature vectors of similar ceramics are more clustered in space, while the feature vectors of different types of ceramics (especially genuine and high-quality imitations) are more separated in space, which significantly improves the model's ability to distinguish subtle differences, especially enhances the accuracy of distinguishing high-quality imitations. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] 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.

[0052] Figure 1 Flowchart of the ceramic authenticity detection method based on deep learning of the present invention;

[0053] Figure 2 A diagram showing the training process of the deep metric learning model of the present invention;

[0054] Figure 3 Flowchart for the construction of triplet training data. DETAILED DESCRIPTION

[0055] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] like Figure 1 As shown, the present invention provides a ceramic authenticity detection method based on deep learning, comprising the following steps:

[0057] S1, acquiring ceramic images and performing preprocessing;

[0058] S2, constructing triplet training data based on the preprocessed ceramic images;

[0059] S3. Build a ceramic authenticity detection model based on a deep metric learning framework, input the triplet training data into the ceramic authenticity detection model, and optimize the training of the ceramic authenticity detection model using the triplet loss function;

[0060] S4. Evaluate the trained ceramic authenticity detection model to obtain a trained ceramic authenticity detection model.

[0061] The ceramic authenticity detection method based on deep learning provided by the present invention successfully realizes the rapid judgment of the authenticity of ceramics by constructing a deep metric learning model based on the triplet network framework and combining the overall process of ceramic image preprocessing, feature extraction and similarity calculation; the adopted deep metric learning framework can automatically learn the discriminative feature representation of ceramic images, especially through the optimization of the triplet loss function, so that similar ceramic features are clustered and heterogeneous features are kept away, which effectively improves the model's sensitivity to subtle feature differences; the introduced online difficult example mining strategy focuses on optimizing sample pairs that are difficult to distinguish, greatly improving the ability to distinguish high-quality imitations, and providing a convenient, efficient, low-cost and highly accurate intelligent solution for ceramic authenticity identification.

[0062] Specifically, step S1 includes:

[0063] S11. Obtain images of authentic and replica ceramics as the raw dataset. Collect a large number of high-definition images of authentic and replica ceramics, covering both authentic pieces from different eras and kilns, as well as various types of replicas, to ensure a representative and diverse dataset. Authentic ceramic images primarily come from professional museums, archaeological excavations, and authoritative collections, while replica ceramic images include modern imitations and forgeries. Each image is guaranteed to clearly depict the glaze color, texture, and microscopic features of the ceramic surface.

[0064] Data augmentation was performed on the original dataset. This data augmentation process included geometric transformation, color perturbation, and texture simulation. Geometric transformation included random rotation (±15°), translation (±10%), and scaling (0.9-1.1 times); color perturbation included adjusting brightness (±20%), contrast (±15%), and saturation (±10%); and texture simulation included adding Gaussian noise (σ = 0.01) to simulate signs of erosion.

[0065] S12. Normalize the data-enhanced ceramic image and standardize it using the pre-calculated mean and standard deviation to obtain a pre-processed ceramic image. Normalize the enhanced image to eliminate the influence of different shooting equipment and lighting conditions, specifically including: mapping the pixel value from the range of 0-255 to the range of 0-1, and calculating the mean and standard deviation for the three color channels (R, G, B) of the image:

[0066]

[0067] Where c∈{R,G,B} represents the color channel of the image, N is the total number of images, H is the height of the image, and W is the width of the image. Represents the pixel value of color channel c at position (x, y) in the i-th image. The image is normalized based on the calculated mean and standard deviation, and then the highlight area is masked to eliminate specular reflection interference.

[0068] Specifically, step S2 includes:

[0069] S21, constructing a triplet function, and using the triplet function to construct samples of the preprocessed ceramic images in batches to form a triplet set, where each triplet includes an anchor sample, a positive sample of the same type as the anchor sample, and a negative sample of a different type from the anchor sample;

[0070] Specifically, anchor samples are authentic products with typical characteristics selected as reference benchmarks; positive samples are samples that belong to the same category as the anchor samples but have natural variations; negative samples include non-similar samples such as modern imitations and products from different kilns. The triple function is as follows:

[0071] Satisfy condition a i ,p i ∈C k ,

[0072] Among them, C k represents the set of genuine ceramics of type k, M is the total number of valid triples, a i represents the anchor sample, p i represents the positive sample, n i represents negative samples.

[0073] In step S21, the hard example mining principle is adopted when constructing the triplet set. In each training batch, negative samples with a moderate distance from the anchor sample in the feature space are preferentially selected. Neither negative samples that are too easy to distinguish (to avoid low training efficiency) nor negative samples that are too difficult to distinguish (to avoid unstable model training) are selected. Specifically, it includes:

[0074] First, randomly select anchor sample a, which is a genuine ceramic image;

[0075] Randomly select a positive sample p from the same authentic ceramic images as the anchor sample a;

[0076] According to the principle of hard example mining, negative sample n is selected. By calculating the feature similarity between the anchor sample a under the current model and each sample in the candidate negative sample set, the sample with the highest feature similarity to the anchor sample a is preferentially selected as the negative sample n. The similarity is calculated using the cosine similarity metric function.

[0077] The selected anchor sample a, positive sample p, and negative sample n are combined into a triplet, and the triplet is added to the triplet set.

[0078] S22. Input the triplet set into the data generator to generate triplet training data.

[0079] Use the data generator to generate triplet training data, such as Figure 3 As shown, the specific details are as follows: Input parameters include triplet set T = {(a i ,p i ,n i )}, where i ranges from 1 to N, containing N triplets; batch size b. First, initialize the image preprocessing function preprocess(·), and then enter an infinite loop to continuously generate data batches: at the beginning of each round, randomly shuffle the order of samples in the triplet set T to ensure the randomness of the data in each training cycle, and iterate in batches for the entire dataset T. The number of cycles is the total number of samples N divided by the batch size b (usually rounded down to represent the number of complete batches). Extract the current batch B from the set T, which contains triplet samples with indexes from (k-1)b to kb. Initialize three empty sets A, P, N to store batch data, and traverse the current batch B k Each triple in (a i ,p i ,n i ): The preprocessed anchor sample a i Add to the set A, and add the preprocessed positive sample p i Add to the set P, and add the preprocessed negative samples n i Add to the set N and end the sample processing in the batch. Convert the collected image data A, P, N into a standard multi-dimensional array (tensor) format.

[0080] Each ceramic image in each triplet is subjected to single image normalization processing, including adjusting the image size, normalizing the pixel value to the range of [0, 1], and performing channel-level normalization on the image using a channel normalization parameter to obtain triplet training data.

[0081] The ceramic image is resized and then pixel normalized to map all pixel values ​​from the original 0-255 range to 0-1, making the model training more stable. c ) and standard deviation (σ c ), standardize the image to eliminate interference caused by different shooting equipment, lighting conditions, etc. The standardization calculation formula is:

[0082]

[0083] Where c∈{R,G,B} represents the three color channels of the image. The normalization process processes each color channel separately, subtracting the mean of the red, green, and blue channels and dividing them by their standard deviations. For highlight areas that may appear during the normalization process, masking techniques are used to eliminate interference from specular reflections, further improving the quality and consistency of the preprocessed images. When processing multispectral image data containing infrared bands, this channel is normalized separately. Through the above preprocessing steps, the ceramic images are converted into standardized images suitable for analysis by deep metric learning models. This not only eliminates the influence of external factors on the detection results, but also provides standardized data input for subsequent feature extraction, laying the foundation for accurately determining the authenticity of ceramics.

[0084] Furthermore, step S3 specifically includes:

[0085] S31. Constructing a ceramic authenticity detection model based on a deep metric learning framework;

[0086] Furthermore, the network structure of the ceramic authenticity detection model includes: an input layer for receiving the input ceramic image; a ResNet 50 layer, which uses a pre-trained ResNet50 network, in which the first 30% of the convolutional layers are frozen; a global average pooling layer, which is used to convert the feature map into a 2048-dimensional vector; a first fully connected layer, containing 128 neurons, using the ReLU activation function; a Dropout layer with a dropout rate of 0.5; a second fully connected layer, containing 64 neurons, using the ReLU activation function; a normalization layer, which normalizes the output features to obtain a normalized feature vector.

[0087] Table 1 Network parameter configuration table

[0088]

[0089] As shown in Table 1, the network parameter configuration table details the settings for each model layer. First, the input layer has a size of 192×192×3, specified by the parameter Config.IMG_SIZE. Next, the ResNet 50 layer uses a pretrained ResNet50 network with the first 30% of layers frozen, specifically implemented as base_model.layers[:int(0.3L)]. Subsequently, a global average pooling layer is used to compress the feature map into a 2048-dimensional vector (corresponding to the code base_model.output.shape[-1]). After that, the first fully connected layer with 128 units and using the ReLU activation function is added, followed by a Dropout layer with a dropout rate of 0.5 to prevent overfitting. The second fully connected layer contains 64 units and also uses the ReLU activation function. Finally, a normalization layer is used to implement L2 normalization (calling the tf.math.12normalize function), completing the construction of the entire model structure.

[0090] S32. Use the triplet training data as the training set data to train the ceramic authenticity detection model, take the triplet loss function as the optimization target, and combine it with the online difficult example mining strategy to make the similar ceramic features clustered and the heterogeneous ceramic features separated.

[0091] Furthermore, the model is trained using the triplet training data as the training set data. The specific steps are as follows: first, the batch size is set, and then the triplet loss function is used as the optimization target. The loss function is calculated in each forward propagation, and then the model parameters are updated through backpropagation, gradually forming an effective feature space, so that the features of ceramics of the same category are clustered together, while the features of ceramics of different categories are kept away from each other. The calculation formula of the triplet loss function is:

[0092]

[0093] Where s(·) is the cosine similarity metric function, defined as Represents the feature vector of the anchor sample in the i-th triple; represents the feature vector of the positive sample in the i-th triplet, represents the feature vector of the negative sample in the i-th triplet, margin is the set boundary parameter, and its value is 0.8; B is the batch size.

[0094] During the model training process, an online hard case mining strategy is introduced, and in each training batch, the samples that meet the To improve the model's ability to identify subtle differences, the specific implementation is as follows:

[0095] During the optimization training process, the data generator is used to dynamically generate data before each training batch:

[0096] Calculate the feature representation of samples in the batch based on the current model parameters; dynamically select difficult triplet samples that meet the condition s(a,p)>s(a,n); and update the triplet data composition in the training batch in real time;

[0097] Calculate the triplet loss for the difficult triplet samples in each training batch and update the model parameters by gradient descent;

[0098] As model parameters are continuously updated, the hard example mining strategy continuously selects difficult examples within the current model state for training. This process is dynamic because, as the model's capabilities improve, previously difficult examples may become easier to distinguish, while new difficult examples are identified. The system continues this training process until the model converges, as indicated by the performance indicators on the validation set stabilizing or reaching a preset threshold. At this point, training ends.

[0099] Through the online difficult example mining strategy, the model can concentrate computing resources to process the most challenging samples, avoiding wasting computing power on simple samples that have already been learned.

[0100] To address the common learning rate selection difficulties and training instability problems in deep learning model training, this paper introduces an adaptive learning strategy that dynamically adjusts the learning rate according to the training process to balance the model convergence speed and stability. The strategy includes: setting an initial learning rate η0; adopting a step-by-step decay strategy, and dynamically adjusting the learning rate η(t) according to the number of training steps t in the following manner:

[0101]

[0102] Among them, γ is the attenuation coefficient, which is 0.96; k is the attenuation step size, which is 1000; is the initial learning rate, which is 5×10 -5 .

[0103] When the learning rate falls below the preset minimum value, the training process is terminated. The attenuation coefficient is dynamically adjusted based on changes in validation set performance. If performance improvement slows, the attenuation coefficient is increased; if performance fluctuates significantly, the attenuation coefficient is decreased. When the model is detected to be stuck in a local optimum, a learning rate restart is implemented. The restart condition is: the change in the F1 score of three consecutive validation set evaluations is less than the threshold. Restart operation: Temporarily increase the learning rate to 5 times the current value. A maximum of two learning rate restarts are allowed per training cycle. Through this adaptive learning strategy, the model training process can be flexibly adjusted according to the actual training status, avoiding common overfitting and underfitting problems while taking into account training efficiency and model performance.

[0104] Furthermore, during the model training process, after each training step or a fixed number of steps (e.g., 500 steps), a comprehensive evaluation is performed on the validation set. The evaluation indicators include: calculating the model performance indicators; performing feature similarity analysis on the feature vectors of the extracted positive and negative samples to achieve feature space visualization and analyze the distribution of samples of different categories in the feature space. The details are as follows:

[0105] Model performance indicators include accuracy (the proportion of correct predictions), precision (the proportion of predicted positive samples that are actually positive), recall (the proportion of actual positive samples that are correctly predicted), F1 score (the harmonic mean of precision and recall), and AUC value (the area under the ROC curve, used to measure classification ability). Among them, the calculation formula for accuracy is (TP+TN) / (TP+TN+FP+FN); the calculation formula for precision is (TP) / (TP+FP); the calculation formula for recall is (TP) / (TP+FN); the calculation formula for F1 score is 2×(precision+recall) / (precision+recall); the calculation formula for AUC value is

[0106] In addition to numerical indicators, feature space visualization analysis is also performed to intuitively display the learning results of the model. Feature space visualization analysis includes:

[0107] All samples in the validation set are used to extract 64-dimensional feature vectors through the trained model. The high-dimensional feature vectors are mapped to a two-dimensional space using the t-SNE dimensionality reduction technique, and a two-dimensional scatter plot is output. The color of the point represents the category (such as genuine / imitation, or more detailed classification of age and kiln). Visualization result analysis: Observe the distribution of samples of different categories, including the tight clustering of similar genuine products (such as Song Dynasty official kiln samples); check whether samples from different dynasties present a reasonable distribution pattern, such as a radial distribution according to the evolution of age; analyze the position of the imitation samples in the feature space to see whether they are distributed at the edge of the genuine cluster; identify abnormal points in the feature space, which may represent incorrectly labeled samples or special cases. Perform feature similarity analysis on the extracted positive and negative sample feature vectors, and use the kernel density estimation method to draw the similarity distribution curve. The calculation formula of the kernel density estimation method is:

[0108]

[0109] Where K(·) is the Gaussian function, h is the bandwidth parameter, n is the number of samples, and x i is the similarity value of the i-th sample. Kernel density estimation can intuitively show the difference in the similarity distribution between genuine and fake pairs and evaluate the model's discriminative ability.

[0110] Based on the comprehensive evaluation results, the model is dynamically adjusted. The specific strategies are as follows:

[0111] (1) Adjust the network structure: If the performance of the validation set is poor and the performance of the training set is good (overfitting): increase the Dropout rate, add L2 regularization, and reduce the number of neurons in the fully connected layer; If the performance of both the training set and the validation set is poor (underfitting): increase the width of the fully connected layer, unfreeze more convolutional layers, and increase the model capacity; If the recognition rate of some categories is particularly low: add an attention mechanism to the feature extraction network to enhance the perception of key areas.

[0112] (2) Adjustment of loss function parameters: According to the cosine similarity distribution curve, optimize the boundary parameter margin of the triplet loss function. If the internal similarity of the authentic samples is generally low, reduce the margin value; if the similarity distribution of the authentic and fake samples overlaps seriously, increase the margin value.

[0113] (3) Training strategy optimization: Adjust the learning rate strategy according to the learning curve. If the fluctuation is large, reduce the initial learning rate; optimize the online difficult example mining parameters, such as adjusting the difficult example screening ratio and difficulty coefficient; if it is found that some samples are always misclassified, conduct in-depth analysis and adjust the annotation.

[0114] (5) Sample reconstruction and enhancement: Targeted reconstruction of triplet samples of categories with poor verification performance, adding challenging boundary samples, such as high-simulation products and authentic products with clear age characteristics, and adding special training samples for kiln products that are easily confused in history (such as official kilns and Ru kilns).

[0115] In this paper, the trained ceramic authenticity detection model can convert input ceramic images into discriminative 64-dimensional feature vectors. The position of these feature vectors in the feature space effectively reflects the authenticity of the ceramic. Combined with the constructed authentic feature benchmark library, this model can achieve rapid and accurate identification of unknown ceramic samples, providing a reliable basis for subsequent similarity calculations and authenticity determinations.

[0116] 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 ceramic authenticity detection method based on deep learning, characterized in that: The following steps are involved: S1, acquiring ceramic images and performing preprocessing; S2, constructing triplet training data based on the preprocessed ceramic images; S3. Build a ceramic authenticity detection model based on a deep metric learning framework, input the triplet training data into the ceramic authenticity detection model, and optimize the training of the ceramic authenticity detection model using the triplet loss function; S4. Evaluate the trained ceramic authenticity detection model to obtain a trained ceramic authenticity detection model.

2. A ceramic authenticity detection method based on deep learning according to claim 1, characterized in that: Step S1 specifically includes: S11, acquiring ceramic images, including authentic ceramic images and imitation ceramic images, and performing data enhancement on the ceramic images, wherein the data enhancement includes geometric transformation, color perturbation, and texture simulation; S12. Normalize the ceramic image after data enhancement and standardize it using the pre-calculated mean and standard deviation to obtain a pre-processed ceramic image.

3. A ceramic authenticity detection method based on deep learning as claimed in claim 2, characterized in that: In step S12, the calculation formulas for the pre-calculated mean and standard deviation are as follows: Where c∈{R,G,B} represents the color channel of the image, N is the total number of images, H is the height of the image, and W is the width of the image. Represents the pixel value of color channel c at position (x, y) in the i1th image.

4. The ceramic authenticity detection method based on deep learning according to claim 2, characterized in that: Step S2 includes: S21, constructing a triplet function, and using the triplet function to construct samples of the preprocessed ceramic images in batches to form a triplet set, where each triplet includes an anchor sample, a positive sample of the same type as the anchor sample, and a negative sample of a different type from the anchor sample; S22. Input the triplet set into the data generator to generate triplet training data.

5. The ceramic authenticity detection method based on deep learning according to claim 4, characterized in that: The triple function is as follows: Satisfy condition a i ,p i ∈C k , Among them, C k represents the set of genuine ceramics of type k, M is the total number of valid triples, a i represents the anchor sample, p i represents the positive sample, n i represents negative samples.

6. The ceramic authenticity detection method based on deep learning according to claim 4, characterized in that: The network structure of the ceramic authenticity detection model in step S3 includes: The input layer is used to receive the input ceramic image; For the feature extraction backbone, we use a pre-trained ResNet50 network and freeze the first 30% of the convolutional layers. A global average pooling layer is used to convert the feature map into a 2048-dimensional vector; The first fully connected layer contains 128 neurons and uses the ReLU activation function; Dropout layer, with a dropout rate of 0.5; The second fully connected layer contains 64 neurons and uses the ReLU activation function; Normalization layer normalizes the output features to obtain the normalized feature vector.

7. The ceramic authenticity detection method based on deep learning according to claim 4, characterized in that: In step S3, the calculation formula of the triplet loss function is: Where s(·) is the cosine similarity measurement function, Represents the feature vector of anchor sample a in the i-th triple; represents the feature vector of the positive sample p in the i-th triplet, represents the feature vector of negative sample n in the i-th triplet, margin is the set boundary parameter, and its value is 0.8; B is the batch size.

8. The ceramic authenticity detection method based on deep learning according to claim 1, characterized in that: In step S4, evaluating the trained ceramic authenticity detection model includes: Calculate one or more evaluation metrics selected from accuracy, precision, recall, F1 score, and AUC value; Perform feature space visualization analysis to observe the distribution of samples of different categories in the feature space, and perform similarity distribution analysis on the extracted feature vectors.

9. The ceramic authenticity detection method based on deep learning according to claim 5, characterized in that: The hard example mining principle is adopted when constructing the triple set in step S21, specifically including: First, randomly select anchor sample a, which is a genuine ceramic image; Randomly select a positive sample p from the same authentic ceramic images as the anchor sample a; According to the principle of hard example mining, negative sample n is selected. By calculating the feature similarity between the anchor sample a under the current model and each sample in the candidate negative sample set, the sample with the highest feature similarity to the anchor sample a is preferentially selected as the negative sample n. The similarity is calculated using the cosine similarity metric function. The selected anchor sample a, positive sample p, and negative sample n are combined into a triplet, and the triplet is added to the triplet set.

10. The ceramic authenticity detection method based on deep learning according to claim 7, characterized in that: In step S3, when optimizing and training the ceramic authenticity detection model, a difficult example mining strategy is adopted, which specifically includes: During the optimization training process, the data generator is used to dynamically generate data before each training batch: Calculate the feature representation of samples in the batch based on the current model parameters; dynamically select difficult triplet samples that meet the condition s(a,p)>s(a,n); and update the triplet data composition in the training batch in real time; Calculate the triplet loss for the difficult triplet samples in each training batch and update the model parameters by gradient descent; As the model parameters are continuously updated, the difficult example mining strategy continues to select difficult samples under the current model state for training until the model converges and the training ends.

Images