Deep learning-based ancient porcelain aging trace identification system

Abstract

The present application belongs to the technical field of cultural relic identification, and particularly relates to an ancient porcelain aging trace identification system based on deep learning. The system collects high-resolution microscopic images of the glaze surface and exposed body of ancient porcelain through an image acquisition module, identifies the glaze concentric circle scratch features using a glaze scratch detection network based on double-flow feature extraction, identifies the exposed body limonite coloration features using a multi-spectral fusion exposed body limonite analysis network, realizes the automatic detection and quantitative determination of the two features of ancient porcelain aging, and outputs clear binary classification identification results of old and new artifacts based on the double-feature joint determination logic. The present application can accurately capture natural aging traces that are difficult to imitate artificially, eliminate the subjectivity and ambiguity of traditional identification, improve the identification efficiency and consistency, and thus provide reliable technical support for the rapid screening and accurate dating of massive ancient porcelain cultural relics.

Description

Technical Field

[0001] This invention belongs to the technical field of cultural relic identification, and in particular relates to a deep learning-based system for identifying aging traces on ancient ceramics. Background Technology

[0002] Ancient ceramics appraisal is a core link in the protection, inheritance and standardized operation of cultural relics market. At present, the appraisal of ancient ceramics relies on a comprehensive analysis of the aging traces of the body and glaze, the glaze coating and luster, the shape characteristics and the marks and decorations. However, the judgment of old and new ceramics lacks a unified scale and standard and remains at the level of vague concepts.

[0003] Based on the principle of natural sedimentation over time, ancient ceramics over a century old possess two core characteristics that cannot be artificially replicated. First, when ancient ceramics emerge from the kiln, the glaze has a Mohs hardness of 7. Over time, hydrogen ions in the air replace alkali metal ions on the glaze, forming a silica gel film on the surface, gradually decreasing the hardness until it falls below 7 after about a century. At this point, quartz particles in the air, with a constant Mohs hardness of 7, impact and rub against the glaze, creating regular concentric circular natural scratches. These scratches become more pronounced with age, exhibiting a concentric circle pattern under light, a key to identifying century-old artifacts that forgers cannot replicate. Second, the exposed clay body of ancient ceramics contains iron, which reacts chemically with water and oxygen in the natural environment to form limonite. The color of limonite gradually deepens over time, developing a distinct hue after about a century. Simultaneously, the clay body's hardness is already below 7 when it emerges from the kiln. The impact and polishing of the uneven exposed clay surface by air by quartz particles over a century results in polished surfaces, both flat and recessed, exhibiting a uniform, bright reflection under sidelight and a smooth feel. The deepening color of limonite and the polishing marks on the body constitute the unique imprints that time bestows upon ancient ceramics, and together they serve as reliable evidence for determining whether a piece is old or new.

[0004] This method focuses on the natural imprints left by centuries of time on ancient ceramics, eliminating interference from easily imitated factors such as shape, decoration, and marks, and establishing a clear criterion that a piece with two distinct characteristics is an antique, while the absence of these characteristics indicates a new piece. However, this identification method still relies on manual observation and subjective judgment by appraisers using magnifying glasses or microscopes, resulting in low efficiency, difficulty in identifying subtle features, and a lack of quantitative standards. Therefore, it is insufficient to meet the needs for rapid screening and accurate identification of a vast number of ancient ceramic artifacts. With the rapid development of deep learning and computer vision technologies, acquiring high-resolution microscopic images of glaze scratches and limonite mineralization at exposed clay surfaces, and using deep learning models to automatically detect and quantify dual features, transforming scientific identification methods based on natural spatiotemporal imprints into efficient, objective, and reproducible intelligent identification systems, has become a pressing technical issue in the field of ancient ceramic scientific authentication. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a deep learning-based system for identifying aging traces in ancient ceramics. This invention aims to solve the long-standing problems in the field of ancient ceramic authentication, such as reliance on subjective experience, lack of unified scientific standards, low authentication efficiency, and difficulty in identifying subtle aging features. It proposes a deep learning-based method for identifying aging traces in ancient ceramics. By acquiring high-resolution microscopic images of the glaze and exposed clay of ancient ceramics, a deep learning model for dual-flow feature extraction and feature fusion is introduced. Concentric circle scratch detection networks on the glaze and limonite color analysis networks on the exposed clay are constructed to automatically identify and quantify the dual characteristics of centuries of aging in ancient ceramics. Based on the joint judgment logic of the dual features, a clear binary classification result of new or old pieces is output, completely resolving the ambiguity in traditional authentication. This invention transforms the scientific authentication method based on the principle of natural spatiotemporal sedimentation into a reproducible and quantifiable intelligent detection method, providing reliable technical support for the rapid screening, accurate dating, and preservation of massive amounts of ancient ceramic artifacts.

[0006] This invention provides a deep learning-based method for identifying aging marks on ancient ceramics, applicable to the identification of the age of ancient ceramic artifacts. The identification objects include complete glazed ancient ceramic objects, fragments, and porcelain shards. The system includes: an image acquisition module, used to acquire microscopic images of concentric circular scratches on the glaze surface of ancient ceramics and microscopic images of limonite mineralization in the exposed clay area of ​​ancient ceramics. The glaze surface microscopic images use dark-field imaging to enhance the contrast of the scratches, making the fine arc-shaped scratches formed by the friction of quartz particles clearly distinguishable. The exposed clay microscopic images use bright-field imaging to realistically restore the color and surface reflective characteristics of limonite. At the same time, visible light band images and 850nm near-infrared band images are acquired.

[0007] The image preprocessing module is used to perform noise reduction, illumination normalization, and multi-focal image depth fusion on the microscopic image. It uses a semantic segmentation network to segment and extract the scratch area in the glaze microscopic image to generate standardized glaze scratch image blocks. It also uses a semantic segmentation network to segment and extract the exposed body area in the exposed body microscopic image to generate standardized exposed body image blocks.

[0008] The deep learning inference module has a built-in trained glaze scratch detection network and exposed limonite analysis network. The glaze scratch detection network receives glaze scratch image blocks and outputs the probability of concentric circle scratches on the glaze surface. The exposed limonite analysis network receives exposed limonite image blocks and outputs the probability of exposed limonite mineralization.

[0009] The dual-feature joint determination module is used to receive the probability of the existence of concentric circle scratches on the glaze and the probability of the existence of exposed limonite mineralization. Based on the preset dual-feature joint determination threshold logic, when both features are determined to exist, the old device identification result is output; otherwise, the new device identification result is output.

[0010] The authentication report generation module is used to generate electronic authentication reports that include the artifact number, microscopic images, dual-feature detection probability values, judgment conclusions, and confidence scores.

[0011] Furthermore, the glaze scratch detection network adopts a dual-stream feature extraction structure. The first feature extraction stream uses ConvNeXt-Tiny as the backbone network to extract depth feature vectors, and the second feature extraction stream extracts gradient direction histogram feature vectors of scratch texture. The depth feature vectors and texture feature vectors are fused and then output as probabilities through a classification layer.

[0012] Furthermore, the exposed limonite analysis network adopts a multispectral fusion input structure, which uses a channel-level stitching of visible light band images and 850nm near-infrared band images as network input. After feature extraction by the EfficientNetV2-S backbone network, a channel attention module and a spatial attention module are used to enhance the modeling ability of color features and reflectivity uniformity features, and output probabilities.

[0013] Furthermore, the determination logic of the dual-feature joint determination module is as follows: the scratch determination threshold is set to 0.65, the limonite determination threshold is set to 0.65, and when the probability of concentric circular scratches on the glaze surface is greater than or equal to the scratch determination threshold and the probability of exposed limonite mineralization is greater than or equal to the limonite determination threshold, the object is determined to be an old object, and a quantitative confidence score S is generated. When the above conditions are not met, the object is determined to be a new object.

[0014] Furthermore, the training process of the glaze scratch detection network includes the following steps: Step A1: Data Collection and Labeling: 120 ancient ceramic artifacts over 100 years old, confirmed by experts, from the collections of the Palace Museum, Henan Museum, Kaifeng Museum, and Pingyuan Museum were selected. 3,000 microscopic images of the glaze surface were obtained as positive samples, and 2,000 microscopic images of the glaze surface of modern antique-style ceramics were collected as negative samples. Two or more ceramic appraisal experts performed pixel-level semantic labeling on the concentric circle scratch areas in the positive samples. The labeling content included the scratch arc segments and the central cluster area. Step A2: Construct a dual-stream feature extraction structure: Construct a first feature extraction stream with ConvNeXt-Tiny as the backbone network, input the glaze scratch image patch, and output a 1024-dimensional depth feature vector. Construct a second feature extraction stream with the scratch texture direction gradient histogram feature as input, and use a three-layer fully connected network to output a 256-dimensional texture feature vector. Concatenate the depth feature vector and the texture feature vector at the channel level to form a 1280-dimensional fused feature vector, which is then input into the classification layer. Step A3: Introduce the concentric circle geometric constraint loss function: Define the geometric constraint loss function as the weighted sum of the variance of the curvature of the scratch arc segment and the dispersion of the center point set; Step A4: Model Training: The ConvNeXt-Tiny backbone network is initialized with ImageNet pre-trained weights. The AdamW optimizer is used, with the initial learning rate set to 1e-4 and the batch size set to 32. Data augmentation strategies such as random rotation, brightness jitter, and Gaussian noise injection are adopted. The network is trained for 100 epochs, and the optimal model weights are selected on the validation set to obtain the glaze scratch detection network.

[0015] Furthermore, the training process of the exposed limonite analysis network includes the following steps: Step B1: Data collection and labeling: Select the exposed clay parts of the above-mentioned ancient ceramic artifacts that are over 100 years old in the collection, and acquire microscopic images of the exposed clay in the visible light band and the 850nm near-infrared band respectively. Obtain 2500 sets of positive sample images and collect 1800 sets of microscopic images of exposed clay from modern imitations as negative samples. Experts will classify and label the degree of limonite mineralization. Step B2: Construct a multispectral fusion input structure: The visible light image and the 850nm near-infrared band image are stitched together at the channel level to form a 6-channel input tensor. EfficientNetV2-S is used as the backbone network. The first 3 channels of the 6-channel input layer are initialized with pre-trained RGB weights, and the last 3 channels are initialized with Xavier. A channel attention module and a spatial attention module are introduced after the backbone network. Step B3: Loss function design: Use the focal loss function and introduce contrastive loss in the intermediate layer to enhance the compactness of intra-class features; Step B4: Model training. The SGD optimizer was used with a momentum of 0.9 and an initial learning rate of 5e-3. Cosine annealing was used for learning rate scheduling, and the batch size was set to 64. The training was conducted for 120 epochs to obtain the analysis network for exposed limonite.

[0016] The beneficial effects achieved by this invention are as follows: 1. This invention transforms the dual-feature identification method of concentric circle scratches on glaze and exposed limonite mineralization based on the principle of natural spatiotemporal sedimentation into an automated detection system driven by a deep learning model. This realizes the scientific, standardized and intelligent identification of ancient ceramics. Through dual-stream network reasoning on microscopic images, the system can accurately capture the characteristics of natural aging over a century that are difficult to imitate artificially. It outputs a clear judgment that the presence of these two features indicates an old piece and the absence of these two features indicates a new piece. This completely eliminates the root of vague concepts and subjective disputes in traditional identification and fills the intelligent gap in the field of scientific authentication of ancient ceramics. 2. This invention introduces a concentric circle geometric constraint loss function and a multispectral fusion input strategy into the deep learning model. The concentric circle geometric constraint loss function is specifically optimized for the concentric circle regularity of natural scratches on the glaze surface, enhancing the network's ability to distinguish between natural concentric circle scratches and artificial random scratches. The multispectral fusion input strategy utilizes the near-infrared band to enhance the capture of limonite spectral absorption characteristics. The application of dual-stream feature extraction structure and attention mechanism enables the network to simultaneously focus on the geometric morphology of scratches and the color and reflective texture of exposed clay, significantly improving the detection sensitivity and anti-interference ability of fine aging traces. 3. This invention constructs a complete intelligent identification system process from microscopic image acquisition, preprocessing, deep learning inference to dual-feature joint judgment and report generation, realizing the automation and batch operation of ancient ceramic identification. The system can be applied to cultural heritage institutions, auction houses and customs inspection scenarios, greatly improving the identification efficiency and consistency of ancient ceramic cultural relics, and providing reliable technical support for cultural relic identity authentication, collection management, market regulation and prevention of cultural relic loss. Attached Figure Description

[0017] The accompanying drawings are provided to further understand the present invention and form part of the specification. They are used together with the embodiments of the present invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 This is a schematic diagram of the overall process of the deep learning-based ancient ceramic aging trace identification system proposed in this invention; Figure 2 This is a flowchart of the dual-feature joint determination logic in this invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those familiar to those skilled in the art. Furthermore, any methods and materials similar to or equivalent to those described herein may be applied to this invention. The preferred embodiments and materials described herein are for illustrative purposes only and do not limit the scope of this application.

[0021] Example 1: This example provides a deep learning-based system for identifying aging traces in ancient ceramics. It is applied to the identification of old and new ancient ceramic artifacts. The identification objects include complete glazed ancient ceramic objects, fragments, and porcelain shards. The system in this example includes an image acquisition module, an image preprocessing module, a deep learning inference module, a dual-feature joint judgment module, and an identification report generation module.

[0022] The image acquisition module acquires microscopic images of concentric circular scratches on the glaze surface of ancient ceramics, performs microscopic imaging on a designated area of ​​the glaze surface, sets the optical magnification to 10x, and uses dark-field imaging to enhance the contrast of the scratches, making the fine arc-shaped scratches formed by the friction of quartz particles clearly visible. At the same time, it acquires microscopic images of limonite mineralization in the exposed clay area of ​​ancient ceramics, performs microscopic imaging on a designated area of ​​the exposed clay surface, sets the optical magnification to 20x, and uses bright-field imaging to realistically restore the color and surface reflective characteristics of limonite, acquiring visible light band images and 850nm near-infrared band images.

[0023] The image preprocessing module performs Gaussian filtering for noise reduction, illumination normalization based on the gray-world hypothesis, and multi-focal image depth fusion on the microscopic image. The U-Net semantic segmentation network is used to segment and extract the scratch region in the glaze microscopic image, generating a standardized 256×256 pixel glaze scratch image block. The U-Net semantic segmentation network is also used to segment and extract the exposed clay region in the exposed clay microscopic image, generating a standardized 256×256 pixel exposed clay image block.

[0024] The deep learning inference module incorporates a trained glaze scratch detection network and an exposed limonite analysis network. The glaze scratch detection network receives glaze scratch image blocks and outputs the probability of concentric circle scratches on the glaze surface. The exposed limonite analysis network receives exposed limonite image blocks and outputs the probability of exposed limonite mineralization.

[0025] The dual-feature joint judgment module sets the scratch judgment threshold to 0.65 and the goethite judgment threshold to 0.65. When the probability of concentric circular scratches on the glaze surface is ≥0.65 and the probability of goethite mineralization on the exposed body is ≥0.65, the object is judged to be an old object, and a quantitative confidence score S is generated. When the above conditions are not met, the object is judged to be a new object.

[0026] The identification report generation module generates an electronic identification report that includes the artifact number, a microscopic image thumbnail, dual-feature detection probability values, judgment conclusions, and confidence scores.

[0027] Example 2, based on Example 1, describes the training process of the glaze scratch detection network, which includes the following steps: This embodiment also provides a method for preparing a low-carbon emission coal-based composite liquid fuel, which specifically includes the following steps: Step A1: Data Collection and Labeling: 120 ancient ceramic artifacts over 100 years old, confirmed by experts, from the collections of the Palace Museum, Henan Museum, Kaifeng Museum, and Pingyuan Museum were selected. 3,000 microscopic images of the glaze surface were obtained as positive samples, and 2,000 microscopic images of the glaze surface of modern antique-style ceramics were collected as negative samples. Two or more ceramic appraisal experts performed pixel-level semantic labeling on the concentric circle scratch areas in the positive samples. The labeling content included the scratch arc segments and the central cluster area. Step A2: Construct a dual-stream feature extraction structure: Construct a first feature extraction stream with ConvNeXt-Tiny as the backbone network, input the glaze scratch image patch, and output a 1024-dimensional depth feature vector. Construct a second feature extraction stream with the scratch texture direction gradient histogram feature as input, and use a three-layer fully connected network to output a 256-dimensional texture feature vector. Concatenate the depth feature vector and the texture feature vector at the channel level to form a 1280-dimensional fused feature vector, which is then input into the classification layer. Step A3: Introduce the concentric circle geometric constraint loss function: Define the geometric constraint loss function L_geo as the weighted sum of the variance of the curvature of the scratch arc segment and the dispersion of the center point set. The calculation formula is as follows: ,in This is the set of curvatures of the detected scratch arc segments. For the curvature variance, The total loss function is calculated as follows: (The formula is missing from the original text.) ,in Cross-entropy is used for classification loss; Step A4: Model Training: The ConvNeXt-Tiny backbone network is initialized with ImageNet pre-trained weights. The AdamW optimizer is used, with the initial learning rate set to 1e-4 and the batch size set to 32. Data augmentation strategies such as random rotation, brightness jitter, and Gaussian noise injection are adopted. The network is trained for 100 epochs, and the optimal model weights are selected on the validation set to obtain the glaze scratch detection network.

[0028] Example 3, based on Example 1, describes the training process of the exposed limonite analysis network, which includes the following steps: Step B1: Data collection and labeling: Select the exposed clay parts of the above-mentioned ancient ceramic artifacts that are over 100 years old in the collection, and acquire microscopic images of the exposed clay in the visible light band and the 850nm near-infrared band respectively. Obtain 2500 sets of positive sample images and collect 1800 sets of microscopic images of exposed clay from modern imitations as negative samples. Experts will classify and label the degree of limonite mineralization. Step B2: Construct a multispectral fusion input structure: The visible light image and the 850nm near-infrared band image are stitched together at the channel level to form a 6-channel input tensor. EfficientNetV2-S is used as the backbone network. The first 3 channels of the 6-channel input layer are initialized with pre-trained RGB weights, and the last 3 channels are initialized with Xavier. A channel attention module and a spatial attention module are introduced after the backbone network. Step B3: Loss Function Design: Use the focal loss function, calculated as follows: ,in To predict probabilities for the model, a contrastive loss is introduced in the intermediate layer to enhance the compactness of intra-class features; Step B4: Model training. The SGD optimizer was used with a momentum of 0.9 and an initial learning rate of 5e-3. Cosine annealing was used for learning rate scheduling, and the batch size was set to 64. The training was conducted for 120 epochs to obtain the analysis network for exposed limonite.

[0029] Example 4: This example is based on Examples 1-3. In this example, the overall system performance is verified.

[0030] An independent test set was collected, containing microscopic image data of 50 antique ceramic pieces from the museum's collection and 50 modern imitations. The antique pieces spanned from the Song to the Qing dynasties. Batch identification tests were conducted using the system of this invention. The test results showed that the system had a recall rate of 96% for antique pieces, a precision rate of 94% for new pieces, and an overall judgment accuracy rate of 95%. The average testing time for a single piece was approximately 3 minutes. Compared with the blind test results of three senior ceramic appraisal experts on the same test set, the experts' average accuracy rate was 82%, and the consistency of judgment among the three experts was 78%. The system of this invention demonstrated significant advantages in terms of accuracy, consistency, and efficiency.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based system for identifying aging traces on ancient ceramics, characterized by: The system includes: an image acquisition module, an image preprocessing module, a deep learning inference module, a dual-feature joint judgment module, and an identification report generation module; The image acquisition module is used to acquire microscopic images of concentric circular scratches on the glaze surface of ancient ceramics and microscopic images of limonite mineralization in the exposed body area of ​​ancient ceramics. The microscopic images of the glaze surface are enhanced by dark field imaging to enhance the contrast of the scratches, while the microscopic images of the exposed body are restored by bright field imaging to restore the color of limonite and the surface reflective characteristics. At the same time, visible light band images and 850nm near-infrared band images are acquired. The image preprocessing module is used to perform noise reduction, illumination normalization, and multi-focal image depth fusion on the microscopic image. It uses a semantic segmentation network to segment and extract the scratch area in the glaze microscopic image to generate standardized glaze scratch image blocks. It also uses a semantic segmentation network to segment and extract the exposed body area in the exposed body microscopic image to generate standardized exposed body image blocks. The deep learning inference module has a built-in trained glaze scratch detection network and exposed limonite analysis network. The glaze scratch detection network receives glaze scratch image blocks and outputs the probability of concentric circle scratches on the glaze surface. The exposed limonite analysis network receives exposed limonite image blocks and outputs the probability of exposed limonite mineralization. The dual-feature joint determination module is used to receive the probability of the existence of concentric circular scratches on the glaze and the probability of the existence of exposed limonite mineralization. Based on the preset dual-feature joint determination threshold logic, when both features are determined to exist, the old device identification result is output; when neither feature is determined to exist, the new device identification result is output. The authentication report generation module is used to generate electronic authentication reports that include the artifact number, microscopic images, dual-feature detection probability values, judgment conclusions, and confidence scores.

2. The deep learning-based system for identifying aging traces in ancient ceramics according to claim 1, characterized in that: The glaze scratch detection network adopts a dual-stream feature extraction structure. The first feature extraction stream uses ConvNeXt-Tiny as the backbone network to extract depth feature vectors, and the second feature extraction stream extracts gradient direction histogram feature vectors of scratch texture. The depth feature vector and texture feature vector are fused and then output as probabilities through a classification layer. The exposed limonite analysis network adopts a multispectral fusion input structure, which uses a channel-level stitching of visible light band images and 850nm near-infrared band images as network input. After feature extraction by the EfficientNetV2-S backbone network, the channel attention module and spatial attention module are used to enhance the modeling ability of color features and reflectivity uniformity features, and output probability.

3. The deep learning-based system for identifying aging traces in ancient ceramics according to claim 2, characterized in that: The determination logic of the dual-feature joint determination module is as follows: the scratch determination threshold is set to 0.65, the goethite determination threshold is set to 0.65, when the probability of concentric circular scratches on the glaze is ≥0.65 and the probability of goethite mineralization on the exposed body is ≥0.65, the object is determined to be an old object, and a quantitative confidence score S is generated. When the above conditions are not met, the object is determined to be a new object.

4. The deep learning-based system for identifying aging traces on ancient ceramics according to claim 3, characterized in that: The training process of the glaze scratch detection network includes the following steps: Step A1: Data collection and annotation. Collect microscopic images of the glaze of modern antique-style ceramics as negative samples. Two or more ceramic appraisal experts perform pixel-level semantic annotation on the concentric circle scratch areas in the positive samples. The annotation content includes the scratch arc segments and the circle center cluster areas. Step A2: Construct a dual-stream feature extraction structure. Construct the first feature extraction stream with ConvNeXt-Tiny as the backbone network, input the glaze scratch image patch, and construct the second feature extraction stream with the scratch texture direction gradient histogram feature as input. Concatenate the depth feature vector and texture feature vector at the channel level to form a 1280-dimensional fused feature vector, which is then input into the classification layer. Step A3: Introduce the concentric circle geometric constraint loss function; Step A4: Model training. The ConvNeXt-Tiny backbone network is initialized with ImageNet pre-trained weights. The AdamW optimizer is used to select the optimal model weights on the validation set to obtain the glaze scratch detection network.

5. The deep learning-based system for identifying aging traces on ancient ceramics according to claim 4, characterized in that: The training process of the exposed limonite analysis network includes the following steps: Step B1: Data collection and labeling. Select the exposed clay parts of ancient ceramic artifacts from the collection that are over 100 years old, obtain microscopic images of the exposed clay to obtain positive sample images, and collect microscopic images of the exposed clay of modern imitations as negative samples. Experts will then classify and label the degree of limonite mineralization. Step B2: Construct a multispectral fusion input structure by performing channel-level stitching of the visible light image and the 850nm near-infrared band image to form a 6-channel input tensor. Use EfficientNetV2-S as the backbone network. Initialize the first 3 channels of the 6-channel input layer with pre-trained RGB weights and initialize the last 3 channels with Xavier. Introduce a channel attention module and a spatial attention module after the backbone network. Step B3: Loss function design. The focus loss function is adopted, and contrastive loss is introduced in the intermediate layer to enhance the compactness of intra-class features; Step B4: Model training. The SGD optimizer was used with a momentum of 0.9 and an initial learning rate of 5e-3. Cosine annealing was used to schedule the learning rate to obtain the analysis network for exposed limonite.

Images