Intelligent identification method and system for archaeological site carbonized plant seeds
APSNet solves the problem of identifying carbonized plant seeds by extracting high-frequency information through a size-aware module and Fourier transform, and combines this with texture information to guide network learning, achieving efficient and accurate cross-class discrimination.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-04-10
AI Technical Summary
Existing deep learning methods are ineffective in identifying carbonized plant seeds with highly similar morphologies. Traditional research methods are time-consuming, inefficient, and subject to subjectivity, making it impossible to achieve a systematic and quantitative identification of carbonized grain seeds.
The APSNet deep learning architecture is adopted. High-frequency information is extracted through size perception module and Fourier transform, and texture information is combined to guide network learning. Features of different granularities are learned in layers, and cross-class discrimination ability is improved by using asynchronous dual-domain classification structure.
It significantly improves the accuracy and generalization ability of carbonized plant seed identification, can automatically sense differences in seed size and texture features, and enhances the ability to identify morphologically closely related categories.
Smart Images
Figure CN121837849A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of archaeobotany and automated detection, and in particular relates to a method and system for intelligent identification of carbonized plant seeds from archaeological sites. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Carbonized grain seeds, as the most direct food remains of agricultural civilizations, carry crucial information about agricultural production and dietary structure, and have therefore long been a core object of archaeobotanical research. Current research methods heavily rely on expert experience, typically involving excavation, flotation, classification, and analysis. Becoming an archaeobotanical expert capable of conducting independent research often requires years of professional training and practice. Especially in the classification and analysis of carbonized seeds, researchers must distinguish subtle differences in texture, size, and embryonic regions to determine species, a process that is not only time-consuming and inefficient but also prone to subjectivity. This limitation makes it difficult for traditional research to achieve systematization and quantification. Current deep learning methods utilize the powerful learning capabilities of deep networks. Although deep learning has made progress in archaeological subfields such as Bronze Age determination, fragment reconstruction, mural restoration, and ancient text recognition, existing classification methods do not prioritize differences in size and texture details when faced with a vast number of morphologically similar plant seeds, rendering them unsuitable for the practical scenarios corresponding to the identification of carbonized grain seeds. Summary of the Invention
[0004] To address at least one of the technical problems mentioned above, this invention provides an intelligent identification method and system for carbonized plant seeds from archaeological sites. By strengthening the expression of size and texture information, using affine modeling of size difference to model contour information as a size difference representation, and utilizing texture information to guide the network to focus on the detailed features between carbonized seeds with similar appearances, the method improves the ability to distinguish between different categories.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for intelligent identification of carbonized plant seeds at archaeological sites, comprising the following steps: Acquire carbonized plant seed image data and extract high-frequency information from the carbonized plant seed image data; The original carbonized plant seed images and extracted high-frequency information are affinely fused into the backbone network channel by channel, guiding the backbone network to learn multi-granular features. Based on multi-granularity features and a trained multi-head classifier, features of different granularities are learned hierarchically. Discriminative features of the global domain and local domain are learned simultaneously at different levels. The discriminative features learned at each level are then fused to obtain the classification result.
[0006] Furthermore, the extraction of high-frequency information from the carbonized plant seed image data includes: Fourier transform is applied to convert the carbonized plant seed image to the frequency domain to obtain the spectral representation of the original image; Based on the spectral representation of the original image, high-frequency components are separated using a frequency domain mask to obtain the high-frequency components; High-frequency information in the spatial domain is obtained by reversing the high-frequency components to return to the spatial domain.
[0007] Furthermore, the original carbonized plant seed image and the extracted high-frequency information are affinely fused channel by channel into the backbone network, guiding the backbone network to learn multi-granularity features, including: Based on the extracted high-frequency information and the preprocessed image, an initial feature channel affine transformation is performed. The resulting affine feature map is then input into the backbone network to obtain the backbone feature map. The obtained affine feature map and backbone feature map are then subjected to multiple feature channel affine transformations and corresponding backbone network feature extractions to obtain the affine feature map and backbone feature map of each level. This includes passing the affine feature map and backbone feature map of the previous level through the next level channel affine transformation and corresponding backbone network feature extraction to obtain the affine feature map and backbone feature map of the next level.
[0008] Further, the step of obtaining the next-level affine feature map and backbone feature map by passing the previous-level affine feature map and backbone feature map through the next-level channel affine transformation and the corresponding backbone network feature extraction includes: Connect the affine feature map and the backbone feature map of the previous level along the channel dimension, and attach the fused information to the last layer of the backbone feature map of this level. By processing the fused features through channel interactions, the network searches for pairwise relationships between the original features and size information of the backbone through interactive operations. Specifically, each channel is compressed into a scalar to represent its global importance, and local convolutions are applied along the channel dimension to model the interaction between adjacent channels. The resulting scalar interaction matrix is projected back to the cascaded feature map, thereby establishing size pairing between the fused information and each channel. Convolution is used to simulate MLP operations, and channel-based affine transformations are performed. Based on the calculated pairing information, the size and texture difference information are combined to obtain the affine feature map and backbone feature map of the corresponding level.
[0009] Furthermore, the hierarchical learning of features of different granularities based on multi-granularity features and the trained multi-head classifier includes: The backbone feature map extracted at the fourth level is input into the first classification head, and the backbone feature maps extracted at the second and third levels are input into the corresponding second and third classification heads. Then, the scale feature maps output by each classifier are concatenated and input into the fourth classification head. Finally, the results of the four classification heads are fused to generate the final prediction.
[0010] Furthermore, the first classification head adopts a dual-domain classification head and designs a dual-domain voting mechanism to obtain classification results simultaneously from both channel and spatial directions.
[0011] Furthermore, the training process of the multi-head classifier includes: The first stage of training ends when the features pass through the second layer of the backbone network, obtaining coarse-grained features that are sensitive to small-scale targets. The second stage of training ends when the features pass through the three-layer bottleneck structure, and the coarse-grained features are processed to improve the recognition effect of mesoscale targets. The third stage of training uses a combined classification head, Head_Com, which processes both channel domain information and spatial domain information to generate a composite supervision signal that fuses channel domain classification loss and spatial domain contrast loss. The fourth stage of training concatenates the multi-scale features obtained in the first three stages along the channel dimension and generates the final classification prediction by fusing the classification head.
[0012] A second aspect of the present invention provides an intelligent identification and carbonization system for carbonized plant seeds from archaeological sites, comprising: The high-frequency information extraction module is used to extract high-frequency information from the acquired carbonized plant seed image data. The multi-granularity feature learning module is used to affinely fuse the original carbonized plant seed image and the extracted high-frequency information into the backbone network channel by channel, guiding the backbone network to learn multi-granularity features. The contrastive learning module is used to learn features of different granularities in layers based on multi-granularity features and a trained multi-head classifier. It simultaneously learns global and local discriminative features at different levels and fuses the discriminative features learned at each layer to obtain the classification result.
[0013] A third aspect of the present invention provides a computer-readable storage medium.
[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent identification method for carbonized plant seeds at an archaeological site as described above.
[0015] A fourth aspect of the present invention provides a computer device.
[0016] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the intelligent identification method for carbonized plant seeds at archaeological sites as described above.
[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the challenge of classifying carbonized plant seeds under a microscope, where inter-class differences are small but intra-class differences are large. It proposes for the first time a "size-texture" collaborative discrimination framework. APSNet embeds a size-aware module into its fine-grained backbone network, using Fourier high-frequency filtering to capture seed scale and micro-texture information in real time. Through channel-level affine transformation, scale cues are seamlessly injected into general features, enabling the network to automatically perceive size without additional annotation. By enhancing the representation of size and texture information, size difference affine transformation models contour information as a size difference representation. Simultaneously, texture information guides the network to focus on detailed features among similarly carbonized seeds, thereby improving cross-class discrimination capabilities.
[0018] This invention proposes an asynchronous dual-domain classification structure to further explore the "channel-space" two-dimensional discrimination clues. In the first and second stages, the asynchronous dual-domain classification structure maintains the traditional classification structure. In the third stage, channel domain voting and spatial domain comparison are used to calculate the supervision signal. In the fourth stage, multi-head fusion and multi-scale prediction preserve both global semantics and enhance local details. This progressive learning amplifies subtle differences and significantly improves APSNet's ability to recognize morphologically close categories.
[0019] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0021] Figure 1 This is a flowchart of a method for intelligent identification of carbonized plant seeds at archaeological sites, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the SP module structure provided in an embodiment of the present invention; Figure 3 This is a diagram of the ADD classification header structure provided in an embodiment of the present invention. Detailed Implementation
[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0025] To address the challenges in ancient plant seed classification, this invention proposes a deep learning architecture capable of automatically perceiving size differences, termed the Ancient Plant Seed Classification Network (APSNet). APSNet primarily consists of three parts: a SizePerception (SP) module, a backbone network, and an Asynchronous Dual Domain Classification Head (ADD). The core of the SP module lies in its designed SizeFeature Extraction (SFE) module, which uses Fourier transform to obtain high-frequency information from images, thereby capturing subtle differences that are difficult to distinguish using traditional texture and morphological features. Subsequently, Feature Channel Affine (FCA) further models contour information as a representation of size differences, while simultaneously utilizing texture information to guide the network to focus on detailed features between seeds with similar carbonized appearances, thus improving cross-class discrimination capabilities.
[0026] During the layer-by-layer feature propagation and abstraction process, APSNet's backbone network, in conjunction with four staged classification heads, achieves staged supervised training of multi-granular features. This design helps the network learn both global and local discriminative features at different levels simultaneously. In this process, the Dual Domain Classification Header (Head_Com) further models the fused features output by the SP module through contrastive learning, effectively strengthening the expression of size difference information. This gives APSNet better robustness and generalization ability in complex ancient plant seed classification tasks.
[0027] Example 1 like Figure 1As shown in the figure, this embodiment provides a method for intelligent identification of carbonized plant seeds at archaeological sites, including the following steps: Step 1: Acquire image data of ancient plant carbonized seeds and preprocess them to obtain preprocessed image data; Initial image data was obtained by capturing 3×H×W images of ancient plant seeds under a 1.6x magnification microscope using a professional archaeological experimental microscope. The initial image data was then preprocessed: first, the images were resized to squares by center cropping; then, the cropped images were randomly horizontally flipped and randomly stitched together; finally, the image data was normalized using the mean and variance. This data preprocessing yielded image data beneficial for network processing and learning, improving the model's generalization ability. During network training, B images were stitched together to form B×3×H×W data and fed into the network for learning. During network inference, the input was 1×3×H×W.
[0028] Step 2: Extract high-frequency information from ancient plant carbonized seed image data; In this embodiment, high-frequency information of plant seed images is extracted through low-complexity filtering operations, thereby providing size and texture details, which facilitates subsequent feature extraction guided by mixed information.
[0029] Specifically, this includes: after acquiring the preprocessed three-channel image B×3×H×W, such as... Figure 2 As shown, the SFE module applies Fourier transform to convert the image to the frequency domain, decomposing it into amplitude and phase components at the pixel level, as follows: , in, This represents a B×3 two-dimensional input matrix. Indicates pixel position, It is a complex exponential basis function. For the corresponding The corresponding spectrum representation, where H represents high, For width, The values are taken starting from 0 and continuing up to H-1. Two-dimensional matrix coordinates Coordinates transformed to the frequency domain.
[0030] Then use a frequency domain mask. Separate the high-frequency components, which are used to represent fine-grained details such as size and texture, while retaining the high frequencies: , in, For high-frequency components, if Then the frequency is preserved, while Then the frequency is filtered out.
[0031] It is obtained by reversing the high-frequency components to return to the spatial domain, and is represented as: , in, This represents the high-frequency image after the inverse Fourier transform.
[0032] In the SP module, this mixture of size and texture detail spatially guides the backbone network, such as... Figure 2 As shown, the high-frequency information of B×3×H×W is fused with a convolutional layer with Kernel_size=1 and Stride=1 to obtain a high-frequency information spatial fusion matrix of B×1×H×W, which is used as one of the inputs of the FCA module.
[0033] SFE can be represented as:
[0034] in, Represents the spatial fusion matrix of high-frequency information. The stride of the convolution kernel. The kernel size is the convolution kernel size. for, It is a frequency domain mask.
[0035] Step 3: Based on the extracted high-frequency information and the preprocessed image, perform initial feature channel affine transformation. Input the obtained affine feature map into the backbone network to obtain the backbone feature map. Then, perform multiple feature channel affine transformations and corresponding backbone network feature extractions on the obtained affine feature map and backbone feature map to obtain the affine feature map and backbone feature map of each level. Finally, perform the affine feature map and backbone feature map of the previous level through the next level channel affine transformation and corresponding backbone network feature extraction to obtain the affine feature map and backbone feature map of the next level. APSNet guides the backbone network to automatically learn inter-class differences by aligning the mixed information generated by SFE through FCA. FCA achieves this by performing channel affine transformation.
[0036] This embodiment includes three levels of feature channel affine mapping and corresponding backbone network feature extraction; the specific feature channel affine mapping and corresponding backbone network feature extraction process at each level includes the following steps: Step 301: Based on the extracted high-frequency information and the preprocessed image, perform initial feature channel affine transformation and backbone network feature extraction to obtain the initial feature channel affine feature map and backbone feature map; Specifically: FCA first receives the feature map B×1×H×W output from SFE and the preprocessed image; Then, the two feature maps are connected along the channel dimension to obtain the initial feature channel affine map, and the fused information, i.e. the initial feature channel affine map, is attached to the last layer of the backbone feature map to obtain the first-level backbone feature map. Step 302: The affine feature map and backbone feature map obtained from the first level are processed through the feature channel affine transformation of the next level and the corresponding backbone network feature extraction to obtain the affine feature map and backbone feature map of that level; for example... Figure 2 As shown, it specifically includes: FCA first receives the outputs from the previous layer SFE and Backbone, which are feature maps of shapes B×1×H×W and B×N×H×W; Then, the two feature maps are connected along the channel dimension, and the fused information is appended to the last layer of the backbone feature map of this layer to obtain a feature map of B×(N+1)×H×W. Then, the fused features are processed through channel interaction, enabling the network to search for pairwise relationships between the original features of the backbone and key information such as size through interactive operations; Specifically, each channel is compressed into a scalar to represent its global importance, and then local convolutions are applied along the channel dimension to model the interactions between adjacent channels. The resulting scalar interaction matrix is projected back onto the concatenated feature maps, thus establishing a size pairing between the fused information and each channel; Finally, Conv with Kernel_size=1 and Stride=1 is used to simulate MLP operations, perform channel-based affine transformation, and obtain the corresponding affine feature map and backbone feature map based on the calculated pairing information, combined with the size and texture difference information.
[0037] Because APSNet employs a staged training strategy to capture multi-granularity information, FCA requires affine transformations of size and texture features at different stages. Therefore, the fused information is aligned using CBR (convolution, batch normalization, ReLU activation) blocks before moving to the next stage.
[0038] The characteristic channel radiation FCA has the same structure and can be represented as: , in, This indicates the output of the SFE. This represents the intermediate output of the backbone network. Indicates ECA channel attention. This indicates splicing along the channel dimension; Because the seed objects analyzed in archaeobotany come from uncontrollable environments, the original fine-grained feature distribution is disturbed. Traditional classification methods only consider fine-grained features from a macroscopic perspective, failing to accurately capture the discriminative features exhibited by carbonization and long-term storage, thus limiting classification performance. Unlike existing technologies, this invention addresses the classification challenge of small inter-class differences and large intra-class differences in carbonized plant seeds under a microscope, proposing for the first time a "size-texture" collaborative discrimination framework. APSNet embeds a size-aware (SP) module into its fine-grained backbone network, using Fourier high-frequency filtering to capture seed scale and microtexture information in real time, and seamlessly injecting scale cues into general features through channel-level affine transformation, enabling the network to automatically perceive size without additional annotation.
[0039] Step 4: Based on the backbone feature maps extracted from each level, multiple classification heads learn features of different granularities, and the features of different granularities are fused to obtain the final classification prediction result; After the SP module guides the backbone network to perceive size and texture details, a segmented classification structure is introduced to perceive multi-granularity features. In this embodiment, four classification heads are used to learn features of different granularities hierarchically, enabling the extraction of subtle discriminative cues from visually similar categories. The four classification heads include a dual-domain classification head (Head_Com) and two ordinary classification heads. A concatenated category header (Head_Con); This represents a traditional classification header, with a specific structure as follows: Figure 3 As shown, the dual-domain classification header Head_Com incorporates a dual-domain voting mechanism to obtain classification results simultaneously from both channel and spatial directions. The use of the spatial domain compensates for the limitations of traditional methods in that they lack explicit supervision of size information.
[0040] The backbone feature map extracted at the fourth level is input into the dual-domain classification head Head_Com, and the backbone feature maps extracted at the second and third levels are input into the corresponding ordinary classification heads. Then, the scale feature maps output by each classifier are concatenated using the concatenated classification head Head_Con, and finally the results of the four classification heads are fused to generate the final prediction.
[0041] The training process for the four classifier heads consists of four stages; Specifically, the first training phase of the ADD classification structure ends when passing through the second layer of the backbone network. At this point, the information captured by APSNet is coarse-grained, sensitive only to small-scale targets. The second training phase ends when passing through the third-layer bottleneck. In this step, the coarse-grained features are further processed; the expansion of APSNet's receptive field and the enhancement of features at the language level improve the recognition performance of medium-scale targets. This is because it is desirable to explicitly process spatial information, such as size.
[0042] Therefore, Head_Com is chosen in the third stage because it contains the most information. For example... Figure 3 As shown, the channel field branch in Head_Com is consistent with the first and second level classification headers.
[0043] Specifically, the output from the backbone is processed through max pooling. Large-scale feature maps are generated to produce a batch of 1024-dimensional vectors, which are then used to generate classification predictions via an MLP. The channel-domain output is optimized using the cross-entropy loss function. , in, It is the number of categories. It is a sign function (0 or 1). The observation belongs to the category The predicted probability, where N is the number of channels.
[0044] Therefore, spatial domain branching is used to adjust the channel information and retain spatial information as an additional loss term.
[0045] Specifically, Head_Com first will The features are mapped to a 1000-dimensional vector for each sample to control computational complexity. Then, for each sample, one channel dimension is selected from the 1000-channel vector to represent spatial information. This step is implemented using an MLP operation, preserving the spatial representation of the dimensions. After applying L2 normalization to project the vector onto a unit hypersphere, the cosine similarity matrix is calculated. Subsequently, the corresponding penalty matrix is constructed to obtain the loss value.
[0046] in, This represents the L2-normalized features. Indicates the first Feature vectors of seed samples Indicates the first The label of each sample Represents the total number of samples. Masks representing pairs of samples of the same class (positive sample pairs: ), Masks representing different classes of sample pairs (negative sample pairs: ).
[0047] Therefore, Head_Com can be represented as: , in, This represents the traditional category header, identical to Head_Con, Head1, and Head2. Indicates average pooling. This indicates the output of Backbone.
[0048] Using spatial distance as a pseudo-monitoring signal, the comparison loss is evaluated. Combining cross-entropy loss from a dual-domain perspective and spatial distance loss Generate the final monitoring signal: , Finally, to balance the predictions across the three scales, APSNet employs a standard multi-head architecture.
[0049] The fourth training phase concatenates the 3-scale feature maps along the channel dimension to generate a fused classification prediction, while combining the outputs of the four classification heads to produce the final prediction. Therefore, the total loss can be expressed as: , , , , in, This indicates the losses in the first phase. This indicates the losses in the second phase. This indicates the losses in the third phase. This indicates the losses in the fourth stage; and This represents the cross-entropy loss between Head1 and Head2. This represents the concatenation cross-entropy loss.
[0050] Most training methods complete feature supervision at different granularities through only a single backpropagation signal transmission. However, for ancient plant seeds, the complex feature distribution requires targeted learning of features at different granularities. Progressive training methods can achieve this, but the lack of spatial domain signal supervision prevents them from co-supervising the "size-texture" relationship. Therefore, an asynchronous dual-domain classification (ADD) structure is proposed to further explore the "channel-space" two-dimensional discriminative cues. In the first and second stages, the ADD structure maintains the traditional classification structure. In the third stage, channel domain voting and spatial domain comparison are used to calculate the supervision signal. The fourth stage uses multi-head fusion and multi-scale prediction to preserve global semantics while enhancing local details. This progressive learning amplifies subtle differences and significantly improves APSNet's ability to identify morphologically close categories.
[0051] Extensive experiments on large-scale ancient plant seed datasets demonstrate that this invention provides comparative results between APSNet and the best algorithms in three classification domains: the traditional GhostNetv2 method, the LOS method for long-tailed distribution classification tasks, and the FET-FGVC method for fine-grained classification tasks, as shown in Table 1. Table 1 Comparative Experiment
[0052] As can be seen, the proposed method achieves the best classification performance, improving accuracy and F1 score by 5.3% and 12.1% respectively compared to the previous best method.
[0053] Example 2 This embodiment provides an intelligent identification system for carbonized plant seeds from archaeological sites, including: The high-frequency information extraction module is used to extract high-frequency information from the acquired carbonized plant seed image data. The multi-granularity feature learning module is used to affinely fuse the original carbonized plant seed image and the extracted high-frequency information into the backbone network channel by channel, guiding the backbone network to learn multi-granularity features. The contrastive learning module is used to learn features of different granularities in layers based on multi-granularity features and a trained multi-head classifier. It simultaneously learns global and local discriminative features at different levels and fuses the discriminative features learned at each layer to obtain the classification result.
[0054] It should be noted that the specific implementation of the intelligent identification system for carbonized plant seeds in this embodiment of the invention is similar to the specific implementation of the intelligent identification method for carbonized plant seeds from an archaeological site in this embodiment of the invention. Please refer to the description in the method section for details. To reduce redundancy, it will not be repeated here.
[0055] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent identification method for carbonized plant seeds from an archaeological site as described above.
[0056] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the above-described method for intelligent identification of carbonized plant seeds from an archaeological site.
[0057] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0058] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0059] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0060] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0061] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for intelligent identification of carbonized plant seeds from archaeological sites, characterized in that, Includes the following steps: Acquire carbonized plant seed image data and extract high-frequency information from the carbonized plant seed image data; The original carbonized plant seed images and extracted high-frequency information are affinely fused into the backbone network channel by channel, guiding the backbone network to learn multi-granular features. Based on multi-granularity features and a trained multi-head classifier, features of different granularities are learned hierarchically. Discriminative features of the global domain and local domain are learned simultaneously at different levels. The discriminative features learned at each level are then fused to obtain the classification result.
2. The intelligent identification method for carbonized plant seeds at archaeological sites as described in claim 1, characterized in that, The extraction of high-frequency information from carbonized plant seed image data includes: Fourier transform is applied to convert the carbonized plant seed image to the frequency domain to obtain the spectral representation of the original image; Based on the spectral representation of the original image, high-frequency components are separated using a frequency domain mask to obtain the high-frequency components; High-frequency information in the spatial domain is obtained by reversing the high-frequency components to return to the spatial domain.
3. The method for intelligent identification of carbonized plant seeds at archaeological sites as described in claim 1, characterized in that, The process involves channel-by-channel affine fusion of the original carbonized plant seed image and extracted high-frequency information into the backbone network, guiding the backbone network to learn multi-granular features, including: Based on the extracted high-frequency information and the preprocessed image, an initial feature channel affine transformation is performed. The resulting affine feature map is then input into the backbone network to obtain the backbone feature map. The obtained affine feature map and backbone feature map are then subjected to multiple feature channel affine transformations and corresponding backbone network feature extractions to obtain the affine feature map and backbone feature map of each level. This includes passing the affine feature map and backbone feature map of the previous level through the next level channel affine transformation and corresponding backbone network feature extraction to obtain the affine feature map and backbone feature map of the next level.
4. The intelligent identification method for carbonized plant seeds at archaeological sites as described in claim 3, characterized in that, The process of extracting the next-level affine feature map and backbone feature map from the previous-level affine feature map and backbone feature map through the next-level channel affine transformation and the corresponding backbone network feature extraction includes: Connect the affine feature map and the backbone feature map of the previous level along the channel dimension, and attach the fused information to the last layer of the backbone feature map of this level. By processing the fused features through channel interactions, the network searches for pairwise relationships between the original features and size information of the backbone through interactive operations. Specifically, each channel is compressed into a scalar to represent its global importance, and local convolutions are applied along the channel dimension to model the interaction between adjacent channels. The resulting scalar interaction matrix is projected back to the cascaded feature map, thereby establishing size pairing between the fused information and each channel. Convolution is used to simulate MLP operations, and channel-based affine transformations are performed. Based on the calculated pairing information, the size and texture difference information are combined to obtain the affine feature map and backbone feature map of the corresponding level.
5. The intelligent identification method for carbonized plant seeds at archaeological sites as described in claim 1, characterized in that, The hierarchical learning of features of different granularities based on multi-granularity features and a trained multi-head classifier includes: The backbone feature map extracted at the fourth level is input into the first classification head, and the backbone feature maps extracted at the second and third levels are input into the corresponding second and third classification heads. Then, the scale feature maps output by each classifier are concatenated and input into the fourth classification head. Finally, the results of the four classification heads are fused to generate the final prediction.
6. The method for intelligent identification of carbonized plant seeds from archaeological sites as described in claim 5, characterized in that, The first classification head adopts a dual-domain classification head and designs a dual-domain voting mechanism to obtain classification results from both channel and spatial directions simultaneously.
7. The intelligent identification method for carbonized plant seeds at archaeological sites as described in claim 1, characterized in that, The training process of the multi-head classifier include: The first stage of training ends when the features pass through the second layer of the backbone network, obtaining coarse-grained features that are sensitive to small-scale targets. The second stage of training ends when the features pass through the three-layer bottleneck structure, and the coarse-grained features are processed to improve the recognition effect of mesoscale targets. The third stage of training uses a combined classification head, Head_Com, which processes both channel domain information and spatial domain information to generate a composite supervision signal that fuses channel domain classification loss and spatial domain contrast loss. The fourth stage of training concatenates the multi-scale features obtained in the first three stages along the channel dimension and generates the final classification prediction by fusing the classification head.
8. A smart identification system for carbonized plant seeds from archaeological sites, characterized in that, include: The high-frequency information extraction module is used to extract high-frequency information from the acquired carbonized plant seed image data. The multi-granularity feature learning module is used to affinely fuse the original carbonized plant seed image and the extracted high-frequency information into the backbone network channel by channel, guiding the backbone network to learn multi-granularity features. The contrastive learning module is used to learn features of different granularities in layers based on multi-granularity features and a trained multi-head classifier. It simultaneously learns global and local discriminative features at different levels and fuses the discriminative features learned at each layer to obtain the classification result.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for intelligent identification of carbonized plant seeds at archaeological sites as described in any one of claims 1-7.
10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for intelligent identification of carbonized plant seeds at archaeological sites as described in any one of claims 1-7.