Tooth detection and FDI identification system and method

Abstract

The invention discloses a tooth detection and FDI recognition system and method. The tooth detection and FDI recognition system comprises a mixed frequency domain-space attention module, a Transform decoding module, a dental arch perception query graph priori module and a category uniqueness gating module. According to the invention, the mixed frequency domain-space attention module improves the tooth structure characterization capability under the conditions of light reflection, fuzziness, noise and shielding, and improves the detection robustness and positioning precision, thereby supporting the Transform decoding module to carry out accurate positioning and preliminary classification; the dental arch perception query graph priori submodule and the category uniqueness gating submodule cooperate with the Transform decoding module to work, and jointly ensure that the final output of the Transform decoding module meets the clinical logic requirements of dental arch topological continuity and FDI number global uniqueness.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis of oral images and computer vision technology, specifically to a tooth detection and FDI recognition system and method. Background Technology

[0002] Oral imaging plays a vital role in clinical diagnosis and treatment, automated electronic dental record (EDR) entry, and remote dental care. To achieve structured interpretation of oral images, it is typically necessary to first locate and standardize the numbering of teeth. Compared to natural scene images, intraoral photographs / intraoral scan renderings / panoramic images often exhibit specular reflections, overexposure, motion blur, occlusion, similar gingival and tooth colors, and significant variations in perspective and scale, leading to instability in tooth margins and texture features. Furthermore, the dense density and similar morphology of teeth exacerbate confusion in detection and classification.

[0003] Currently, the methods for structured interpretation of oral images include: Anchor-based detectors (such as Faster R-CNN, SSD, YOLO, etc.) typically output a large number of candidate boxes and use non-maximum suppression (NMS) to remove duplicates. Then, each box is independently classified to obtain a tooth position number. Because teeth are closely spaced and have similar shapes, parameters such as NMS threshold and anchor scale need to be frequently tuned. In reflective / blurred / occluded scenes, candidate box scores are unstable, which can easily lead to false suppression or retention of incorrect boxes. Furthermore, multiple boxes of the same FDI category may be predicted repeatedly, requiring additional regularization post-processing (such as sorting by x-coordinate, grouping by dental arch, and retaining the highest score for duplicate categories) to correct these issues.

[0004] Transformer detectors such as DETR and its deformable DETR model object detection as ensemble prediction, using Hungarian matching for one-to-one training supervision, thus reducing dependence on anchors and NMS. However, these models typically predict class distributions independently and lack explicit constraints on "globally unique FDI numbers" and "dental arch topological order," making them prone to duplicate numbering, cross-dental arch confusion, and local sequence swapping under oral imaging noise and cross-device domain offset.

[0005] Some dental methods introduce adjacency graphs / relationship branches or perform graph optimization and sequence matching during the inference stage to correct the numbering. However, these methods often rely on the initial detection quality, have complex structures, high inference overhead, and are unstable in graph edge construction under strong noise conditions, which can easily lead to error propagation.

[0006] Therefore, it is necessary to develop a new method for tooth detection and identification under complex oral imaging conditions to improve the accuracy of tooth detection. Summary of the Invention

[0007] To address the aforementioned issues, this invention aims to provide a tooth detection and FDI identification system and method. The hybrid frequency domain-spatial attention module of this invention enhances the representation of tooth structures under conditions of reflection, blur, noise, and occlusion, thereby improving detection robustness and positioning accuracy. This supports the Transformer decoding module in performing accurate positioning and preliminary classification. The dental arch perception query map prior submodule and the category uniqueness gating submodule work in conjunction with the Transformer decoding module to ensure that the final output of the Transformer decoding module meets the clinical logical requirements of dental arch topological continuity and globally unique FDI numbers.

[0008] This invention is achieved through the following technical solution: A tooth detection and FDI recognition system includes: a hybrid frequency domain-spatial attention module (HFSA), a Transformer decoding module, a dental arch perception query graph prior module (QGP), and a category uniqueness gating module (UniGate); The hybrid frequency domain-spatial attention module receives oral image input feature maps and enhances the oral image feature maps flowing through that location; The Transformer decoding module contains a fixed number of learnable queries. It iteratively updates the queries through a multi-layer decoder and uses the enhanced features of the hybrid frequency domain-spatial attention module as encoding memory to output the tooth detection box and category prediction corresponding to each query. The dental arch perception query map prior module works in conjunction with the Transformer decoding module to construct query adjacency relationships during the self-attention phase of the Transformer decoding module and generate an attention mask based on the reference point of the query to restrict query interaction to a local neighborhood that conforms to the dental arch anatomy. The category uniqueness gating module works in conjunction with the Transformer decoding module to apply a differentiable unique constraint to the category prediction output by the Transformer decoding module during the model training phase, so as to suppress the same tooth category from being predicted repeatedly by multiple queries.

[0009] The hybrid frequency-space attention module includes frequency-domain and spatial branches for parallel processing; The frequency domain branch is used to: perform a fast Fourier transform on the input features to obtain the spectrum; decompose the spectrum according to a preset frequency band threshold to obtain low-frequency components and at least one high-frequency band component; after performing an inverse Fourier transform on each frequency band component, generate spatial modulation weights through depthwise separable convolution, perform weighted modulation on the corresponding spatial domain frequency band components, and finally merge all modulated frequency band components to obtain frequency domain enhanced features. The spatial branching is used to: extract strip direction features by using mutually orthogonal strip convolution kernels, and integrate them to obtain spatial enhancement features, taking into account the characteristics of the transverse extension of the dental arch and the longitudinal structure of the tooth. Frequency domain enhancement features and spatial enhancement features are fused through a selective kernel fusion mechanism, and then output through residual connection.

[0010] The construction of the dental arch perception query graph prior module for query adjacency relationships specifically includes one of the following two methods: 1) Calculate the Euclidean distance between any two query reference points, and select the k+1 nearest queries for each i as the allowed attention set. The k+1 nearest queries may include themselves. 2) Based on the ordinate of the reference point of the current decoding layer query vector, the query is divided into the upper dental arch group and the lower dental arch group; within each group, a one-dimensional parameter along the dental arch curve of each query is estimated based on the reference point coordinate; based on the one-dimensional parameter, k nearest neighbor queries are selected for each query within its own dental arch group to construct a k nearest neighbor graph; a self-attention mask is generated based on the k nearest neighbor graph, and in the self-attention calculation of the Transformer decoding module, only the query is allowed to focus on its k nearest neighbors within the same dental arch group, blocking interaction with other queries.

[0011] The category uniqueness gating module is used to construct an approximately one-to-one soft matching matrix between the set of FDI categories that appear in the current image and the main query set, and to add the log probability of the matrix as a penalty term to the unnormalized classification score (logits).

[0012] It also includes an input and preprocessing module: receiving oral images X (RGB intraoral photos or grayscale panoramic images, etc.), performing image scaling, normalization, optional brightness / contrast correction and cropping, to obtain the network oral image input tensor.

[0013] It also includes an output module, which receives the output of the Transformer decoding module, obtains the normalized bounding box and FDI category confidence, and restores the normalized bounding box to pixel coordinates according to the input image size, and finally outputs the location and numbering results of each tooth.

[0014] It also includes a multi-scale feature extraction network, which is used to extract multi-scale feature maps of the input oral cavity image and output multi-scale feature maps to the hybrid frequency domain-spatial attention module. The multi-scale feature extraction network includes a convolutional neural network backbone and a feature fusion network FPN / PAN. The convolutional neural network backbone extracts multi-layer feature maps C3, C4, C5, etc. from the input oral cavity image. The feature fusion network FPN / PAN further fuses the multi-layer feature maps to obtain multi-scale feature maps P3, P4, P5, etc., to take into account different scale targets such as anterior teeth / molars.

[0015] The reference point is the normalized two-dimensional coordinate point corresponding to each query, used for deformable attention sampling and bounding box regression initialization; Deformable attention is a mechanism for aggregating attention on a small number of sampled points across a multi-scale feature map.

[0016] A method for tooth detection and FDI identification includes the following steps: acquiring oral images to be analyzed; The oral images are input into the dental detection and FDI recognition system as described in the preceding item; The output of the tooth detection and FDI recognition system is analyzed to obtain the bounding box of each tooth in the oral image and its corresponding FDI number.

[0017] An electronic device comprising the tooth detection and FDI recognition system as described above.

[0018] An electronic storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the tooth target detection method as described above. The electronic storage medium is a non-volatile computer-readable storage medium, such as a solid-state drive (SSD), hard disk drive (HDD), read-only memory (ROM), flash memory, or cloud storage.

[0019] Compared with the prior art, the present invention has the following advantages and beneficial effects: The hybrid frequency domain-spatial attention module of this invention improves the ability to represent tooth structures under conditions of reflection, blur, noise, and occlusion, enhancing detection robustness and localization accuracy, thereby supporting the Transformer decoding module in accurate localization and preliminary classification. The arch perception query map prior submodule and the category uniqueness gating submodule work in conjunction with the Transformer decoding module to ensure that the final output of the Transformer decoding module meets the clinical logical requirements of arch topological continuity and globally unique FDI numbering. In summary, this invention achieves the following significant effects: 1. Reduced bounding box regression error, lower false negative and false positive rates, and improved detection accuracy and robustness; 2. Better conformity to the globally unique FDI and anatomical order logic, significantly reducing duplicate numbering, cross-arch confusion, and left-right order disorder; and 3. Clinically usable detection and numbering results can be obtained without complex heuristic post-processing, improving deployment stability and cross-device generalization ability. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 : Schematic diagram of the overall system framework of this invention; Figure 2 : Schematic diagram of the HFSA hybrid frequency domain-spatial attention module structure; Figure 3 : Schematic diagram of dental arch QGP query graph prior and attention mask generation; Figure 4 Flowchart of the method for dental examination and FDI numbering; Figure 5 UniGate category uniqueness gating diagram. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0022] Example 1

[0023] A tooth detection and FDI recognition system includes: a hybrid frequency domain-spatial attention module, a Transformer decoding module, a dental arch perception query map prior module, and a category uniqueness gating module; The hybrid frequency domain-spatial attention module receives oral image input feature maps and enhances the oral image feature maps flowing through that location; The Transformer decoding module contains a fixed number of learnable queries. It iteratively updates the queries through a multi-layer decoder and uses the enhanced features of the hybrid frequency domain-spatial attention module as encoding memory to output the tooth detection box and category prediction corresponding to each query. The dental arch perception query map prior module works in conjunction with the Transformer decoding module to generate an attention mask based on the reference point of the query during the self-attention phase of the Transformer decoding module, so as to restrict the query interaction to a local neighborhood that conforms to the dental arch anatomy. The category uniqueness gating module works in conjunction with the Transformer decoding module to apply a differentiable unique constraint to the category prediction output by the Transformer decoding module during the model training phase.

[0024] Specifically, such as Figure 2 As shown, the HFSA module includes parallel frequency domain branches and spatial branches, and performs channel-level adaptive weighting through selective kernel fusion (SK-Fusion). Finally, it outputs enhanced features through residual methods.

[0025] (1) Frequency domain branch (FBMLite): The input feature x∈R^{B×C×H×W} is subjected to a two-dimensional Fast Fourier Transform (FFT) in the frequency domain to obtain the spectrum Xf. Based on a preset set of frequency band thresholds, a concentric low-pass / band-pass mask is constructed to decompose Xf stepwise, obtaining low-frequency components and several high-frequency band components. Each frequency band component is then subjected to an inverse Fourier Transform (iFFT) back to the spatial domain, and spatial weights w_k(x) are generated using depthwise separable convolution and modulated to form frequency-domain enhanced features. x_freq = x_low + Σ_k w_k(x) ⊙ band_k(x).

[0026] Preferably, the spatial weights are activated by Sigmoid and scaled to the (0,2) range; a switch can be set to control whether the low-frequency components are also modulated (enabled in the preferred embodiment); to improve numerical stability, FFT / iFFT and weight calculation adopt a high-precision floating-point format (such as float32) to ensure numerical stability.

[0027] (2) Spatial branch (StripBlock): Based on the characteristics of the lateral extension of the dental arch and the longitudinal structure of the tooth, the spatial branch uses depth convolution in the (1×k) and (k×1) directions (preferably k=11) to extract the strip direction features, and then integrates them through 1×1 convolution to obtain x_sp.

[0028] (3) SK fusion and residual output: Global average pooling is applied to x_freq and x_sp, followed by two 1×1 convolutions to generate 2C channel weights. Softmax is then applied to obtain the channel weights α and β for the two branches, achieving x_fused = α⊙x_freq + β⊙x_sp. Finally, a 1×1 convolution out_proj is used to map back to C channels, and the output is performed using a residual method. x_out = x + tanh(a) · out_proj(x_fused), Where 'a' is the learnable residual scaling parameter (the initial value of the residual scaling parameter is set to 0), so that it remains stable in the early stage of training and gradually improves in the later stage.

[0029] Specifically, such as Figure 3 As shown, the dental arch perception query graph prior module, i.e. the QGP module, is used to construct sparse query adjacency relationships during the Transformer decoding stage, and generate a self-attention mask M based on this, thereby restricting query interactions to a locally anatomically reasonable range.

[0030] (1) Input and data cleaning: QGP receives the reference point P_ref∈[0,1]^{B×Q×2} (two-dimensional normalized coordinates) from the main query. To avoid the influence of outliers, it is preferable to replace the reference point with a non-numeric (NaN) value and truncate its numerical range.

[0031] (2) Adjacency relationship construction strategy: QGP provides at least two optional strategies: Strategy A (Global Euclidean KNN): Calculate the Euclidean distance D_ij=||p_i-p_j||_2 between any two query reference points, and select the (k+1) nearest queries (including itself) for each i as the allowed attention set N_k(i); Strategy B (arch-aware grouped KNN, enabled in the preferred embodiment): The query is divided into upper arch group (y<0.5) and lower arch group (y≥0.5) based on the reference point ordinate y; within each arch group, a one-dimensional parameter t_i along the arch is estimated for each query (preferably obtained through an arch curve fitting function; when the number of points is insufficient, it can degenerate to using the x coordinate as t_i), and KNN is performed on the one-dimensional distance |t_i-t_j| to select the nearest neighbor set N_k(i).

[0032] (3) Mask generation and attention injection: Construct a Boolean mask M∈{0,1}^{B×Q×Q}, where M_{ij}=0 indicates that i is allowed to focus on j, and M_{ij}=1 indicates that focus is blocked; then convert it into an additive mask A, where the blocking position is assigned a value of -∞ and the allowed position is 0, and inject it into the multi-head self-attention.

[0033] (4) Gated residual update: To avoid excessive prior strength in the early stages of training leading to convergence difficulties, a learnable gating parameter λ (initially set to -5) is preferably set, and the QGP update strength is controlled by σ(λ): q ← q + σ(λ)·(q_qgp -q). This QGP update preferably applies only to the main query and not to the denoised query, and can be enabled only in some decoding layers (e.g., layers 1 and 3) to balance stability and constraint strength.

[0034] (5) Numerical stability: To avoid overflow in half-precision, mask transformation and QGP self-attention calculation preferably use float32 and automatic mixed precision is turned off during mixed precision training.

[0035] Specifically, the category uniqueness gating module, such as Figure 5As shown, UniGate is used to impose differentiability constraints on the uniqueness logic of "each FDI category appears at most once within the same image" during the training phase. Its core idea is to construct an approximately one-to-one soft matching matrix between the set of FDI categories that appear within the current image and the main query set, and write back the log probability of the matrix as a penalty term to the classification logits.

[0036] (1) Input: The main query classification score matrix S∈R^{Qm×C} (Qm is the number of main queries, and C is the number of FDI categories). Based on the annotation, the set of categories where the image appears is obtained, C_b⊆{1…C}, and the molecular matrix S_b∈R^{Qm×|C_b|} is extracted.

[0037] (2) Sinkhorn row and column normalization: Transpose S_b to |C_b|×Qm and perform temperature scaling to obtain Z=S_b^T / τ_t. Perform K Sinkhorn iterations in the logarithmic domain (the number of iterations K can be set to 5): alternately normalize the rows and columns with log-sum-exp so that exp(Z) approximates the birandom matrix P∈[0,1]^{|C_b|×Qm}, which can be interpreted as a soft permutation relationship of "category-query". Optionally, multi-inquiry heads can be used for parallel computation. (3) Logarithmic penalty write-back: Transpose log(P) back to Qm×|C_b| and weight it according to the intensity coefficient s_t to form Δ=s_t·log(P). Write back only for the category dimensions that appear: S_b ← S_b + Δ. Since log(P) is negative for non-occupied positions, it can naturally suppress the logits of repeatedly occupied categories, thereby inhibiting the same FDI category from being predicted with high scores by multiple queries at the same time.

[0038] (4) Parameter scheduling: To ensure training stability, τ_t and s_t are automatically scheduled according to the training progress. Preferably, τ_t is gradually reduced from τ_init=1.0 to τ_end=0.05 by exponential annealing; s_t adopts a warm-up strategy: the first 5% of steps are set to 0 (UniGate is disabled), and then the intensity is linearly increased to the maximum intensity (the intensity coefficient can be set to 1.0) in the 5% to 15% of steps, and then remains unchanged.

[0039] (5) Scope of application: UniGate is preferably applied only to the main query and not to the denoised query (which is segmented based on the length information of the denoised query). The inference phase can be turned off by default to avoid extra calculations, or it can be turned on as an optional prior and post-processing as needed.

[0040] This invention employs an ensemble prediction training paradigm. For each training image, the decoder outputs Qm main query predictions (and optional denoised query predictions). The optimal one-to-one matching relationship M between the prediction set and the labeled tooth set is solved using Hungarian matching, ensuring that each labeled tooth matches at most one query.

[0041] The preferred detection loss consists of classification loss and regression loss: (1) Classification loss L_cls: For matched queries, calculate the cross-entropy / focus loss of their FDI category; unmatched queries are trained as "no target" categories; (2) Regression loss L_reg: For the matched query, calculate the L1 distance between the predicted box and the ground truth box; (3) Shape consistency loss L_giou: For the matched query, calculate the generalized IoU loss to improve the positioning accuracy.

[0042] The total loss can be expressed as: L = λ_cls·L_cls + λ_1·L_reg + λ_giou·L_giou + L_aux.

[0043] L_aux is an optional auxiliary loss, including: auxiliary supervision of the decoder intermediate layer, denoising training loss, and optional dental arch sequence consistency loss (e.g., penalizing the FDI numbering order according to the estimated t_i sequence within each dental arch).

[0044] During training, UniGate performs uniqueness gating on the classification logits, and QGP applies topological constraints to the query self-attention, thereby simultaneously optimizing detection accuracy and numbering consistency in end-to-end training.

[0045] In implementation, an output module may also be included to receive the output of the Transformer decoding module, process the tooth detection boxes and category predictions, and output the normalized or pixel-based bounding box of each tooth, its FDI number, and confidence score. The output set can be filtered according to a confidence score threshold. Specifically, the inference output and interface are as follows: During the inference phase, after the oral images to be analyzed are input into the system, the output is a set of tooth detection results, each of which contains at least: (1) Tooth bounding box (bbox); (2) FDI number (or its category probability distribution); (3) Confidence score.

[0046] Preferably, low-confidence results can be filtered by a score threshold, and the output can be sorted by dental arch (upper / lower) and along the dental arch parameter t, which facilitates direct writing into electronic medical records or visualization annotation. Since this invention uses an ensemble prediction method, deduplication results can usually be obtained without NMS; if duplicate boxes still exist in specific scenarios, simple IoU threshold merging can be performed on the output set as an alternative post-processing.

[0047] In implementation, it may also include an input and preprocessing module: receiving oral images X (RGB intraoral photographs or grayscale panoramic films, etc.), performing image scaling, normalization, optional brightness / contrast correction and cropping, to obtain the network oral image input tensor.

[0048] In implementation, a multi-scale feature extraction network is also included to extract multi-scale feature maps from the input oral cavity image and output multi-scale feature maps to the hybrid frequency domain-spatial attention module. The multi-scale feature extraction network includes a convolutional neural network backbone and a feature fusion network FPN / PAN. The convolutional neural network backbone extracts multi-layer feature maps C3, C4, C5, etc. from the input oral cavity image. The feature fusion network FPN / PAN further fuses the multi-layer feature maps to obtain multi-scale feature maps P3, P4, P5, etc., to take into account different scale targets such as anterior teeth / molars.

[0049] Example 2

[0050] A tooth detection and FDI numbering system (which can be deployed as a software system on a server, workstation or mobile terminal, or integrated into an image acquisition device as an oral image analysis device) is used to output the bounding box and corresponding FDI number of each tooth from the input oral image.

[0051] like Figure 1 As shown, the system includes at least: (1) Input and preprocessing module: Receives oral images X (RGB intraoral photos or grayscale panoramic films, etc.), performs image scaling, normalization, optional brightness / contrast correction and cropping, and obtains network input tensors; (2) Multi-scale feature extraction module: includes a convolutional neural network backbone and a feature fusion network FPN / PAN. The backbone extracts multi-layer feature maps C3, C4, C5, etc. from the input image; the feature fusion network FPN / PAN uses a dynamic upsampling operator (DySample) to replace traditional interpolation, generates upsampling kernels based on content awareness, and further fuses them to obtain multi-scale feature maps (P3, P4, P5) to restore tooth edge details while maintaining low computational cost; (3) Hybrid Frequency Domain-Spatial Attention Enhancement Module (HFSA): Configured at the encoding end (e.g., the output of the selected Backbone / Encoder layer) and / or each output layer of FPN to enhance multi-scale features, highlight tooth edges and detailed textures, and suppress reflections and noise; (4) Transformer decoding module: contains a fixed number of queries (Query Set, QS), which are iteratively updated through a multi-layer decoder. The decoder preferably uses deformable cross attention, which samples and aggregates on multi-scale feature maps, and outputs the bounding box regression results and FDI classification distribution corresponding to each query; (5) Query Graph Prior Module (QGP): During the self-attention phase of the decoder, a dental arch k-nearest neighbor graph is constructed based on the query reference point and an attention mask is generated to restrict query interactions to occur only in the local neighborhood of the same dental arch. (6) Uniqueness gate module UniGate: During the training phase, a differentiable unique constraint is applied to the classification logits, so that each FDI category is occupied by a unique main query, reducing duplicate numbering conflicts. (7) Output module: Outputs the bounding box of each tooth (e.g., normalized form of (cx,cy,w,h) or pixel form of (x1,y1,x2,y2), FDI number, and confidence level, and can filter the output set according to the confidence level threshold. An electronic device includes the tooth detection and FDI recognition system as described above. An electronic storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the tooth target detection method as described above. The electronic storage medium is a non-volatile computer-readable storage medium, such as a solid-state drive (SSD), hard disk drive (HDD), read-only memory (ROM), flash memory, or cloud storage space.

[0052] Example 3

[0053] A method for tooth detection and FDI numbering, which can be executed by the system of Embodiment 2 above, is as follows: Figure 4 As shown, the preferred method includes the following steps: S101: Acquire the oral cavity image X to be analyzed and preprocess it to obtain the network input; S102: The preprocessed input is fed into the multi-scale feature extraction module to obtain a multi-layer feature map; S103: The multi-layer feature maps are enhanced using the HFSA module to obtain an enhanced feature pyramid; S104: Initialize a fixed number of query vectors and their reference points, and use the enhanced feature pyramid as the input encoding memory to the Transformer decoder; S105: In each decoding layer, the query is first updated through the parallel update unit and then fused in the "query fusion" unit. Then, the self-attention update is constrained by the attention mask generated by the QGP module to achieve query interaction with consistent dental arch topology. S106: During the training phase, UniGate uniqueness gate is applied to the logits output by the classification branch to form a soft one-to-one occupancy constraint on the FDI categories appearing in the current image. S107: Output the bounding box and FDI category distribution for each main query, and obtain the final tooth detection box and its FDI number after threshold filtering.

[0054] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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 tooth detection and FDI identification system, characterized in that, include: Hybrid frequency domain-spatial attention module Transformer decoding module, dental arch perception query graph prior module, and category uniqueness gating module; The hybrid frequency domain-spatial attention module receives oral image input feature maps and enhances the oral image feature maps flowing through that location; The Transformer decoding module contains a fixed number of learnable queries. It iteratively updates the queries through a multi-layer decoder and uses the enhanced features of the hybrid frequency domain-spatial attention module as encoding memory to output the tooth detection box and category prediction corresponding to each query. The dental arch perception query map prior module works in collaboration with the Transformer decoding module to construct query adjacency relationships during the self-attention phase of the Transformer decoding module and generate an attention mask based on the reference point of the query to restrict query interaction to a local neighborhood that conforms to the dental arch anatomy. The category uniqueness gating module works in conjunction with the Transformer decoding module to impose a differentiable unique constraint on the category prediction output by the Transformer decoding module during the model training phase, in order to prevent the same tooth category from being predicted repeatedly by multiple queries.

2. The tooth detection and FDI identification system according to claim 1, characterized in that, Mixed frequency domain - space The inter-attention module includes frequency domain branches and spatial branches for parallel processing; The frequency domain branch is used to: perform a fast Fourier transform on the input features to obtain the spectrum; decompose the spectrum according to a preset frequency band threshold to obtain low-frequency components and at least one high-frequency band component; perform an inverse Fourier transform on each frequency band component to obtain spatial domain frequency band features; at the same time, based on the input features, generate spatial modulation weights through depthwise separable convolution, perform weighted modulation on the features of each frequency band component, and finally merge all the modulated frequency band components to obtain frequency domain enhanced features. The spatial branching is used to: extract strip direction features by using mutually orthogonal strip convolution kernels, and integrate them to obtain spatial enhancement features, taking into account the characteristics of the transverse extension of the dental arch and the longitudinal structure of the tooth. Frequency domain enhancement features and spatial enhancement features are fused through a selective kernel fusion mechanism, and then output through residual connection.

3. The tooth detection and FDI identification system according to claim 1, characterized in that, The dental arch sensing The construction of the query graph prior module for query adjacency relationships specifically includes one of the following two methods: 1) Calculate the Euclidean distance between any two query reference points, and select the k+1 nearest queries for each i as the allowed attention set. The k+1 nearest queries may include themselves. 2) Based on the ordinate of the reference point of the current decoding layer query vector, the query is divided into the upper dental arch group and the lower dental arch group; within each group, a one-dimensional parameter along the dental arch curve of each query is estimated based on the reference point coordinate; based on the one-dimensional parameter, k nearest neighbor queries are selected for each query within its own dental arch group to construct a k nearest neighbor graph; a self-attention mask is generated based on the k nearest neighbor graph, and in the self-attention calculation of the Transformer decoding module, only the query is allowed to focus on itself and its k nearest neighbors within the same dental arch group, blocking interaction with other queries.

4. The tooth detection and FDI identification system according to claim 1, characterized in that, The category is unique. The sex-gating module is used to construct an approximately one-to-one soft matching matrix between the set of FDI categories that appear in the current image and the main query set, and to add the log probability of the matching matrix as a penalty term to the classification unnormalized score (logits).

5. The tooth detection and FDI identification system according to claim 1, characterized in that, It also includes input and preprocessing. Processing module: Receives oral images, performs size scaling, normalization, optional brightness / contrast correction and cropping on the oral images, and obtains the network oral image input tensor.

6. The tooth detection and FDI identification system according to claim 1, characterized in that, It also includes an output module. It is used to receive the output of the Transformer decoding module, process the tooth detection box and category prediction, output the normalized form or pixel form of the bounding box of each tooth, FDI number and confidence score, and can filter the output set according to the confidence score threshold.

7. The tooth detection and FDI identification system according to claim 1, characterized in that, It also includes multi-scale features The extraction network is used to extract multi-scale feature maps from the input oral cavity image and output multi-scale feature maps to the hybrid frequency domain-spatial attention module. The multi-scale feature extraction network includes a convolutional neural network backbone and a feature fusion network FPN / PAN. The convolutional neural network backbone extracts multi-layer feature maps from the input oral cavity image. The feature fusion network FPN / PAN further fuses the multi-layer feature maps to obtain multi-scale feature maps.

8. A method for tooth detection and FDI identification, characterized in that, Includes the following steps: Obtain oral images to be analyzed; The oral imaging is input into the dental detection and FDI recognition system as described in any one of claims 1-7; The output of the tooth detection and FDI recognition system is analyzed to obtain the bounding box of each tooth in the oral image and its corresponding FDI number.

9. An electronic device, characterized in that, Includes the dental detection and FDI identification system as described in any one of claims 1-7.

10. An electronic storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the tooth target detection method as described in claim 8.

Images