Occlusal Model Center Point Detection Method and Device Based on Prior Knowledge

Abstract

The present invention discloses a method and device for detecting the center point of a dental model based on prior knowledge, including: performing pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; extracting shallow feature information of sample image information through the image encoder and determining a target feature map corresponding to the sample image information; generating aggregated feature information for each pair of teeth based on second prior knowledge for judging tooth categories and the target feature map and through a deep detection network; and generating a center point prediction result for the pair of teeth according to the aggregated feature information of each pair of teeth. It can be seen that implementing the present invention can achieve intelligent positioning and intelligent detection of the tooth center points of the dental model by combining prior knowledge, which is beneficial to improving the accuracy of detecting the tooth center points and the efficiency of detecting the tooth center points.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent detection, and particularly relates to a method and device for detecting the center point of a dental model based on prior knowledge. Background Art

[0002] In the prior art, the calibration of the tooth center point can greatly improve the diagnosis speed of orthodontic dentists for patients' conditions, and can improve the treatment experience of patients. The calibration of the tooth center point has been widely applied in the field of dental medicine. For example, in orthodontic treatment, it is necessary to check the tooth loss situation and tooth arrangement status of patients, and classify and locate each tooth. However, currently, the calibration of the tooth center point usually requires dentists to manually calibrate medical images, which not only has the problem of low positioning efficiency but also has the problem of low positioning accuracy. It can be seen that it is particularly important to provide a new center point detection method to improve the efficiency and accuracy of calibrating the tooth center point. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and device for detecting the center point of a dental model based on prior knowledge, which can realize the intelligent positioning and intelligent detection of the tooth center point of the dental model by combining prior knowledge, is beneficial to improving the accuracy of tooth center point detection, and is beneficial to improving the efficiency of tooth center point detection.

[0004] To solve the above technical problem, in the first aspect of the present invention, a method for detecting the center point of a dental model based on prior knowledge is disclosed, and the method includes:

[0005] Performing a pre-training operation on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; wherein, the first prior knowledge is used to perform a pre-training operation on the image encoder so that the image encoder performs a corresponding self-learning operation;

[0006] Extracting shallow feature information of sample image information through the image encoder, and determining a target feature map corresponding to the sample image information based on the shallow feature information, wherein the target feature map includes feature maps of each tooth pair included in the sample image information;

[0007] Generating aggregated feature information of each tooth pair based on second prior knowledge and the target feature map, and through a deep detection network; wherein, the second prior knowledge includes pre-determined prior knowledge for judging tooth categories;

[0008] For each of the tooth pairs, a center point prediction result for the tooth pair is generated according to the aggregated feature information of the tooth pair; wherein, the center point prediction result includes classification result information of the tooth pair and / or a center point prediction heat map of the tooth pair.

[0009] As an alternative implementation manner, in the first aspect of the present invention, before performing a pre-training operation on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence, the method further includes:

[0010] Obtain a set of sample image information, wherein the set of sample image information includes a plurality of sample image information, and the set of sample image information includes positive sample image information and negative sample image information. Among them, the positive sample image information includes dental arch model image information and tooth region rendering image information from the same patient, and the negative sample image information includes dental arch model image information and tooth region rendering image information from different patients;

[0011] The performing a pre-training operation on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence includes:

[0012] Perform a pre-training operation on a pre-determined image encoder based on first prior knowledge and the set of sample image information to obtain a pre-training result, and perform an update operation on the image encoder based on the pre-training result to obtain an image encoder trained to convergence, so as to constrain the shallow output of the image encoder;

[0013] Wherein, the updated image encoder includes prior knowledge information for distinguishing tooth regions and gum regions.

[0014] As an alternative implementation manner, in the first aspect of the present invention, after generating a center point prediction result for each of the tooth pairs according to the aggregated feature information of the tooth pair, the method further includes:

[0015] Obtain a target center point heat map, and the target center point heat map includes a center point heat map for each of the tooth pairs;

[0016] Generate a target center point prediction heat map according to the center point prediction heat map for each of the tooth pairs;

[0017] Calculate a difference parameter between the target center point heat map and the target center point prediction heat map, and based on the difference parameter, perform an update on the deep detection network to update the deep detection network;

[0018] Among them, the difference parameters include one or more of a semantic overlap penalty loss parameter, a cross-entropy loss parameter, and a mean squared error loss parameter, and the parameter update operation includes a stochastic gradient descent algorithm operation.

[0019] As an alternative implementation, in the first aspect of the present invention, the shallow feature information of the sample image information includes the shallow feature information of each tooth pair included in the sample image;

[0020] Among them, determining the target feature map corresponding to the sample image information based on the shallow feature information includes:

[0021] Determine each tooth pair included in the sample image information. For each tooth pair, input the shallow feature information of the tooth pair into a preset multi-head attention model so that the multi-head attention model outputs a spatial attention map of the region where the tooth pair is located;

[0022] For each tooth pair, calculate the feature map of the tooth pair according to the spatial attention map of the region where the tooth pair is located and the preset weighted network deep features;

[0023] Determine the target feature map corresponding to the sample image information according to the feature maps of all the tooth pairs;

[0024] Among them, the multi-head attention model includes at least one lightweight encoder, and each lightweight encoder has a semantic overlap penalty loss function matching the lightweight encoder. The semantic overlap penalty loss functions corresponding to different lightweight encoders are different, and the semantic overlap penalty loss function is used to constrain the lightweight encoder corresponding to the semantic overlap penalty loss function to collect the regional feature information of the semantic region corresponding to the lightweight encoder.

[0025] As an alternative implementation, in the first aspect of the present invention, generating the aggregated feature information of each tooth pair based on the second prior knowledge and the target feature map through a deep detection network includes:

[0026] For each pair of tooth pairs, determine the target feature information from the feature map of the tooth pair according to the second prior knowledge information and the target feature map, and the number of the target feature information is at least one;

[0027] For each pair of tooth pairs, perform a preset feature aggregation operation on all the target feature information of the tooth pair through a deep detection network to obtain the aggregated feature information of the tooth pair;

[0028] Among them, the feature aggregation operation includes one or more of a feature splicing operation and a fully connected layer aggregation operation.

[0029] As an alternative implementation, in the first aspect of the present invention, for each of the tooth pairs, generating a center point prediction result for the tooth pair according to the aggregated feature information of the tooth pair includes:

[0030] For each of the tooth pairs, performing a calculation operation on the aggregated feature information of the tooth pair through at least one pre-determined convolutional layer according to the aggregated feature information of the tooth pair to obtain a feature calculation result of the tooth pair;

[0031] For each pair of the tooth pairs, determining a center point prediction heat map of the tooth pair and classification result information of the tooth pair according to the feature calculation result of the tooth pair.

[0032] As an alternative implementation, in the first aspect of the present invention, the method further includes:

[0033] Inputting the determined target dental arch model rendering image into the pre-determined deep detection network, and performing a feature extraction operation on the target dental arch model rendering image through the deep detection network to obtain target rendering feature information;

[0034] Based on the target rendering feature information and the deep detection network, generating a target center point prediction result matching the target dental arch model rendering image; wherein, the target center point prediction result includes center point position information of each target tooth corresponding to the target dental arch model rendering image and center point classification information of each target tooth corresponding to the target dental arch model rendering image.

[0035] The second aspect of the present invention discloses a dental arch model center point detection device based on prior knowledge, and the device includes:

[0036] A training module, configured to perform pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; wherein, the first prior knowledge is used to perform pre-training operations on the image encoder so that the image encoder performs corresponding self-learning operations;

[0037] An extraction module, configured to extract shallow feature information of sample image information through the image encoder;

[0038] A determination module, configured to determine a target feature map corresponding to the sample image information based on the shallow feature information, wherein the target feature map includes feature maps of each tooth pair included in the sample image information;

[0039] A generation module, configured to generate aggregated feature information for each of the tooth pairs based on second prior knowledge and the target feature map through a deep detection network; wherein, the second prior knowledge includes prior knowledge determined in advance for judging tooth categories; for each of the tooth pairs, a center point prediction result of the tooth pair is generated according to the aggregated feature information of the tooth pair; wherein, the center point prediction result includes classification result information of the tooth pair and / or a center point prediction heat map of the tooth pair.

[0040] As an optional implementation manner, in the second aspect of the present invention, the apparatus further includes:

[0041] An acquisition module, configured to acquire a set of sample image information before the training module performs a pre-training operation on a predetermined image encoder based on first prior knowledge to obtain an image encoder trained to convergence, wherein the set of sample image information includes a plurality of sample image information, and the set of sample image information includes positive sample image information and negative sample image information, wherein the positive sample image information includes dental arch model image information and tooth region rendering image information from the same patient, and the negative sample image information includes dental arch model image information and tooth region rendering image information from different patients;

[0042] The specific manner in which the training module performs a pre-training operation on a predetermined image encoder based on first prior knowledge to obtain an image encoder trained to convergence includes:

[0043] Performing a pre-training operation on a predetermined image encoder based on first prior knowledge and the set of sample image information to obtain a pre-training result, and performing an update operation on the image encoder based on the pre-training result to obtain an image encoder trained to convergence, so as to constrain the shallow output of the image encoder;

[0044] Wherein, the updated image encoder includes prior knowledge information for distinguishing tooth regions and gum regions.

[0045] As an optional implementation manner, in the second aspect of the present invention, the acquisition module is further configured to, after the generation module generates a center point prediction result for each of the tooth pairs according to the aggregated feature information of the tooth pair, acquire a target center point heat map, where the target center point heat map includes the center point heat map of each of the tooth pairs; and generate a target center point prediction heat map according to the center point prediction heat map of each of the tooth pairs;

[0046] The apparatus further includes:

[0047] A calculation module, configured to calculate a difference parameter between the target center point heat map and the predicted heat map of the target center point, and based on the difference parameter, perform a network update operation on the deep detection network to update the deep detection network;

[0048] Wherein, the difference parameter includes one or more of a semantic overlap penalty loss parameter, a cross-entropy loss parameter, and a mean square error loss parameter, and the parameter update operation includes a stochastic gradient descent algorithm operation.

[0049] As an optional implementation manner, in the second aspect of the present invention, the shallow feature information of the sample image information includes the shallow feature information of each pair of teeth included in the sample image;

[0050] The specific manner in which the determination module determines the target feature map corresponding to the sample image information based on the shallow feature information includes:

[0051] Determine each pair of teeth included in the sample image information. For each pair of teeth, input the shallow feature information of the pair of teeth into a preset multi-head attention model, so that the multi-head attention model outputs a spatial attention map of the area where the pair of teeth is located;

[0052] For each pair of teeth, calculate the feature map of the pair of teeth according to the spatial attention map of the area where the pair of teeth is located and the preset weighted network deep feature;

[0053] Determine the target feature map corresponding to the sample image information according to the feature maps of all the pairs of teeth;

[0054] Wherein, the multi-head attention model includes at least one lightweight encoder, and each lightweight encoder has a semantic overlap penalty loss function matching the lightweight encoder. The semantic overlap penalty loss functions corresponding to different lightweight encoders are different, and the semantic overlap penalty loss function is used to constrain the lightweight encoder corresponding to the semantic overlap penalty loss function to collect the regional feature information of the semantic region corresponding to the lightweight encoder.

[0055] As an optional implementation manner, in the second aspect of the present invention, the specific manner in which the generation module generates the aggregated feature information of each pair of teeth based on the second prior knowledge and the target feature map through the deep detection network includes:

[0056] For each pair of teeth, determine target feature information from the feature map of the pair of teeth according to the second prior knowledge information and the target feature map, and the number of the target feature information is at least one;

[0057] For each pair of the tooth pairs, a preset feature aggregation operation is performed on all the target feature information of the tooth pair through a deep detection network to obtain the aggregated feature information of the tooth pair;

[0058] Wherein, the feature aggregation operation includes one or more of a feature splicing operation and a fully connected layer aggregation operation.

[0059] As an optional implementation manner, in the second aspect of the present invention, the specific manner in which the generation module generates the center point prediction result of each tooth pair according to the aggregated feature information of the tooth pair includes:

[0060] For each tooth pair, a calculation operation is performed on the aggregated feature information of the tooth pair through at least one pre-determined convolutional layer according to the aggregated feature information of the tooth pair to obtain the feature calculation result of the tooth pair;

[0061] For each pair of the tooth pairs, according to the feature calculation result of the tooth pair, the center point prediction heat map of the tooth pair and the classification result information of the tooth pair are determined.

[0062] As an optional implementation manner, in the second aspect of the present invention, the extraction module is further configured to input the determined target dental arch model rendering diagram into the pre-determined deep detection network, and perform a feature extraction operation on the target dental arch model rendering diagram through the deep detection network to obtain target rendering feature information;

[0063] The generation module is further configured to generate a target center point prediction result matching the target dental arch model rendering diagram based on the target rendering feature information and the deep detection network; wherein, the target center point prediction result includes the center point position information of each target tooth corresponding to the target dental arch model rendering diagram and the center point classification information of each target tooth corresponding to the target dental arch model rendering diagram.

[0064] A third aspect of the present invention discloses another dental arch model center point detection device based on prior knowledge, and the device includes:

[0065] A memory storing executable program code;

[0066] A processor coupled to the memory;

[0067] The processor calls the executable program code stored in the memory and executes the dental arch model center point detection method based on prior knowledge disclosed in the first aspect of the present invention.

[0068] A fourth aspect of the present invention discloses a computer storage medium storing computer instructions which, when called, are used to execute the prior - knowledge - based dental arch model center point detection method disclosed in the first aspect of the present invention.

[0069] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0070] In the embodiments of the present invention, a pre - training operation is performed on a pre - determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence. The shallow - layer feature information of the sample image information is extracted by the image encoder, and the target feature map corresponding to the sample image information is determined. Based on the second prior knowledge for determining tooth categories and the target feature map, which are pre - determined, and through a deep detection network, the aggregated feature information of each tooth pair is generated. For each tooth pair, according to the aggregated feature information of the tooth pair, a center point prediction result of the tooth pair is generated, where the center point prediction result includes classification result information of the tooth pair and / or a center point prediction heat map of the tooth pair. It can be seen that implementing the present invention can realize intelligent positioning and intelligent detection of the tooth center points of the dental arch model by combining prior knowledge, which is beneficial to improving the accuracy of tooth center point detection and the efficiency of tooth center point detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0072] Figure 1 is a schematic flowchart of a prior - knowledge - based dental arch model center point detection method disclosed in an embodiment of the present invention;

[0073] Figure 2 is a schematic flowchart of another prior - knowledge - based dental arch model center point detection method disclosed in an embodiment of the present invention;

[0074] Figure 3 is a schematic structural diagram of a prior - knowledge - based dental arch model center point detection device disclosed in an embodiment of the present invention;

[0075] Figure 4 is a schematic structural diagram of another prior - knowledge - based dental arch model center point detection device disclosed in an embodiment of the present invention;

[0076] Figure 5It is a schematic structural diagram of another dental model center point detection device based on prior knowledge disclosed in the embodiments of the present invention. Detailed implementation manners

[0077] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0078] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or terminal that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or terminals.

[0079] Referring to "embodiment" herein means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The phrase appears in various positions in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0080] The present invention discloses a dental model center point detection method and device based on prior knowledge, which can realize intelligent positioning and intelligent detection of the tooth center points of the dental model in combination with prior knowledge, is beneficial to improving the accuracy of tooth center point detection, and is beneficial to improving the efficiency of tooth center point detection. The following will be described in detail respectively.

[0081] Embodiment 1

[0082] Please refer to Figure 1 , Figure 1 It is a schematic flowchart of a dental model center point detection method based on prior knowledge disclosed in the embodiments of the present invention. Among them, Figure 1 The described dental model center point detection method based on prior knowledge can be applied to a dental model center point detection device based on prior knowledge, or can also be applied to a local server or a cloud server for dental model center point detection based on prior knowledge. The embodiments of the present invention do not make limitations. As Figure 1As shown, the method for detecting the center point of the dental arch model based on prior knowledge may include the following operations:

[0083] 101. Perform pre-training on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence.

[0084] In an embodiment of the present invention, the first prior knowledge is used to perform pre-training on the image encoder so that the image encoder performs corresponding self-learning operations.

[0085] In an embodiment of the present invention, optionally, the first prior knowledge includes prior knowledge that enables the image encoder to distinguish between the tooth region and the gingival region.

[0086] In an embodiment of the present invention, optionally, by performing pre-training on the image encoder with the first prior knowledge, the image encoder trained to convergence can be used to distinguish between the region corresponding to the gingival part included in the tooth alignment and the region corresponding to the tooth part.

[0087] In an embodiment of the present invention, optionally, the shallow layer of the encoder network masters the prior knowledge of distinguishing between the tooth region and the gingival region after pre-training; further, the parameters of the shallow layer of the encoder network can be frozen in subsequent training to better preserve the pre-training knowledge; at this time, in subsequent training, the image encoder only needs to encode the rendering of the dental arch model; if it is necessary to continue to update the parameters to better adapt to the downstream tasks of tooth center point positioning and classification, then in subsequent training, a set of renderings of the dental arch model and a rendering of the tooth region are still input simultaneously, and the loss function of the shallow layer of the encoder is calculated.

[0088] 102. Extract the shallow feature information of the sample image information through the image encoder, and determine the target feature map corresponding to the sample image information based on the shallow feature information.

[0089] In an embodiment of the present invention, the target feature map includes the feature maps of each tooth pair included in the sample image information.

[0090] In an embodiment of the present invention, the feature map of each tooth pair can be determined through a shallow feature algorithm and a pre-determined weighted deep feature network.

[0091] In an embodiment of the present invention, optionally, the feature information extracted by the shallow feature algorithm is relatively similar to the input information, and the feature information extracted by the shallow feature algorithm contains more information of pixel points. For example, the feature information extracted by the shallow feature algorithm contains some color, texture, edge, and corner information of the image; further, the receptive field of the shallow feature algorithm is small, and the overlapping area of the receptive field is also small, which can ensure capturing more details.

[0092] In an embodiment of the present invention, optionally, a weighted deep feature network is used to extract deep feature information, and the feature information extracted by the weighted deep feature network is relatively close to the output. Moreover, the feature information extracted by the weighted deep feature network contains some coarse-grained information and more abstract information, that is, semantic information. Further, the receptive field of the weighted deep feature network is increased, the overlapping area between receptive fields is increased, the image information is compressed, and the weighted deep feature network obtains some information about the integrity of the image, and stronger semantic information can be obtained.

[0093] 103. Based on the second prior knowledge and the target feature map, and through a deep detection network, aggregate feature information of each tooth pair is generated.

[0094] In an embodiment of the present invention, the second prior knowledge includes the prior knowledge determined in advance for judging tooth categories.

[0095] In an embodiment of the present invention, optionally, the second prior knowledge may be determined according to the prior knowledge of human dentists for judging tooth categories.

[0096] In an embodiment of the present invention, optionally, the prior knowledge information is used to fuse the deep tooth feature maps. Further, the prior knowledge information can constrain the symmetry of the tooth positions and can constrain the adjacent judgment criteria for tooth classification. Further, the prior knowledge information can also endow the network with the tooth part and the gum part.

[0097] 104. For each tooth pair, according to the aggregate feature information of the tooth pair, a center point prediction result of the tooth pair is generated.

[0098] In an embodiment of the present invention, the center point prediction result includes the classification result information of the tooth pair and / or the center point prediction heat map of the tooth pair.

[0099] In an embodiment of the present invention, further optionally, after generating the center point prediction result of each tooth pair according to the aggregate feature information of the tooth pair for each tooth pair, the method may further include:

[0100] Performing a permutation operation on all tooth pairs according to the center point prediction heat map of each tooth pair to obtain a permutation result, and generating a target prediction heat map based on the permutation result, where the target prediction heat map includes the center point prediction heat maps after all teeth are arranged.

[0101] It can be seen that the implementation Figure 1The described prior-knowledge-based dental model center point detection method can perform pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; extract shallow feature information of sample image information through the image encoder and determine the target feature map corresponding to the sample image information; generate aggregated feature information for each tooth pair based on the second prior knowledge for judging tooth categories and the target feature map, and through a deep detection network; generate a center point prediction result for the tooth pair according to the aggregated feature information for each tooth pair. It can perform an image enhancement process on the tooth part and the gum part through the first prior knowledge to initially distinguish the area corresponding to the tooth part and the area corresponding to the gum part, and extract and fuse the deep feature information through the second prior knowledge to generate the center point prediction result for the tooth pair. It can achieve intelligent positioning and intelligent detection of the tooth center points in the dental model based on different types of prior knowledge, and can synergistically promote the improvement of the detection quality of the tooth center points through prior knowledge, a deep detection network, and the predicted heat map, which is beneficial to improving the accuracy of detecting the tooth center points and the efficiency of detecting the tooth center points.

[0102] Embodiment 2

[0103] Please refer to Figure 2 , Figure 2 , which is a schematic flowchart of another prior-knowledge-based dental model center point detection method disclosed in the embodiments of the present invention. Among them, Figure 2 The described prior-knowledge-based dental model center point detection method can be applied to a prior-knowledge-based dental model center point detection device, or to a local server or a cloud server for prior-knowledge-based dental model center point detection. The embodiments of the present invention do not make any limitations. As Figure 2 shown, the prior-knowledge-based dental model center point detection method may include the following operations:

[0104] 201. Obtain a set of sample image information.

[0105] In the embodiments of the present invention, the set of sample image information includes several pieces of sample image information. The set of sample image information includes positive sample image information and negative sample image information. Among them, the positive sample image information includes dental model image information and tooth region rendering image information from the same patient, and the negative sample image information includes dental model image information and tooth region rendering image information from different patients.

[0106] In an embodiment of the present invention, optionally, for example, assume that there are B "original model rendering images" from different patients during training. At this time, each "original model rendering image" will have a corresponding "tooth part rendering image". Thus, the possible combination ways of the input images of the network are B*B kinds. Since each "original model rendering image" and its own "tooth part rendering image" come from the same patient, the number of positive sample pairs is B at this time, and the number of negative sample pairs is B^2 - B. This example is the simplest case during training. In actual operation, the sources of the B "original model rendering images" are not restricted.

[0107] In an embodiment of the present invention, further optionally, the way to obtain the sample image information set can be obtained in real time, can be obtained at regular intervals according to a preset time period, or can be obtained when a training operation needs to be performed on the image encoder. The embodiment of the present invention does not make specific limitations.

[0108] 202. Based on the first prior knowledge and the sample image information set, perform a pre-training operation on the pre-determined image encoder to obtain a pre-training result, and based on the pre-training result, perform an update operation on the image encoder to obtain an image encoder trained to convergence, so as to constrain the shallow output of the image encoder.

[0109] In an embodiment of the present invention, further optionally, the positive sample image information and the negative sample image information are used to perform corresponding training operations on the image encoder, and while performing the corresponding training operations on the image encoder, calculate the shallow loss function of the image encoder.

[0110] 203. Through the image encoder, extract the shallow feature information of the sample image information, and based on the shallow feature information, determine the target feature map corresponding to the sample image information.

[0111] 204. Based on the second prior knowledge and the target feature map, and through the deep detection network, generate the aggregated feature information of each tooth pair.

[0112] 205. For each tooth pair, generate the center point prediction result of the tooth pair according to the aggregated feature information of the tooth pair.

[0113] In an embodiment of the present invention, for the detailed description of steps 203 - 205, please refer to the other descriptions of steps 102 - 104 in Embodiment 1. The embodiment of the present invention will not elaborate.

[0114] It can be seen that the implementation Figure 2The described method for detecting the center point of the dental arch model based on prior knowledge can perform pre-training operations on the image encoder by obtaining a set of sample image information containing several positive sample image information and negative sample image information, and can use the positive sample image information and negative sample image information to perform corresponding training operations on the image encoder, which is beneficial to improving the accuracy of the shallow loss function of the calculated image encoder, thereby improving the accuracy and intelligence of the training operations performed on the image encoder, and is beneficial to improving the accuracy of the image encoder trained to convergence, and further beneficial to improving the subsequent intelligent positioning and intelligent detection of the tooth center points in the dental arch model based on different types of prior knowledge, and is beneficial to being able to synergistically promote the improvement of the detection quality of the tooth center points through prior knowledge, a deep detection network, and the predicted heat map, beneficial to improving the accuracy of the detection of the tooth center points, and beneficial to improving the efficiency of the detection of the tooth center points.

[0115] In an optional embodiment, for each pair of teeth, after generating the center point prediction result of the pair of teeth according to the aggregated feature information of the pair of teeth, the method further includes:

[0116] Obtain a target center point heat map, where the target center point heat map includes the center point heat map of each pair of teeth;

[0117] Generate a target center point prediction heat map according to the center point prediction heat map of each pair of teeth;

[0118] Calculate the difference parameter between the target center point heat map and the target center point prediction heat map, and based on the difference parameter, perform a network update operation on the deep detection network to update the deep detection network;

[0119] Wherein, the difference parameter includes one or more of a semantic overlap penalty loss parameter, a cross-entropy loss parameter, and a mean square error loss parameter, and the parameter update operation includes a stochastic gradient descent algorithm operation.

[0120] In this optional embodiment, optionally, the difference parameter between the target center point heat map and the target center point prediction heat map can be calculated by a multi-layer perceptron; wherein, the multi-layer perceptron uses the calculation method of the cross-entropy loss function for the specific tooth category of each pair of teeth predicted, and calculates the difference parameter between the true category and the true category of each pair of teeth; it should be noted that the multi-layer perceptron (MLP, Multilayer Perceptron) is a feedforward artificial neural network model that maps multiple input data sets to a single output data set.

[0121] In this optional embodiment, optionally, the purpose of the semantic overlap penalty loss is to constrain the attention heads to focus on different regions and avoid different attention heads from repeatedly extracting the same information; the cross entropy loss function (Cross Entropy Loss) is a commonly used loss function in machine learning, which can be used to measure the distance between two distributions. The significance of the cross entropy loss function is that it can help the machine learning model estimate the prediction probability density function that is closest to the target distribution from the training samples; the mean squared error loss parameter is a commonly used loss function in machine learning, and its role is to measure the gap between the model's predicted value and the true value. In machine learning, training data is usually used to train the model, and then test data is used to evaluate the model's performance, and the mean squared error loss function parameter is used to evaluate the model's performance on the test data.

[0122] It can be seen that implementing this optional embodiment can obtain the target center point heat map, predict the heat map according to the center points of each tooth pair to generate the target center point prediction heat map, calculate the difference parameter between the target center point heat map and the target center point, and perform a network update operation on the deep detection network based on the difference parameter to update the deep detection network, which is beneficial to improving the accuracy and reliability of subsequent deep feature information extraction through the deep detection network, and is also beneficial to improving the intelligence and efficiency of subsequent deep feature information extraction through the deep detection network. Thus, the deep detection network and the predicted heat map cooperate to promote the improvement of the center point detection quality, which is further beneficial to improving the accuracy of tooth center point detection and the efficiency of tooth center point detection.

[0123] In another optional embodiment, the shallow feature information of the sample image information includes the shallow feature information of each tooth pair included in the sample image;

[0124] Among them, determining the target feature map corresponding to the sample image information based on the shallow feature information includes:

[0125] Determine each tooth pair included in the sample image information. For each tooth pair, input the shallow feature information of the tooth pair into a preset multi-head attention model so that the multi-head attention model outputs the spatial attention map of the region where the tooth pair is located;

[0126] For each tooth pair, calculate the feature map of the tooth pair according to the spatial attention map of the region where the tooth pair is located and the preset weighted network deep features;

[0127] Determine the target feature map corresponding to the sample image information according to the feature maps of all tooth pairs;

[0128] Among them, the multi-head attention model includes at least one lightweight encoder. For each lightweight encoder, there is a semantic overlap penalty loss function that matches this lightweight encoder. The semantic overlap penalty loss functions corresponding to different lightweight encoders are different. The semantic overlap penalty loss function is used to constrain the lightweight encoder corresponding to this semantic overlap penalty loss function to collect the regional feature information of the semantic region corresponding to this lightweight encoder.

[0129] In this optional embodiment, optionally, the preset multi-head attention model includes multiple groups of lightweight encoders with non-shared parameters. Among them, the lightweight encoder with non-shared parameters can be a single-layer convolutional network. Further, each lightweight encoder is an attention head. Each lightweight encoder inputs the shallow features of the image encoder and outputs the spatial attention map of the region where a specific pair of teeth is located. After weighting the spatial attention map with the deep features of the encoder network, the local features of a specific pair of teeth can be obtained.

[0130] In this optional embodiment, further optionally, to constrain each attention head to focus on regions with different semantics, each attention head is constrained by the semantic overlap penalty loss function. To improve the stability of the training process and accelerate the network convergence speed, residual connections are used for the deep features of the encoder network to ensure that the output result of the multi-head attention module does not lose the rich semantic information generated by the image encoder. Further, the purpose of the semantic overlap penalty loss is to constrain the attention heads to focus on different regions and avoid different attention heads from repeatedly extracting the same information.

[0131] It can be seen that implementing this optional embodiment can input the shallow feature information of each pair of teeth into the preset multi-head attention model so that the multi-head attention model outputs the spatial attention map of the region where this pair of teeth is located, and calculate the feature map of each pair of teeth according to the spatial attention map of the region where each pair of teeth is located and the preset weighted network deep features. It is possible to calculate the feature map of each pair of teeth through the preset multi-head attention model and the preset weighted network deep features, which is beneficial to improving the intelligence and efficiency of outputting the spatial attention map of the region where each pair of teeth is located, and is beneficial to improving the accuracy and reliability of outputting the spatial attention map of the region where each pair of teeth is located. Thus, it is beneficial to improve the accuracy and reliability of calculating the feature map of each pair of teeth, and further beneficial to improving the accuracy and efficiency of subsequent detection of the tooth center point, and also beneficial to improving the quality of detecting the tooth center point.

[0132] In another optional embodiment, based on the second prior knowledge and the target feature map, and through the deep detection network, the aggregated feature information of each pair of teeth is generated, including:

[0133] For each pair of teeth, according to the second prior knowledge information and the target feature map, target feature information is determined from the feature map of the pair of teeth, and the number of the target feature information is at least one;

[0134] For each pair of teeth, a preset feature aggregation operation is performed on all the target feature information of the pair of teeth through a deep detection network to obtain the aggregated feature information of the pair of teeth;

[0135] Wherein, the feature aggregation operation includes one or more of a feature splicing operation and a fully connected layer aggregation operation.

[0136] In this optional embodiment, optionally, the aggregated feature information of each pair of teeth includes the classification information of the pair of teeth and the predicted center point position information of the pair of teeth.

[0137] In this optional embodiment, optionally, the second prior knowledge can be used to perform a corresponding feature extraction operation on the deep feature information. Further optionally, the object for determining the target feature information can be determined by the objective reasoning logic and the second prior knowledge corresponding to a human dentist. Further, the target feature information is information that can be used to judge the classification category of the pair of teeth and the information of the predicted center point position of the pair of teeth.

[0138] In this optional embodiment, further optionally, the feature splicing operation includes simply performing a simple linear transformation on the text modality to convert it into the feature vector size required by the image modality, and then splicing the text feature vector and the image feature vector. Further, the fully connected layer aggregation operation includes performing an aggregation operation on all the target feature information through a fully connected layer to obtain the aggregated feature information.

[0139] It can be seen that implementing this optional embodiment can determine the target feature information from the feature map of each pair of teeth according to the pre-determined prior knowledge information, and perform a preset feature aggregation operation according to the target feature information of the pair of teeth to obtain the aggregated feature information of the pair of teeth, which can improve the accuracy and reliability of determining the target feature information of each pair of teeth, and is beneficial to improving the intelligence and efficiency of determining the target feature information of each pair of teeth. Thus, the aggregated feature information of each pair of teeth can also be obtained through the preset feature aggregation operation, which is beneficial to improving the accuracy and intelligence of obtaining the aggregated feature information of each pair of teeth, and further beneficial to improving the accuracy and reliability of generating the center point prediction heat map of each pair of teeth based on the feature information of each pair of teeth, and is beneficial to improving the accuracy and detection quality of detecting the center point of each pair of teeth.

[0140] In another optional embodiment, for each pair of teeth, generating a center point prediction result for the pair of teeth includes:

[0141] For each pair of teeth, based on the aggregated feature information of the pair of teeth, perform a calculation operation on the aggregated feature information of the pair of teeth through at least one pre-determined convolutional layer to obtain a feature calculation result of the pair of teeth;

[0142] For each pair of teeth, based on the feature calculation result of the pair of teeth, determine the center point prediction heat map of the pair of teeth and the classification result information of the pair of teeth.

[0143] In this optional embodiment, optionally, for each pair of teeth, based on the aggregated feature information of the pair of teeth, perform a calculation operation on the aggregated feature information of the pair of teeth through at least one pre-determined convolutional layer to obtain a feature calculation result of the pair of teeth, including:

[0144] For each pair of teeth, based on the aggregated feature information of the pair of teeth, through a preset deformable convolution model and through a preset projection model, predict the center point area of the pair of teeth through a plurality of convolutional layers, so as to realize that the intermediate features are calculated through a plurality of stacked convolutional layers to obtain a final feature map, where the final feature map is regarded as the predicted center point heat map;

[0145] Among them, the preset deformable convolution model includes a plurality of convolutional layers, and the preset deformable convolution model is used to expand the receptive field of the image decoder without increasing the amount of computation.

[0146] In this optional embodiment, optionally, the number of convolutional layers can be one or multiple; the embodiments of the present invention do not make specific limitations; further, the more the number of convolutional layers, the more accurate the feature calculation result of the pair of teeth obtained.

[0147] In this optional embodiment, further optionally, after completing the prediction of one stage through multiple convolutional layers, the technical solution splices the prediction result with the original input of the decoder and then inputs it into the next stage. Therefore, each stage is based on the prediction result of the previous stage to continuously improve the prediction effect, and finally realizes the prediction of the center point position with high precision.

[0148] It can be seen that implementing this optional embodiment can input the aggregated feature information of each tooth pair into a pre-determined image decoder to obtain the target aggregated feature information of the tooth pair, and perform a calculation operation on the target aggregated feature information of each tooth pair through a pre-determined convolutional layer to obtain the feature calculation result of each tooth pair, and determine the center point prediction heat map of each tooth pair based on the feature calculation result of each tooth pair. It can specifically determine the target aggregated feature information of each tooth pair and specifically determine the feature calculation result of each tooth pair, which is beneficial to improving the accuracy and reliability of determining the target aggregated feature information of each tooth pair and determining the feature calculation result of each tooth pair, and is beneficial to improving the intelligence and efficiency of determining the target aggregated feature information of each tooth pair and determining the feature calculation result of each tooth pair. Thus, it can specifically determine the center point prediction heat map of each tooth pair, which is beneficial to improving the accuracy and reliability of determining the center point prediction heat map of each tooth pair, and is beneficial to improving the intelligence and efficiency of determining the center point prediction heat map of each tooth pair. Furthermore, it is beneficial to realize the intelligent positioning and intelligent detection of tooth center points based on the dental arch model, beneficial to improving the accuracy of detecting tooth center points, and beneficial to improving the efficiency of detecting tooth center points.

[0149] In yet another optional embodiment, the method further includes:

[0150] Input the determined target dental arch model rendering into a pre-determined deep detection network, and perform a feature extraction operation on the target dental arch model rendering through the deep detection network to obtain target rendering feature information;

[0151] Generate a target center point prediction result matching the target dental arch model rendering based on the target rendering feature information and the deep detection network; wherein, the target center point prediction result includes the center point position information of each target tooth corresponding to the target dental arch model rendering and the center point classification information of each target tooth corresponding to the target dental arch model rendering.

[0152] In this optional embodiment, optionally, inputting the determined target dental arch model rendering into a pre-determined deep detection network, and performing a feature extraction operation on the target dental arch model rendering through the deep detection network to obtain target rendering feature information includes:

[0153] Input the determined target dental arch model rendering into a pre-determined deep detection network, and perform a feature extraction operation on the target dental arch model rendering through the deep detection network to obtain target shallow layer feature information, and determine the target local feature information of the target tooth pairs included in the target dental arch model rendering according to the target shallow layer feature information, and generate target rendering feature information according to all the target local feature information.

[0154] In this optional embodiment, optionally, generating a center point prediction result matching the target dental arch model rendering diagram based on the target rendering feature information and the deep detection network includes:

[0155] Based on the target rendering feature information and the deep detection network, determining the tooth pair classification information of each target tooth pair included in the target dental arch model rendering diagram and the tooth pair center point position information of each target tooth pair included in the target dental arch model rendering diagram;

[0156] Based on the tooth pair classification information of each target tooth pair included in the target dental arch model rendering diagram and the tooth pair center point position information of each target tooth pair included in the target dental arch model rendering diagram, generating a center point prediction result of the target dental arch model rendering diagram.

[0157] In this optional embodiment, further optionally, before inputting the determined target dental arch model rendering diagram into the pre-determined deep detection network, the method may further include:

[0158] Obtaining the network loss parameter of the deep detection network, and determining whether the network loss parameter meets the preset model loss condition;

[0159] When it is determined that the network loss parameter meets the preset model loss condition, triggering the operation of inputting the determined target dental arch model rendering diagram into the pre-determined deep detection network, and performing a feature extraction operation on the target dental arch model rendering diagram through the deep detection network to obtain the target rendering feature information;

[0160] When it is determined that the network loss parameter does not meet the preset model loss condition, determining the target network parameters in the deep detection network, determining the network update parameters of each target network parameter, and updating the target network parameters based on the network update parameters to update the deep detection network, and re-triggering the operation of obtaining the network loss parameter of the deep detection network and determining whether the network loss parameter meets the preset model loss condition;

[0161] Among them, determining whether the network loss parameter meets the preset model loss condition includes:

[0162] Determining the network loss value corresponding to the network loss parameter;

[0163] Determining whether the network loss value is greater than or equal to the preset network loss threshold;

[0164] When it is determined that the network loss value is greater than or equal to the preset network loss threshold, determining that the network loss parameter does not meet the preset model loss condition;

[0165] When it is determined that the network loss value is less than a preset network loss threshold, it is determined that the network loss parameter meets the preset model loss condition.

[0166] It can be seen that implementing this optional embodiment can input the target dental arch model rendering diagram into a pre-determined deep detection network, perform feature extraction operations on the target dental arch model rendering diagram through the deep detection network to obtain target rendering feature information, and generate a center point prediction result matching the target dental arch model rendering diagram based on the target rendering feature information and the deep detection network. It is possible to obtain target rendering feature information and then generate a center point prediction result based on the deep detection network trained to convergence, which is beneficial to improving the intelligence and efficiency of obtaining the target rendering feature information and the center point prediction result, and further beneficial to improving the accuracy and reliability of the target rendering feature information and the center point prediction result. It can also realize intelligent positioning and intelligent detection of the tooth center point based on the dental arch model, which is beneficial to improving the accuracy of tooth center point detection and the efficiency of tooth center point detection.

[0167] Embodiment III

[0168] Please refer to Figure 3 , Figure 3 which is a schematic structural diagram of a dental arch model center point detection device based on prior knowledge disclosed in an embodiment of the present invention. As Figure 3 shown, the dental arch model center point detection device based on prior knowledge may include:

[0169] A training module 301, configured to perform pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; wherein, the first prior knowledge is used to perform pre-training operations on the image encoder so that the image encoder performs corresponding self-learning operations;

[0170] An extraction module 302, configured to extract shallow feature information of sample image information through the image encoder;

[0171] A determination module 303, configured to determine a target feature map corresponding to the sample image information based on the shallow feature information, wherein the target feature map includes feature maps of each pair of teeth included in the sample image information;

[0172] A generation module 304, configured to generate aggregated feature information of each pair of teeth based on second prior knowledge and the target feature map, and through a deep detection network; wherein, the second prior knowledge includes pre-determined prior knowledge for judging tooth categories; for each pair of teeth, a center point prediction result of the pair of teeth is generated according to the aggregated feature information of the pair of teeth; wherein, the center point prediction result includes classification result information of the pair of teeth and / or a center point prediction heat map of the pair of teeth.

[0173] It can be seen that implementing Figure 3 the described device can perform pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; extract shallow feature information of sample image information through the image encoder, and determine a target feature map corresponding to the sample image information; generate aggregated feature information for each tooth pair based on second prior knowledge for judging tooth categories and the target feature map, and through a deep detection network; generate a center point prediction result for the tooth pair according to the aggregated feature information of each tooth pair. It can perform an image enhancement process on the tooth part and the gum part through the first prior knowledge to initially distinguish the area corresponding to the tooth part and the area corresponding to the gum part, and extract and fuse the deep feature information through the second prior knowledge to generate a center point prediction result for the tooth pair. It can realize intelligent positioning and intelligent detection of the tooth center points in the dental arch model based on different types of prior knowledge, and can synergistically promote the improvement of the detection quality of the tooth center points through prior knowledge, a deep detection network, and the predicted heat map, which is beneficial to improving the accuracy of detecting the tooth center points and is beneficial to improving the efficiency of detecting the tooth center points.

[0174] In an optional embodiment, as Figure 4 shown, the device further includes:

[0175] An acquisition module 305, configured to acquire a set of sample image information before the training module 301 performs pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence. The set of sample image information includes several pieces of sample image information, and the set of sample image information includes positive sample image information and negative sample image information. The positive sample image information includes dental arch model image information and tooth region rendering image information from the same patient, and the negative sample image information includes dental arch model image information and tooth region rendering image information from different patients;

[0176] The specific manner in which the training module 301 performs pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence includes:

[0177] Perform pre-training operations on a pre-determined image encoder based on first prior knowledge and the set of sample image information to obtain a pre-training result, and perform an update operation on the image encoder based on the pre-training result to obtain an image encoder trained to convergence, so as to constrain the shallow output of the image encoder;

[0178] Among them, the updated image encoder includes prior knowledge information for distinguishing tooth regions and gum regions.

[0179] It can be seen that implementingFigure 4 The described device can perform pre-training operations on the image encoder based on a set of sample image information obtained, which includes several positive sample image information and negative sample image information. And it can use the positive sample image information and negative sample image information to perform corresponding training operations on the image encoder, which is beneficial to improving the accuracy of the shallow loss function of the calculated image encoder, thereby improving the accuracy and intelligence of the training operations on the image encoder. And it is beneficial to improving the accuracy of the image encoder trained to convergence, and further beneficial to improving the subsequent intelligent positioning and intelligent detection of the tooth center points in the dental arch model based on different types of prior knowledge. And it is beneficial to be able to synergistically promote the improvement of the detection quality of the tooth center points through prior knowledge, a deep detection network, and the predicted heat map, beneficial to improving the accuracy of the detection of the tooth center points, and beneficial to improving the efficiency of the detection of the tooth center points.

[0180] In another alternative embodiment, as Figure 4 shown, the acquisition module 305 is further configured to, after the generation module 304 generates the center point prediction result of each tooth pair according to the aggregated feature information of the tooth pair, acquire a target center point heat map, where the target center point heat map includes the center point heat map of each tooth pair; generate a target center point prediction heat map according to the center point prediction heat map of each tooth pair;

[0181] The device further includes:

[0182] A calculation module 306, configured to calculate the difference parameter between the target center point heat map and the target center point prediction heat map, and perform a network update operation on the deep detection network based on the difference parameter to update the deep detection network;

[0183] Wherein, the difference parameter includes one or more of a semantic overlap penalty loss parameter, a cross-entropy loss parameter, and a mean square error loss parameter, and the parameter update operation includes a stochastic gradient descent algorithm operation.

[0184] It can be seen that implementing Figure 4The described device can obtain the heat map of the target center point, predict the heat map based on the center points of each tooth pair to generate the predicted heat map of the target center point, calculate the difference parameter between the heat map of the target center point and the target center point, and perform a network update operation on the deep detection network based on the difference parameter to update the deep detection network, which is beneficial to improving the accuracy and reliability of subsequent deep feature information extraction through the deep detection network, and is also beneficial to improving the intelligence and efficiency of subsequent deep feature information extraction through the deep detection network, thereby promoting the improvement of the center point detection quality through the deep detection network and the predicted heat map, and further being beneficial to improving the accuracy of tooth center point detection and the efficiency of tooth center point detection.

[0185] In yet another alternative embodiment, as Figure 4 shown, the shallow feature information of the sample image information includes the shallow feature information of each tooth pair included in the sample image;

[0186] The specific manner in which the determination module 303 determines the target feature map corresponding to the sample image information based on the shallow feature information includes:

[0187] Determine each tooth pair included in the sample image information. For each tooth pair, input the shallow feature information of the tooth pair into a preset multi-head attention model so that the multi-head attention model outputs the spatial attention map of the area where the tooth pair is located;

[0188] For each tooth pair, calculate the feature map of the tooth pair according to the spatial attention map of the area where the tooth pair is located and the preset weighted network deep features;

[0189] Determine the target feature map corresponding to the sample image information according to the feature maps of all tooth pairs;

[0190] Among them, the multi-head attention model includes at least one lightweight encoder, and each lightweight encoder has a semantic overlap penalty loss function matching the lightweight encoder. The semantic overlap penalty loss functions corresponding to different lightweight encoders are different, and the semantic overlap penalty loss function is used to constrain the lightweight encoder corresponding to the semantic overlap penalty loss function to collect the regional feature information of the semantic region corresponding to the lightweight encoder.

[0191] It can be seen that implementing Figure 4The described device can input the shallow feature information of each tooth pair into a preset multi-head attention model so that the multi-head attention model outputs the spatial attention map of the area where the tooth pair is located, and calculates the feature map of each tooth pair according to the spatial attention map of the area where each tooth pair is located and the preset weighted network deep features. It is possible to calculate the feature map of each tooth pair through the preset multi-head attention model and the preset weighted network deep features, which is beneficial to improving the intelligence and efficiency of outputting the spatial attention map of the area where each tooth pair is located, and is beneficial to improving the accuracy and reliability of outputting the spatial attention map of the area where each tooth pair is located. Thus, it is beneficial to improve the accuracy and reliability of calculating the feature map of each tooth pair, and further beneficial to improving the accuracy and efficiency of subsequent detection of tooth center points, and is also beneficial to improving the quality of tooth center point detection.

[0192] In yet another alternative embodiment, as Figure 4 shown, the specific manner in which the generation module 304 generates the aggregated feature information of each tooth pair based on the second prior knowledge and the target feature map through the deep detection network includes:

[0193] For each pair of tooth pairs, according to the second prior knowledge information and the target feature map, determine the target feature information from the feature map of the tooth pair, and the number of target feature information is at least one;

[0194] For each pair of tooth pairs, perform a preset feature aggregation operation on all the target feature information of the tooth pair through the deep detection network to obtain the aggregated feature information of the tooth pair;

[0195] Among them, the feature aggregation operation includes one or more of a feature splicing operation and a fully connected layer aggregation operation.

[0196] It can be seen that implementing Figure 4 the described device can determine the target feature information from the feature map of each tooth pair according to the pre-determined prior knowledge information, and perform a preset feature aggregation operation according to the target feature information of the tooth pair to obtain the aggregated feature information of the tooth pair, which can improve the accuracy and reliability of determining the target feature information of each tooth pair, and is beneficial to improving the intelligence and efficiency of determining the target feature information of each tooth pair. Thus, it can also obtain the aggregated feature information of each tooth pair through the preset feature aggregation operation, which is beneficial to improving the accuracy and intelligence of obtaining the aggregated feature information of each tooth pair, and further beneficial to improving the accuracy and reliability of generating the center point prediction heat map of each tooth pair based on the feature information of each tooth pair, and is beneficial to improving the accuracy and detection quality of detecting the center point of each tooth pair.

[0197] In yet another alternative embodiment, as Figure 4As shown, the specific manner in which the generation module 304 generates the center point prediction result for each tooth pair according to the aggregated feature information of the tooth pair includes:

[0198] For each tooth pair, according to the aggregated feature information of the tooth pair, perform a calculation operation on the aggregated feature information of the tooth pair through at least one pre-determined convolutional layer to obtain the feature calculation result of the tooth pair;

[0199] For each pair of tooth pairs, according to the feature calculation result of the tooth pair, determine the center point prediction heat map of the tooth pair and the classification result information of the tooth pair.

[0200] It can be seen that implementing Figure 4 The described device can input the aggregated feature information of each tooth pair into a pre-determined image decoder to obtain the target aggregated feature information of the tooth pair, and perform a calculation operation on the target aggregated feature information of each tooth pair through a pre-determined convolutional layer to obtain the feature calculation result of each tooth pair, and determine the center point prediction heat map of each tooth pair based on the feature calculation result of each tooth pair. It can specifically determine the target aggregated feature information of each tooth pair and specifically determine the feature calculation result of each tooth pair, which is beneficial to improving the accuracy and reliability of determining the target aggregated feature information of each tooth pair and determining the feature calculation result of each tooth pair, and is beneficial to improving the intelligence and efficiency of determining the target aggregated feature information of each tooth pair and determining the feature calculation result of each tooth pair. Thus, it can specifically determine the center point prediction heat map of each tooth pair, which is beneficial to improving the accuracy and reliability of determining the center point prediction heat map of each tooth pair, and is beneficial to improving the intelligence and efficiency of determining the center point prediction heat map of each tooth pair. Furthermore, it is beneficial to realize intelligent positioning and intelligent detection of tooth center points based on the dental arch model, beneficial to improving the accuracy of tooth center point detection, and beneficial to improving the efficiency of tooth center point detection.

[0201] In another alternative embodiment, as Figure 4 shown, the extraction module 302 is further configured to input the determined target dental arch model rendering diagram into a pre-determined deep detection network, and perform a feature extraction operation on the target dental arch model rendering diagram through the deep detection network to obtain the target rendering feature information;

[0202] The generation module 304 is further configured to generate a target center point prediction result matching the target dental arch model rendering diagram based on the target rendering feature information and the deep detection network; wherein, the target center point prediction result includes the center point position information of each target tooth corresponding to the target dental arch model rendering diagram and the center point classification information of each target tooth corresponding to the target dental arch model rendering diagram.

[0203] It can be seen that implementing Figure 4 the described device can input the rendering diagram of the target dental arch model into a pre-determined deep detection network, perform feature extraction operations on the rendering diagram of the target dental arch model through the deep detection network to obtain target rendering feature information, and generate a center point prediction result that matches the rendering diagram of the target dental arch model based on the target rendering feature information and the deep detection network. It can obtain the target rendering feature information and then generate the center point prediction result based on the deep detection network trained to convergence, which is beneficial to improving the intelligence and efficiency of obtaining the target rendering feature information and the center point prediction result, and further beneficial to improving the accuracy and reliability of the target rendering feature information and the center point prediction result. It can also realize the intelligent positioning and intelligent detection of the tooth center point based on the dental arch model, which is beneficial to improving the accuracy of detecting the tooth center point and the efficiency of detecting the tooth center point.

[0204] Embodiment 4

[0205] Please refer to Figure 5 , Figure 5 which is a schematic structural diagram of another dental arch model center point detection device based on prior knowledge disclosed in the embodiments of the present invention. As Figure 5 shown, the dental arch model center point detection device based on prior knowledge may include:

[0206] A memory 401 storing executable program code;

[0207] A processor 402 coupled to the memory 401;

[0208] The processor 402 calls the executable program code stored in the memory 401 and executes the steps in the dental arch model center point detection method based on prior knowledge described in Embodiment 1 or Embodiment 2 of the present invention.

[0209] Embodiment 5

[0210] The embodiments of the present invention disclose a computer storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute the steps in the dental arch model center point detection method based on prior knowledge described in Embodiment 1 or Embodiment 2 of the present invention.

[0211] Embodiment 6

[0212] The embodiments of the present invention disclose a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to execute the steps in the dental arch model center point detection method based on prior knowledge described in Embodiment 1 or Embodiment 2.

[0213] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0214] Through the specific descriptions of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, and the storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc memories, magnetic disk memories, tape memories, or any other computer-readable medium that can be used to carry or store data.

[0215] Finally, it should be noted that: The method and device for detecting the center point of a dental jaw model based on prior knowledge disclosed in the embodiments of the present invention only disclose the preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting the center point of a dental arch model based on prior knowledge, characterized in that The method includes: Performing pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; wherein, the first prior knowledge is used to perform pre-training operations on the image encoder so that the image encoder performs corresponding self-learning operations; Extracting shallow feature information of sample image information through the image encoder, and determining a target feature map corresponding to the sample image information based on the shallow feature information, wherein the target feature map includes feature maps of each tooth pair included in the sample image information; Generating aggregated feature information of each tooth pair based on second prior knowledge and the target feature map, and through a deep detection network; wherein, the second prior knowledge includes pre-determined prior knowledge for judging tooth categories; For each tooth pair, generating a center point prediction result for the tooth pair according to the aggregated feature information of the tooth pair; wherein, the center point prediction result includes classification result information of the tooth pair and / or a center point prediction heat map of the tooth pair.

2. The method for detecting the center point of the dental model based on prior knowledge according to claim 1, wherein Before performing pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence, the method further includes: Obtaining a set of sample image information, wherein the set of sample image information includes a plurality of sample image information, and the set of sample image information includes positive sample image information and negative sample image information, wherein the positive sample image information includes dental model image information and tooth region rendering image information from the same patient, and the negative sample image information includes dental model image information and tooth region rendering image information from different patients; The performing pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence includes: Performing pre-training operations on a pre-determined image encoder based on first prior knowledge and the set of sample image information to obtain a pre-training result, and performing an update operation on the image encoder based on the pre-training result to obtain an image encoder trained to convergence, so as to constrain the shallow output of the image encoder; Wherein, the updated image encoder includes prior knowledge information for distinguishing tooth regions and gum regions.

3. The method for detecting the center point of the dental model based on prior knowledge according to claim 2, wherein After generating a center point prediction result for each tooth pair according to the aggregated feature information of the tooth pair, the method further includes: Obtaining a target center point heat map, wherein the target center point heat map includes center point heat maps of each tooth pair; Generating a target center point prediction heat map according to the center point prediction heat maps of each tooth pair; Calculating a difference parameter between the target center point heat map and the target center point prediction heat map, and performing a network update operation on the deep detection network based on the difference parameter to update the deep detection network; Wherein, the difference parameter includes one or more of a semantic overlap penalty loss parameter, a cross-entropy loss parameter, and a mean square error loss parameter, and the parameter update operation includes a stochastic gradient descent algorithm operation.

4. The method for detecting the center point of the dental arch model based on prior knowledge according to claim 3, wherein The shallow feature information of the sample image information includes the shallow feature information of each tooth pair included in the sample image; Among them, determining the target feature map corresponding to the sample image information based on the shallow feature information includes: Determining each tooth pair included in the sample image information. For each tooth pair, input the shallow feature information of the tooth pair into a preset multi-head attention model, so that the multi-head attention model outputs the spatial attention map of the area where the tooth pair is located; For each tooth pair, calculate the feature map of the tooth pair according to the spatial attention map of the area where the tooth pair is located and the preset weighted network deep feature; Determine the target feature map corresponding to the sample image information according to the feature maps of all the tooth pairs; Among them, the multi-head attention model includes at least one lightweight encoder, and each lightweight encoder has a semantic overlap penalty loss function matching the lightweight encoder. The semantic overlap penalty loss functions corresponding to different lightweight encoders are different, and the semantic overlap penalty loss function is used to constrain the lightweight encoder corresponding to the semantic overlap penalty loss function to collect the regional feature information of the semantic region corresponding to the lightweight encoder.

5. The method for detecting the center point of the dentition model based on prior knowledge according to claim 4, wherein Generating the aggregated feature information of each tooth pair based on the second prior knowledge and the target feature map through a deep detection network includes: For each pair of tooth pairs, determine the target feature information from the feature map of the tooth pair according to the second prior knowledge information and the target feature map, and the number of the target feature information is at least one; For each pair of tooth pairs, perform a preset feature aggregation operation on all the target feature information of the tooth pair through a deep detection network to obtain the aggregated feature information of the tooth pair; Among them, the feature aggregation operation includes one or more of a feature splicing operation and a fully connected layer aggregation operation.

6. The method for detecting the center point of the dental arch model based on prior knowledge according to claim 5, characterized in that For each tooth pair, generating the center point prediction result of the tooth pair according to the aggregated feature information of the tooth pair includes: For each tooth pair, perform a calculation operation on the aggregated feature information of the tooth pair through at least one pre-determined convolutional layer according to the aggregated feature information of the tooth pair to obtain the feature calculation result of the tooth pair; For each pair of tooth pairs, determine the center point prediction heat map of the tooth pair and the classification result information of the tooth pair according to the feature calculation result of the tooth pair.

7. The method for detecting the center point of the dental cast based on prior knowledge according to claim 3, wherein The method further includes: Inputting the determined target dental arch model rendering diagram into the pre-determined deep detection network, and performing a feature extraction operation on the target dental arch model rendering diagram through the deep detection network to obtain target rendering feature information; Generating a target center point prediction result matching the target dental arch model rendering diagram based on the target rendering feature information and the deep detection network; among them, the target center point prediction result includes the center point position information of each target tooth corresponding to the target dental arch model rendering diagram and the center point classification information of each target tooth corresponding to the target dental arch model rendering diagram.

8. A dental model center point detection device based on prior knowledge, characterized in that, The device includes: A training module, configured to perform pre-training operations on a pre-determined image encoder based on first prior knowledge to obtain an image encoder trained to convergence; wherein, the first prior knowledge is used to perform pre-training operations on the image encoder so that the image encoder performs corresponding self-learning operations; An extraction module, configured to extract shallow feature information of sample image information through the image encoder; A determination module, configured to determine a target feature map corresponding to the sample image information based on the shallow feature information, wherein the target feature map includes feature maps of each tooth pair included in the sample image information; A generation module, configured to generate aggregated feature information of each tooth pair based on second prior knowledge and the target feature map through a deep detection network; wherein, the second prior knowledge includes pre-determined prior knowledge for judging tooth categories; for each tooth pair, a center point prediction result of the tooth pair is generated according to the aggregated feature information of the tooth pair; wherein, the center point prediction result includes classification result information of the tooth pair and / or a center point prediction heat map of the tooth pair.

9. A dental arch model center point detection device based on prior knowledge, characterized in that, The apparatus includes: A memory storing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory and executes the prior-knowledge-based center point detection method for dental arch models according to any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which are used to execute the prior-knowledge-based center point detection method for dental arch models according to any one of claims 1-7 when called.

Images